Funny enough, this has been a major focus for me lately, so I thought I’d lay out why the BIM opportunity matters and give a taste of a few things I’ve been working on.

There are essentially two ways to think about BIM:

First, in order to understand BIM, you need to understand the utility of BIM in the built world. BIM = Building Information Modeling:

Many teams and personas collaborate on a BIM model. An architect will translate design intent into the building shell and floor layouts, exteriors, and interiors; engineers will plan MEP routing and structural integrity; contractors, schedulers, and more will then use this to plan and subsequently build the building.

BIM is useful. Construction firms are among the most operationally competent organizations in the world, and major projects are often planned and executed with extraordinary rigor. Yet, there’s a fundamental problem with BIM as a discipline:1

It doesn’t move at the speed of agents. We will get to why in a second, but let’s first talk about what this practically means:

First, let’s assume that we are still in the precon/design phase of a project. And a structural engineer upon reviewing the current model indicates to the architect that they need to increase the slab by an inch and a half.

What happens next is that an architect has to revise the model, adjust the slab, extend the walls, and then verify that the downstream drawings still reconcile.

For a seemingly minor update on a building (and this isn’t an overly complex building), an architect can look to spend upwards of 6 hours adjusting the model.

Are there pathways out? Sure. These usually involve scripting in Revit (which Claude Code and others have made easier). To one of a16z’s point, perhaps this represents a pathway to better AI outcomes in AEC with BIM.

However, in my view, this is simply not sufficient. But to get there, perhaps we should first consider the alternative. What do we want agents to be able to do in BIM? Is this about mere time savings or something else altogether?

For a second, let’s assume that we have truly agentic BIM - eg agents actually build and maintain models. What are the second order consequences?

Faster modeling means more models.

This one is easy. If the cost to build full fledged models drops significantly, the number of models proportionally grows. Architects and engineers now propose dozens of design permutations at the precon stage with the capability to turn these into fully engineered models within a matter of hours.

Cost and review time go down with impacts on total cost and project speed.

Construction often orbits around constructability reviews, clash detections, and plenty more to ensure that when a project moves from the design phase to the build phase, the project is actually constructable.

RFIS and other construction processes change

ZeroRFI recently launched as an owner’s representation firm rollup with the goal of enabling “zero RFIs” - eg requests for information. RFIs usually occur when plans/drawings aren’t sufficient to gauge design and build intent for a contractor implementing an aspect of the project. BIMs nominally were supposed to address this by having large amounts of metadata around materials, scheduling, costs, and more. There are now more RFIs than ever.

Solve BIM in an agentic native way and guess what? RFIs don’t go to zero, they simply become questions run against the agentic BIM model.

BIM is upstream of all of this and these are all real time and money costs on a construction project. But they aren’t even the full benefit here.

Let’s keep going around even more speculative (but fully realistic) impacts.

Scheduling:

1. Currently a construction project is scheduled via master schedulers and project controls teams. These teams are essentially oracles that can synthesize the entire project progression, resource constraints, and time constraints and get buildings built on time.

2. There’s been movement towards 4-D BIM scheduling, where a firm takes the schedule and the BIM itself and shows the building itself going up over time in accordance with the schedule. This is currently highly expensive, but highly useful for projects where it may not be obvious from glancing at a model and a schedule if crews will be overlapping work areas leading to slowdown.

What happens when this becomes cheap and agentic? Well it’s obvious that this would completely change scheduling and project controls, but that’s the least of it. In 20 years, these time and geospatial dimensions are the core aspects of how a construction team will manage fleets of robots to deploy into a build. Everything is downstream of this.

And yet, BIM remains a time intensive asset that today has not been amenable to these techniques. Why is that? Let me give you my thoughts, often synthesized from others.

Agentic BIM is a kernel problem

Here’s a fun experiment: what happens if you give an LLM a Revit .rfa file and ask it to perform operations, perhaps lengthen a component. This seems reasonable right?

Agents fail almost completely. It’s essentially hopeless. That alone is not a sufficient analysis. Why do they fail so completely? Is geometry simply too complex for the current agent generation? Not really. There’s more going on here.

First, Revit doesn’t natively support interoperability. And in fact, geometry kernels writ large are not very interoperable. The industry attempt to standardize BIM files, IFC, has been useful but by and large there are conversion issues moving files between different BIM systems. This is primarily due to to how legacy geometry kernels represent geometry.

Revit and other CAD/CAM sofware works by modeling b-reps, boundary-representations. In other words, they model the edges of a surface, plane, or vector and then their kernel implementations subsequently extrapolate the shapes to be built from the boundary-representations. Suffice to say, the specific implementation from one b-rep based system to another can be incredibly lossy. As a result, there’s almost zero training value for models to give them b-rep based files and ask them to, they mostly aren’t learning geometry or geo-spatial awareness, but instead something to the effect of “how one BIM provider would implement geometry”.

The best writer and practitioner on this is Blake Courter:

Above, we assumed that the AI could somehow make sense of the CAD geometry. It turns out, that’s also a hard problem.

Let’s say we want to represent a closed shape like a circle. In traditional CAD, we’d model it as a loop that joins the start and the end of a curve or list of chained curves. There’s going to be at least one point where there’s a “parametric seam.”

Now, we might want to know whether a given point is inside or outside of our shape, which is the most basic test in representing a “solid model”. For example, a 3D printer needs to put material where the shape is and other stuff where it is not. How can we reason about this problem? We need an algorithm!

One algorithm might be to pretend that the point is like a cow in a fenced-in pasture and to test if it can escape. We could shoot a ray from the point off to infinity in one direction. If we cross the boundary of the shape an odd number of times, we must be inside the fence, and if an even number, outside.

But does the algorithm always work? What if one of the rays is just tangent to the shape? What if we make a mistake around the parametric seam? Apparently arbitrary cases cause us to get the answer wrong. And indeed, most traditional CAD kernels answer this kind of question incorrectly often enough to cause issues that require the attention of CAD specialists.

All this arbitrariness becomes an impediment to training AI models on CAD data. Not only does the AI need to understand the design, but also, to do anything more than produce output, it needs to understand how the geometry engine will manipulate the geometry, blowing up the latent space.

The entire article is worth a read, but note here the challenge, which isn’t solely unique to agents: boundary-representations of an object are simply not sufficient for an agent, often an algorithm, or even the current kernels themselves, to sufficiently gain perspective on the actual geometry of an object.2

Multiply this frustration by tens of thousands of objects inside a building file, and we can start to grok the magnitude of the problem with b-rep based systems for representing geo-spatial data across construction and engineering.

Suffice to say, it turns out the problem isn’t just “train agents on existing BIM/CAD data”, the problem is largely “we need a different approach altogether to train agents and ultimately enable the creation of high fidelity geometric assets and geospatial awareness in LLMs.”

But note here, the payoff. This problem currently occurs across every single engineering and construction workflow.

So what is the better path, if one exists at all?

Well, Courter and others have outlined a potential remedy here too: implicit geometry.

Implicit Geometry

If we were to ask ourselves, how would we build better geometry kernels amenable to a) ultimately better models, b) and that are useful for agents as training data and for application, a couple core observations occur:

1. We need to treat geometry as code (I’m blatantly stealing this from Courter by the way, all credit to him).

   1. Practically this means that we know always and without a doubt if the geometry is going to compile. We aren’t reliant on understanding the tens of millions of custom code that geometry kernels have implemented to handle edge case scenarios around b-rep.

   2. LLMs can reason over parametric relationships and augmenting one variable (lengthening a slab) instantly adjust the rest of the model as well.

2. We likely want the geometric representation to be primarily math based. Agents are increasingly great at math and code.

Implicit geometry fulfills both of these conditions. In the interest of brevity, for our purposes, any object can be represented as a mathematical function. Operations on these function can then be performed via code.

Why hasn’t this already been done?

1. There was never a fully compelling reason to move geometry off of b-rep kernels prior to the advance of agents. If humans would still mostly be manipulating a UI in a b-rep vs. f-rep based kernel, there wasn’t a compelling reason to move in this direction.

2. There are some true problems around meshing objects. As Matt Keeter puts it, “good meshing of arbitrary implicit surfaces remains an unsolved problem!”3

But in my view, the time is now to fully figure these problems out. First, while engineering CAD vs. BIM needs are different, the EV from these in the industries themselves are huge. There’s also a colorable argument that simulations of worlds will be built via new f-rep based kernels.

Eg, if implicit geometry and subsequent training on these data sets yields agents that can perform geospatial operations at token speed, the ramifications may be far larger than simply BIM and CAD.

1. Gaming engines: to what extent does implicit geometry enable faster iteration and development of gaming assets and games.

2. Robotics: Does the sim2real problem get solved via enabling mass simulation of real world environments that can be scaffolded in mere minutes and then RL’d against? I don’t know the answer, but in my view the progress that Physical Intelligence and General Intuition are making raise this question.4

And in short that may be right expected payoff. In some sense, agents that can represent the physical world geospatially can train against the same world5

And in the future, it’s not hard to imagine agents in the cloud coordinating with robots on the ground to plan the work to be accomplished, all with the same BIM model.

Anyways, if you want to chat around these ideas, as always please reach out.

Subscribe now

“When you see something technically sweet you go ahead and do it, and you argue about what to do about it, only after you have achieved success.”

Oppenheimer

In the emerging effort to make the full TAM of labor addressable via software and robotics, there is a growing appreciation for how complex the world actually is.

That complexity is not a reason to sit on our hands and wait for things to play out.

We are finally at a moment where there is enough understanding around how model capabilities intersect with data, RL, evaluations, and the broader vendor stack that ambitious organizations can begin to behave differently.

The strategies are increasingly legible. The vendor ecosystem has matured. Open-source models have proliferated. The infrastructure required to pursue durable advantage has now emerged.

And so the implication is fairly straightforward: this is a moment for concentrated research ambition in every vertical.

This is not to suggest that vertical SaaS companies should become labs, but instead the organizations most likely to define their categories from here will be the ones that treat research as a competitive advantage in their vertical. Durable advantages will not come from merely consuming frontier models. It will come from building the internal capacity to turn model capability into a science.

There is almost no better crop of companies to pursue this advantage in theory than current vertical SaaS companies at scale. The revenues, customers, and data they commandeer have been hard won. Now is the moment to parlay this into the future.

And so now is the time for vertical SaaS companies to develop research organizations that can turn model capability into a decisive advantage.

Subscribe now

On Acquisitions

First, it’s important to realize that the current strategies around AI acquisitions have had far more to do with product organizations than research. This is far different from how the labs, data companies, and RL companies have assessed the problem.

While vertical SaaS has prized acquisitions for near-term revenue strategies, lab acquisitions have been built more around the underlying data recipes, harnesses, research, and ML talent inside of their acquisitions.

This is not meant to be a critique of acquisitions ad hoc. These have been valuable acquisitions for accelerating product roadmaps. I’m simply suggesting that most vertical SaaS companies, even following these acquisitions, have insufficient research impulses to create novel agents that endure in the face of base model pressure.

But this isn’t merely about a defensive strategy. There’s one question that matters:

Do you believe the TAM is all of human labor?

Ignore for a second all the secondary questions you ultimately jump to, including margin profile, competitive dynamics, commodity pressure, and the rest.

Do you believe that the TAM is all of human labor?

If the answer is yes, then the question for vertical SaaS companies is simple: how do you go about capturing it in your vertical? Can you create an organization that does not just benefit from favorable tailwinds from base models that others will thrive from as well, but become a company that dictates the frontier for your industry?

If the answer is yes, the payoffs are not merely SAM expansion, but a range of partners to help get you there. In other words, the EV of a competent research organization is extremely high from both a credibility and revenue perspective.

Let’s look right now at how the labs behave. Anthropic and OpenAI are both increasingly going to market with partners: Harvey, Rogo, Sierra, and a litany of others. What do all of these partners have in common? A research-centric approach to the problems they are trying to crack.

Right now, the labs primarily judge traditional software companies as consumers of inference internally. Start addressing the TAM of labor and paths begin to open up.

Who do the new wave of vendors like Applied Compute, data companies, and more try to target? Companies with a tractable vision around the future of work in their vertical.

What are the companies most likely to endure? Companies like ServiceNow, which consistently put out some of the best research around frontier capabilities, small models, and more.

The research arm of these companies is a feature, not a bug.

In short, there are a number of tailwinds that directly benefit organizations that think this way.

In this conception, the base mission of research organizations is to deploy an increasingly complex supply chain of vendors and partners to deliver frontier model capabilities that product and engineering teams can then commercialize. Human data, RL vendors, compute vendors, and I’m sure many more will continue to proliferate.

What are the best examples of these research organizations in vertical-ish markets today?

Three examples: Shopify, Ramp, and ServiceNow.

Ramp Labs is a particularly useful example. Look at the diversity of experiments that Ramp is running, everything from Excel experiments to RollerCoaster Tycoon. The benefit is twofold:

AI is ultimately a talent, data, and compute game. A range of experiments signals to each of these markets that the company is a credible partner on the AI frontier.

Compute markets are trending weirder. What’s the payoff from RollerCoaster Tycoon? An internal team that increasingly knows how to handle long-horizon tasks and model evaluation. What’s the payoff from Excel agent capabilities for a fintech company? That one should be obvious.

The goal of research for all of these organizations is not just to have fun. It is about developing an internal team, muscle, and ultimately a model capability analysis function that can increase the rate of commercializable products.

In many ways, it is no different than what every physical-world company of scale does. Physical product companies spend hundreds of millions on R&D.

The future of software is beginning to look similar, except with payoffs measured in trillions across the economy.

Where should the research organization live?

Should it be concentrated in product, engineering, data, or something else altogether?

While the specifics will differ by organization, here’s my unsolicited take: it should live as its own organization.

There’s at least one reason to frame it this way: the goals of research and other organizations diverge before they converge.

Product organizations are concerned with driving near-term revenue from customers by building and managing products with large uptake. Their time horizons and KPIs are built around existing products, with some portion of time spent on near-term new product opportunities.

Research organizations are primarily concerned with identifying longer-term moats and model capabilities that can be turned into durable revenue. Their concern is not solely near-term product capabilities, but instead on pulling the frontier forward by 12–24 months. With novel model research, data, and evaluations, their goal is to deliver transformational capabilities that can result in long-term durable revenue and products at a pace that other companies in the vertical cannot match.

Of course, in practice this is going to involve a great deal of collaboration, with all aspects of the broader company contributing to AI advancement and becoming increasingly AI-native.

A thriving organization should be managing SLAs with product to deliver better performance at lower cost on a recently launched product, while simultaneously advising product that it can launch a first-to-market product with a newly developed model capability.

Which brings us to the fun part of this piece: where do we start? After all, research should be focused on a tacit goal. So what should research teams actually focus on?

Model capability underlies roadmap

By now, I’m sure everyone has seen Anthropic’s graph around AI coverage by industry and profession. Some important caveats:

Theoretical AI coverage does not mean future AI coverage. It means what current models could theoretically cover today.

The more I look at this, the more I think it’s reductive, which in turn helps clarify the opportunity.

But it is still a great place to start. Here’s my heuristic from the graph:

First, if theoretical model capability is greater than observed AI usage today, research should specialize in agent ergonomics

By agent ergonomics, I mean the tools, sequential reasoning, and environment that the model interacts with.

In other words, a capability overhang means two things:

1. There is likely a product gap in deploying models.

2. The harness around the agent is inadequate for the length of task in the vertical.

As a result, research organizations are responsible for determining exactly where the harness and agent are inadequate and subsequently enabling product organizations to fully focus on the user ergonomics around the product.

The focus here is on evaluations, environments, and tool calling, with a special eye toward determining where the agents begin to fail on sufficiently long-horizon tasks.

If there is a current model capability gap and the work should be addressable via software, research teams should focus on data and RL

This is where a dedicated organization to manage research vendors pays off, in my view.

There is a litany of problems inside any vertical that are mostly the result of models lacking verifiability in the domain, and where the problems are sufficiently multimodal rather than just text-based.

These can become some of the most durable revenue streams for vertical SaaS companies as a result, but they demand commandeering data and likely performing RL in order to experience the payoff.

I have an ever-growing list of these that shall remain secret for now. Suffice it to say, the number of problems that we have spent adequate RL on is sufficiently low that hundreds of millions per vertical in revenue is lying in the gutter waiting for vertical SaaS companies to pick up the crown.

Lastly, if total usage and theoretical model capability are both growing, research organizations should spend their time on cost fidelity

Inference is expensive. What should your product organization do when a product has rapid uptake but the margin profile is less than ideal?

Wouldn’t it be nice to hand this problem off to a research organization to deliver cost improvement through RL? It would be.

Lastly, I’ll say this: plenty of vertical SaaS companies touch physical-world operations where sensors and robotics are going to further expand the SAM of labor.

While it is still too early for research organizations to focus too heavily on this problem, future-facing executives should keep an eye on the trendline around rapid physical AI improvement and be situated to take advantage of it the moment the physical world becomes tractable.

We aren’t there yet, but it is certainly coming. The EV of a research organization as this moment approaches will become even higher.

In short, vertical SaaS companies of scale have an unfair advantage with their current revenues: they can become the research organizations that define the frontier in their industry.

The dilemma is how to do this while managing an existing product and business alongside model capabilities that can cannibalize revenue.

Most companies cannot afford to wait, given the rise of so many vertical AI companies and forward-deployed efforts committed to eating all of your software revenue and the labor budget.

Research organizations are not just a plausible path forward. They may become part of the organizational transformation that enables vertical SaaS companies to grow at disproportionate rates going forward.

Subscribe now

If you’re exploring this, please reach out.

Subscribe now

Anthropic in particular seemingly drops a release every other day that makes writing, debugging, or securing code easier. So of course, the rational response is to say “code moats are dead” and freak out.

But this is obviously a reductive narrative. Software with a true agent play still isn’t being repriced. There are constant whispers from the labs around their software-layer ambitions to compete with Workday, Salesforce, and others—a blatant hint that there is at least some value in the software layer. Even agent applications like Claude Cowork seem to be prioritizing integration surfaces into other apps.

I will concede that software as a category has fewer moats, more distributional questions, and even existential questions about its future.

But let’s make no mistake: there is no conceivable world where systems of record die. Systems of record are going to be here until the sun goes out or the earth becomes one mega datacenter.

After all, what is a system of record? It’s a shared data plane and workflow architecture for firms to coordinate their decisions. Right now that coordination is done by humans. In the future it will increasingly be done by agents.

Agents are not the Borg. There is no collective memory latent in their architecture. And agents, while faster than humans, are still bound by time. Businesses have historically solved the Borg problem and the time problem with systems of record.

The current systems of record debate is obfuscating several themes and drawing conclusions that are unwarranted. I agree with two points:

• Current systems of record are quite obviously limited.

• They were not made for agents. They were not even really made for people. They were made for managers.

That’s going to limit your upside when a sudden mass of agents wants to use your system to actually get things done.

Current systems of record impose a tax on agent productivity rather than being additive to it. No existing system is immune to this. And users are increasingly not going to tolerate your high gross margin business if it has dropped the ball on helping them get more work done productively.

The fact that current systems of record are bad for agents does not logically mean that software productive for agents is going to be thrown out or shouldn’t be built. You need another premise.

And the only premises I can imagine here involve who takes over the system of record. This is ultimately the only question that matters for the terminal value of systems of record and how much of the incoming agent workload they capture.

From my perspective, the system of record question matters more, not less.

It’s so important that it informs everything from agent training to inference costs and all the way to how work gets done by agents in the future.

There are roughly four scenarios.

The labs own every single system of record

In this scenario, your buyer chooses between OpenAI Sales and Claude Sales. Maybe, if they’re a software hipster, they roll their own Kimi open-source Salesforce.

I think the labs will try. I think they will even succeed in a lot of cases. Salesforce isn’t cool. You know what’s cool? OpenAI Sales.

But I also think the labs run into foundational questions with this approach. First, how does the game theory shake out? Do companies have an incentive to fully lock in their AI provider as their system of record?

The existential challenge for labs pre-AGI is that their ambitions in the application layer in every vertical are held in check by the competitive landscape in model training. They are fighting thousand-front wars over inference where customers expect growing improvements across video, images, text, computer use, shopping, and more.

Two insights:

• The labs will attack the system-of-record layer to the degree they need to in order to tackle the true TAM of human labor.

• The labs are well aware of how ill-suited current systems of record are for agents. If this remains a fundamental constraint, they will certainly compete in the app layer or even rebuild for individual customers.

On the flipside, the coding wars have proved that switching costs on the inference layer are near zero. Customers by and large want to commoditize the model layer and will choose the best-performing model with zero loyalty.

This somewhat prohibits any concerted pre-AGI effort to build enterprise-grade integrated applications. The minute you commit to building a CRM and integrated sales agent is the minute you commit to having the best sales evals and the best outcomes for your customers.

The alternative is to let vertical companies fight proxy wars on your behalf. Let them build the neo-systems of record, develop the right evals and datasets, and then buy them outright.

This is in fact what’s playing out with Anthropic and OpenAI competing for the finance, health, law, and coding verticals with a mix of model wars and proxy wars through software partners.

As a result, I’d posit the following: the rumors about the labs entering the application layer are more of a condemnation of current systems of record than a stance that systems of record are going away.

If you move into the system-of-record layer, it’s obviously because you think your agent is capable of performing work at human parity but is not able to because the core system of record doesn’t allow it to.

In the same way, I think the emphasis on computer use by labs is interesting. It’s a concession that if we could only improve agents’ ability to use UIs built for humans, they’d be able to accomplish way more. Maybe the problem isn’t just the UI. Maybe it’s the underlying system of record itself.

A quick sidenote on MCP: MCP was meant to be a Band-Aid over current systems of record and apps. They work, but they still don’t work that well for actual work. Some of this is a training problem, and some of this is that current software products are simply incapable of exposing anything other than database rows and not decisions. It’s not actually that useful to get an API connector into a current system of record. The agents are smart, but if you dropped an outsider into a random company’s CRM and asked them what any of it means, they would look at you like you’re crazy.

When OpenAI hints that it wants to build a CRM, what it’s really saying is: Benioff can’t give our agents what they need.

So how should we view the labs as system-of-record entrants? They’re interested, but content to fight proxy wars. They want partners across verticals that structure themselves as both research partners and product companies.

Custom software dominates to such a large degree that the long tail proliferates without large software megacaps

The second scenario involves coding models essentially solving software as a category. The plausible version, in my view, involves a long tail of developers using coding models to create custom solutions for businesses at scale.

In this case, there’s no CRM category to speak of. There’s simply a sea of custom developers creating massive productivity for individual companies, and they all sort of roll their own software.

You could even imagine some of these custom developers deciding to specialize in a unique customer segment across a vertical. Let’s call them vertical software company-specific software builders. The game sounds familiar, doesn’t it?

If someone develops expertise in deploying custom software inside a vertical, but it’s made even more custom to specific companies, isn’t this just the next evolution of vertical software? Yeah, maybe it actually is.

Now maybe you surmise that Claude itself is going to be just that damn good. You don’t need the team, you just need Claude. And as long as Claude can get the right context for your specific business, how it works, what your priorities are, what would bring you new revenue, what would solve your operations, and what your customers would think about... wait. What if we built a system of record for Claude to build the right system of record and software for your business?

It really is just turtles all the way down, isn’t it? Aren’t we all merely competing to see who gets the true system of record to render agents extraordinarily productive? The layer of abstraction may just start to look a bit different.

In some raw way, the thing that is going to matter for all these approaches is a context graph spanning the decisions of an organization. Everything comes back to intent, even in building software to crystallize business intent.

I ultimately think this direction is highly plausible, but it will normalize to different systems of record that unleash custom software viability for businesses, likely in a vertical. To a large degree, the ontology of businesses in a vertical is similar. In other words, maybe the Palantir approach comes to fruition for the mid-market in a significant way.

Current systems of record prevail in their current state. Long live Salesforce. Long live Veeva. Long live every other incumbent.

I’ll go on record: I think this is actually the most implausible direction for the vast majority of verticals. As I’ve said, I think most current systems of record are simply not up to snuff for agent work.

I’m of course biased. Others like Brendan Keeler have made strong cases for why the systems of record will prevail. Keeler makes the point that Epic has entered the medical scribe wars and simply priced scribing for free to prevent competitors from entering.

This, in my opinion, gets at the heart of the current system-of-record conundrum. Systems of record are entirely capable of developing point solutions for a specific pain point in the industry. But their architecture is not conducive to what I’m coining: decisioning at token speed.

Transcribing a medical visit is a largely different task than coordinating all of care delivery, follow-ups, insurance operations, pharmaceutical operations, and a litany of other matters. Hospitals currently rely upon dozens of specialists coordinating this care across an individual patient. They’re specialists in one area of the business, have dedicated modules for their tasks, and frankly, do a damn good job in this current construct.

This architecture made tons of sense when our limitation was human inference. But agents and humans reason differently, with different ramifications.

It’s not clear that replicating human specializations to agent specializations 1:1 is the right approach. For instance, agents are context hungry and increasingly capable of long-horizon tasks. As long-horizon tasks approach years, will it make sense to have dedicated agent teams that handle one aspect of care delivery for all patients, or one agent per patient that handles all aspects of care delivery?

You can follow this same trendline across every human team in every industry. The answers will differ, but the throughline is clear: if agents are capable of longer-horizon tasks at higher speeds than humans, the supply chains of intelligence may be radically different than our current ones.

But again, this isn’t how current systems are built to operate. Structuring systems for inference requires structuring for decisions, context, and long-horizon tasks. How many current systems of record fit the bill?

Systems of record can build a point solution. That’s no longer the game. The game is which system of record can support 10x the workforce, maybe even 100x the workforce, inside a system making 1000x the decisions per day. Most are not built for this, either from an agent lens or from a human operator who has to monitor these decisions.

Neo-systems of record providers proliferate:

Let me draw a scenario out.

Imagine a world where the agent layer is mostly detached from the system-of-record layer. What would that world look like?

It would probably look like a world where users primarily interface with agents via their phone, in a dedicated app like Claude Cowork, or through a business messaging platform. Companies would spin up dedicated servers for agents where they want to fully automate work. They could seamlessly switch from chatting with their AI boyfriend to performing work with their AI intern.

That world looks highly plausible to me. The lesson from OpenClaw and Claude Code and others is that the more agents can replicate the local compute environment of their user, the more productive they are. Long-running agents likewise imply long-running compute with a wide-reaching span of tools across email, docs, Excel, software systems, and more.

How does a business coordinate all these different agent systems to accomplish real work? I think they’re going to need an excellent system of record.

And if the rate of agents, and the value of their decisions, continue to grow, buyers are going to want someone on the other end of the line to help them manage 10,000 agents working concurrently in their actual business, with corresponding proof that the work product being produced is up to par.

What if the game in front of vertical AI is less about building a voice agent and more about building the neo-systems of record that can ultimately support hundreds, if not tens of thousands, of agents performing valuable work for a business?

Will it make sense to internalize that? Maybe. Doubt it for everyone in the mid-market and down. Can the agents do it themselves if they get smart enough? Yeah, they’d build their own system of record. Turtles all the way down.

I realize this is not the trend. Vertical AI is still mainly shipping product form factors around model capabilities.

But now, I think it’s not ambitious enough. And it’s not even necessarily what buyers need. Vertical AI has spent so much time trying to be slightly better than Claude at vertical work product. That may be the wrong front.

If the MO is flipped and vertical value is decided on concurrent operations managed across local and cloud agents, I think vertical startups will end up more resilient to the model layer.

First, value capture seems likely better in the long run. Price according to the number of agents working concurrently and productively. This may scale to far greater dollar amounts than seat-based pricing ever could.

Engineers now commonly spin up 4–5 Codex or Claude instances to work concurrently on their codebase. I know OpenClaw users spinning up 5–7 sub-agents to work on parallel tasks. What if the number of agents working inside a neo-system of record dwarfs the number of employees at the firm by 4–5x? Account for cloud agents, and this easily could eclipse 10x.

Second, rather than trying to create a harness around current model capabilities, you’re instead spending time on harder research questions in the vertical: namely, how to empower increasingly long-horizon tasks and complex decisioning. These are valuable data and training questions that companies built around this realization will likely encounter and solve far faster than we otherwise would.

And this brings me to my final point. So much of the agent architecture game right now is about building the right harness for agents. Give them the right context, the right tools, and let them recursively work through problems.

I agree. But my contention is that the best harnesses will simply be the best systems of record.

In my view, SaaS and systems of record are not dead. In fact, we need far superior systems of record than what we currently have if we want the agent productivity boom we need across verticals.

Buy the ticket, take the ride

Subscribe now

While there are certainly some companies that have spent time training models, most of the optimizations have been at the prompt level or in the supporting system around LLMs.

This has made quite a bit of sense over the past few years.

1. Vertical AI companies were cognizant that they were going to get a performance upgrade without spending really any time on R&D/ML research. Every new model release was a guaranteed performance upgrade.

2. The foundation labs struck a peace dividend by depreciating their older models rapidly and releasing smaller, cost-optimized models.

So the net result was a race for industry adoption with most vertical AI companies paying less attention than probably is warranted to questions around the model layer or even research-oriented questions around how to create systems that maximized current AI capabilities (a lot of the context graph discussion is around this reality).

Simply put, there wasn’t a clear throughline between the model layer and the research questions being asked around data recipes, algorithms, RL environments, and the value they were looking to create in their vertical.

Very quickly, that’s all begun to change. I’ve spent the past two months talking to a wide-range of players across the compute, data, and RL environments landscape to understand the reasons why this is changing.

Don’t worry, some of these interviews are going to be published very soon. But I first wanted to talk through the factors themselves shaping the market:

Subscribe now

Everyone is getting sophisticated around algorithms as the knowledge disseminates

ML techniques are continuing to proliferate. Before diving into the post-training layer around LLMs, it’s important to note that ML sophistication doesn’t necessarily just mean LLMs. Companies are increasingly building more traditional models as a core part of their infrastructure for whatever applied problem they’re tackling.

But as it applies to LLMs, training is no longer a trade secret only known in the labs. RL techniques and optimizations are now widespread, frameworks like verl and solutions like Tinker have delivered teams the ability to work on these without having to know a ton about the low level engineering. And of course, open source base models like the Qwen series are powerful out of the box and ripe for these recipes.

The barrier to entry on the research side has lowered with the net result that more companies are at least beginning to explore post-training techniques. So far this has been primarily seen amongst the coding players, but there are early indicators that across law, healthcare, finance, and more, the dissemination of algorithms is causing teams to experiment in ways they weren’t a year ago.

At minimum, applied AI companies are thinking about their task as fundamentally a mixture of algorithmic design.

Subscribe now

Compute continues to commodify

While still expensive, the last year has seen the rise of inference hosting services like Baseten, spot markets for compute training develop, and even players beginning to formalize true commodity derivatives products for compute.

Whereas most teams 2 years ago would have had to procure a GPU cluster on their own, they can now interact with compute clusters in a similar manner to traditional cloud services. This naturally enables companies to experiment more + scale up in more capital efficient manners if their smaller model starts to become adopted.

This changes not only the unit economics of training but of model distribution. And since potentially higher model performance can be delivered for even lower COGs than using foundation models, this is becoming more and more attractive.

Data is the last undefined asset and where moats are accruing

I want to be careful here to define data in a particular way. What I am not suggesting here is that raw data itself is a particularly good moat. As everyone is well aware, interoperability has meant that structured + raw data is leaving and entering systems all the time. The storage of data is not really proprietary nor that valuable for applied AI companies. While someone will continue to own data storage (the systems of record), the data itself is not uniquely valuable to training.

What is valuable is what I’m going to call the data recipe - ie how do agents when looking at a problem or task provided work their way through a problem to an economically valuable result.

And this is where all the interesting dynamics in applied AI are headed. Perhaps it’s helpful to first consider how the labs are ascertaining this today:

Expert marketplaces

Expert marketplaces like Mercor, Turing, and others derive their data recipes from having experts in a domain create tasks and rubrics for training models on how to make proper decisions.

Here’s a toy example: A doctor hired through Mercor will create a fact pattern around a patient and then provide a grading rubric for how the model should assess the right decision to make in regards to the patient.

Privately, there is some mounting frustration with the current expert marketplace approach. It’s certainly beneficial for improving models, but there are a wide range of issues with these.

1. This decisioning data isn’t necessarily rooted in economically valuable outcomes. They are derived somewhat synthetically where the outcome is derived from what an expert determines the proper outcome to be. This can be useful in certain domains, but often both the decisioning logic + the tasks are somewhat detached from the relevant domain.

2. Attaching outcomes to tasks is quite difficult. Most of the long-horizon tasks we are wanting to target for training agents are… long-horizon. Modeling the decisioning pattern for tasks that can take weeks or months to resolve can be very difficult.

3. The rubrics are usually somewhat insufficient without intense expert time. The best rubrics occupy hundreds of discrete line items with different rewards attached. This is hard to scale in any labor marketplace format.

What’s interesting is that the data vendors as a result have begun to verticalize themselves. Rising data vendors now may only focus on finance or law and seek to capture key data from their domain in a couple core formats:

1. Raw, unadulterated data - There’s a move to buy/license more human generated data directly from end businesses.

2. Decisioning data: Approaches differ here between full scale expers and ops vs. other approaches using experts to help define what I’m calling the “ontology of the decision” to then help autogenerate rubrics and grading for discrete instantiations of the task

3. Outcomes data: Attaching outcomes to completed work artifacts often over a longer period of time to give models a verified outcome (and thus verifiable reward).

For instance in coding: a common tactic is to buy Github Repos (raw unadulterated data), mine PRs (where decisioning data lives), to then create training environments where decisions around a codebase can be measured, remixed, and then verified.

The net result is the new crop of companies that have focused on the data layer and preparing these data recipes and ontologies are now well positioned to become the product layer around their insights.

In other words, given the failure mode of current expert generated data sourced from labor marketplaces, both labs and enterprises are much more cognizant that data recipes and decision ontologies need to be rethought and sourced from underlying businesses.

As one founder recently put it to me:

The training recipes are more or less commoditized. People talk about all these [algorithms] Whatever… if you train long enough. You’ll get the same result. They’re all kind of the same thing. And these are like optimizations at the margin.

However, if you change your data set, you’ll have a massive difference in outcome.

To date, most of this work is being done in the most economically valuable verticals: law, finance, and healthcare, and customer support.

But if the data recipes, decision ontologies, and subsequent procurement of data are where all the value is going to accrue for both domain-specific models and foundation labs, we can quickly expect other domains to become contested spaces for data recipe superiority and thus model superiority.

Moats will accrue to the companies who think seriously about the ideal data recipes and decision ontologies for their vertical. While it won’t make sense for every particular vertical or every particular task distribution to have a small model trained, you’ll note that these decision ontologies useful for post-training models share a lot of similarities with the context graphs that Jaya and others have begun to write about.

Towards Systems of Training:

I have a hunch that the companies who model themselves as an applied system for producing the data recipes and decision logic for training models will ultimately accrue far more value than those who model themselves more as product-centric companies focused on the UX around AI for a vertical.

Read more

Today, an overview of 2025 and where I think we are headed in 2026.

Subscribe now

My 2025 Chain of Thought: The Age of Dichotomies

Model Progress

Vibes around current gen models are all over the place. Some point to Opus 4.5 being AGI, users on LMArena report that GPT 5.2 is worse than most model releases from the year.1

Ilya reports that the Age of Scaling is over and we are in a new Age of Research. At the same time, scaling models for economically valuable tasks is still in the early innings. The abbreviated lesson of 2025 is that if a) we can obtain reasoning data and b) accompanying evaluations for models that can c) act in economically valuable environments, we can hillclimb both large models and small models on these tasks and capabilities.

What’s the actual number of occupations and economically valuable work that we have catalogued with sufficient nuance for training? My back of the napkin math is that we have probably catalogued with sufficient rigor less than 2% of the labor force and likely less than 5% of contributing GDP?

With current model architectures and training techniques, it’s somewhat rational to believe that we can hillclimb nearly every occupation at least into a heavily AI-assisted occupation with current technology. If we use software engineering as emblematic of what’s possible, we would be looking at expected productivity gains of 50-70% across the entire economy even without continual learning. If some new architecture paradigm cracks problems like continual learning? We are now talking about wholesale replacement.

From this perspective, the last remaining challenge in AI’s proliferation in the economy is getting useful training data. At the end of the day, it’s simply a data acquisition problem.

But how long does that take? It has taken us 9-12 months to get models to perform well on a narrow slice of healthcare, finance, and legal problems. Will the data acquisition rate and learning curves be exponential? Or is this a far more difficult problem than scaling compute?

Markets

In the midst of all of this, we have a venture market that is now potentially frothier than 2021, with most valuations being only rational if you believe we are in the strike zone of imminent wholesale labor automation or if vertical monopolies form.

I take it as somewhat granted that if you can wholly automate a profession or task capability, you can charge close to the average rate of labor for that automation. If you can only assist the labor, you can charge near the going rate of inference. Therein lies the dilemma for companies who even if they are 80% margin businesses still are a ways off from valuations that imply the wholesale automation of labor. You can only ride RAG techniques and base model capabilities so far.

Anecdotally, while everyone chases the explosive growth of Harvey, Abridge, and OpenEvidence, there’s been somewhat less exponential growth in many vertical AI solutions vs. 2024. There are plenty of vendors who are now hovering around the $20-30m mark but are bottlenecked by the sort of foundation model progress or smaller model progress on their end that would lead to far better value capture.

Public markets continue to bifurcate between AI tailwinds and everything else. Many investors are now publicly talking about the continued endurance of systems of record and yet growth in public software looks stalled at around 15%.

Rollups continue, but attention is now shifting to bigger targets. Some huge takeouts are beginning to be talked about. EA may be the canary in the coal mine for what’s to come in 2026. Behind the scenes, there’s growing acceptance that AI workflows with current gen models are insufficient for getting to the Rule of 60, and yet increasing confidence that RL techniques can actually achieve fundamentally better cost structures in these businesses.

Feeling the AGI

Current model sentiment is heavily indexed against coding. Claude Code has mostly solved bespoke application development. One of my friends owns an SMB with nearly no automation much less frontier AI tech in the business. Within two afternoons of Claude Code + doing some systems design, I think we have pretty much transformed his business going into 2026 and I’m fairly bullish on growing it close to 3x next year as an AI-enabled service.

It’s legitimately breathtaking what’s possible with current gen models. And yet, 98% of occupations remain entirely inaccessible to models from a capability perspective. 95%+ of GDP is untouched.

And so in my view, 2025 looked like an Age of Dichotomies stemming from the jaggedness of the data (to use Leandro’s turn of phrase) and the amount of human intervention required in agentic workflows. As good as models are, they’re still immensely dependent on humans as shepherds. In my view, 2026 will be marked as the Age of Deployment in the pursuit of smoothing out the jagged AI experience across the economy.

So some predictions. Let’s hope I’m in the right error bars.

2026 Predictions:

Video Gen x World Models ripen for production and serve as the next critical product wedge.

Video gen currently reminds me a ton of image generation in 2023 where it needed to get about 30% better to be truly transformative.

My baseline expectation is that is about to happen for video generation and once it does, it unlocks similar opportunties in GTM that voice AI did for a number of service industries. World models feel more nascnet, but then again within the past month, we are now able to build pretty accurate sims and new worlds.

If they get 70% better, a whole new host of applications for physical world companies get unlocked instantly.

One example with better world models?

Walkable floor plans but within a VR world model of the home.2

I’m fully expecting video generation to be the 2026 rendition of voice agents. Everyone crowds in with vertical applications and a whole host of industries begin to change.

Systems of Record have new life

The endurance of systems of record has been remarkable. More and more they look like true winners from AI as opposed to losers.

That said, there’s still remarkable few AI products being built by vertical systems of record. There’s the occasional chat interface or RAG search, but this has been limited.

I expect that to change over the next year as systems start to meaningfully consider the data they’re sitting on, smaller model training becomes feasible, and shipped products like Shopify’s SimGym become emblematic of what’s possible.

Small models proliferate

Swe-grep, Applied Compute, the recent acquisition of Parsed, EvenUp’s PiAI all point to a world where vertical applications and enterprises begin to build smaller task-specific models for work they care about. Distillation techniques, infra like Tinker from Thinking Machines, open source libraries like Prime Intellect, all point to a different cost profile for building proprietary models. Many of these models will have durable moats for a number of years. Building a muscle around novel model training will be a prerequisite for existing systems of record looking for an uplift in their valuation.

Terminal UI design starts to bleed into knowledge work.

Paul, CEO of BrowserBase, has a fantastic example of a terminal UI he put together here.

If you want to simplify this: I think the personal compute environments for knowledge workers get more personalized.

The terminal with Claude Code and Skills is so useful that whomever can achieve the functionality of giving knowledge workers personal compute experience tied to the terminal probably ends up with some of the most novel training data there is. Models think in file systems, API calls, context, etc. UI is going to start to match that realization for knowledge work.

I can imagine different forms of this business: everything from boutiques that are centered around customizing personal compute environments for professions to potentially Lovable style opportunities that focus on helping knowledge workers build their ideal personal compute environment.

Training data spend goes to 15B a year

The big lesson of 2025 was to simply bring out of distribution data into distribution. This will becomes an even bigger emphasis for labs and researchers. The industrialization of training data will ensure.3

Evals for Economically Useful Work Proliferate

Rubrics, reasoning problems, in-domain tools and workflows, all end up becoming points of emphasis for labs and application companies. Vertical AI companies will begin to publicly put out benchmarks as they become more research-centric organizations as they too realize that revenue is directly tied to the task capabilities of their models.

I’m also frankly predicting this because it’s something I’m actively working on and we are partnering with some fun companies to do. Reach out if you’re interested.

Somebody builds a gaming support studio as an AI-enabled service

Gaming studios have been bottlenecked on adopting AI because of revolt inside their developer core. The natural play will be for execs to outsource even more tasks to support studios leveraging AI for QA, design, console migrations, voice work, etc to give air cover but get productivity gains.

Support studios take no IP risk and so a business here could fully focus on AI transforming the unit economics of the services. I think this is one of the best opportunities out there if you have the right connections into gaming studios and I fully expect someone to give it a go.

Compute Markets get Weirder:

Smaller models proliferate, spot markets, and derivatives are now out in the wild and fully utilized, and every significant AI team starts building out hedging arms. By end of 2026, at least one significant trading desk is indicating that they are building out a compute desk.

Overall

In my view, the Age of Deployment beats the Age of Research as a descriptor because deployment is one of the only scalable ways to obtain the data necessary for continued model improvement. Much of the economy is still not instrumented for models. So 2026 overall will look less like a single intelligence explosion and more like a long, uneven rollout as teams continue to deploy but with more of an eye towards advancing model capabilities through their deployments vs. commercializing some latent ability models already have.

At the same time, 2026 will also see the rise of AI veganism. Model progress has coincided with tech exhaustion amongst the general populace. Life is more expensive, the American Dream feels more contested, and junior jobs are becoming automated. All of that will lead to a growing set of opportunities in the physical world that has nothing to do with AI. And I think it’s probably best that these opportunities remain illegible. After all, if they’re on the internet, the models will soon know about them. For now, they should remain secrets.

Subscribe now

There’s a chance this week will be looked back on as one of the most important weeks in vertical software. We got more tractable signals on the economic value of base model improvements and we saw what might be the coolest AI product release of the year with big ramifications for how vertical software will innovate going forward.

So let’s dive in:

GPT-5.2 is crushing the evals, including GDPVal

The current model trope is that they crush on evals across math, science, and coding and then are dramatically less powerful in real knowledge work situations.

OpenAI has been hard at work here creating GDPval as an index of economically valuable capabilities and then training models around knowledge work problems that matter: financial modeling, nursing scenarios, and more.

Based on the reported results, GPT-5.2 looks like one of the first models that, given the right context, performs meaningfully better out of the box on real knowledge-work tasks.

In my view this indicates two things:

1. If you can define the necessary knowledge work pattern, labs and service providers essentially have zero limitations on getting models to solve the knowledge work capability gap through a combination of mid-training and post-training techniques.

2. It also highlights the continuing importance of AI operating inside of systems and databases. Ultimately, base model improvements are incredibly desirable from the perspective of vertical software. And further base model improvements are likely to imply far more partnerships with vertical software companies and vertical AI companies in order to achieve more economically useful activity and then distribute it.

In short, the natural deployment path for AI in the majority of industries lies within systems of record and systems of action.

There’s much more that could be said here, but I’ll leave it for next week when we do predictions for 2026.

Subscribe now

AI penetration in vertical and functional categories is still nascent

A couple graphs for you:

Menlo catalogued functional and vertical revenues.

OpenAI released their report on the State of Enterprise AI.

Can we use these two graphs to determine some interesting takeaways around where dollars and builders will be headed? Likely yes.

First, manufacturing looks like an industry ready for some breakout vertical AI solutions. Despite absolute dollars being low, the number of interactions with models inside the industry is growing well above the median.

Second, construction has meaningful interest from enterprise customers, but both the dollars being spent and the value being realized are low across the industry.

In short if we measure the labor spend in nearly all these industries vs. the current AI spend, it’s appropriate to say that we are still in Year Zero for AI adoption in the vast majority of industries even with extraordinary growth rates. From a demand perspective, this doesn’t look bubbly, especially with models continuing to steadily improve.

Is Vertical Software Undervalued?

Meritech put together a 2025 Markets Update highlighting the emerging gap between software winners and losers in the AI world.

I’ll highlight some slides here:

In the public markets, AI winners are growing market cap at extraordinary rates.

Public SaaS is hovering around 15% growth rates.

AI Tailwinds are everything. Companies with AI tailwinds are growing faster, with higher multiples, and greater market caps.

The Labor Thesis is still present.

Lastly, in the private markets, it’s AI or bust.

Does this spread between SaaS companies and AI valuations open up arbitrage opportunities?

Jesse Tinsley, acquirer of Mainstreet and other software products, thinks the answer is yes.

Everyone is talking about AI saying SaaS is dead is well… dead wrong…
The Biggest winners will take Legacy SaaS companies and turn them into AI first companies.

Why legacy SaaS?
1) niche training data
2) refactor everything
3) capture instant distribution (revenue and customers)
4) skip 4-5 years to build and capture market share

The pace paradox suggests buying vs building right now and the tradeoff isnt even close.

I largely agree. Again, we will leave it for 2026 predictions, but I think this is going to become one of the largest stories of 2026: the realization that existing vertical software companies and systems of record more broadly are not going away.

SimGym

Shopify launched my favorite product of the week: SimGym.

It’s got an amazing core idea. Ecommerce companies optimize for conversion rates. That has led to an entire industry of companies providing A/B optimization tests for websites; literally going pixel by pixel to figure out what leads to a 0.2% conversion uplift.

The idea behind SimGym: run those simulations via AI agents that can provide instantaneous feedback behind what leads to a conversion.

I already love this. But think through what this really implies: Shopify just created an RL environment for every single site across its merchant base to train new agents on. You can imagine further models being trained to actively optimize from real user interactions, an autopilot that automatically deploys better sites that lead to more conversions, etc.

That should translate into more consumer spend which Shopify can monetize through payments, but it also should equal more subscription revenue as firms already pay thousands for this sort of service.

Underlying all of this is the millions of user interactions and other data sets that Shopify already has as a system of record. That itself is data that any individual store can’t obtain. Unless you’re a giant enterprise doing hundreds of millions in sales, you simply don’t have the data to meaningfully optimize. And so not only is this just an insanely cool product, but it democratizes state of the art A/B testing for the entire industry.

And I think, we are getting closer to understanding why systems of record are about to have their second act. Every single vertical software company will become a training gym for agents that meaningfully expand their addressable software market. And because they can aggregate and have already aggregated necessary data across the entire industry, they are about to experience a true second act as this very data becomes necessary for creating novel agent paradigms.

The TAM remains all of Human Labor.

That’s all for this week. Have a great weekend!

Today, we’ve got a roundup of the three biggest stories of the week:

Subscribe now

Thrive Holdings partners with OpenAI

This is a killer partnership for both sides of the deal. Labs need data, industry environments, and human feedback and AI-rollups need talent for model improvement, strategy, and more.

That said, I think you could draw an inference that everyone’s timelines on AI development should be extended and this deal brings into question most rollup strategies with generative AI.

To put it succinctly, if you need OpenAI to be on your cap table and you need premier AI talent in order to get meaningful unit economics improvement inside these businesses, the costs associated with that talent inside the business are going to skyrocket.

At some point if it’s truly this difficult to get industry-specific agents to work, why own the underlying businesses vs. pioneer an agent lab that works to assemble the data, model finetuning, and human feedback without the drag from operating the existing businesses on top? There’s signs that real moats for agent labs are developing due to this complexity.

That said, you could also paint a picture of a compounding advantage in any rollup strategy that features OpenAI on the cap table. Maybe it’s difficult to implement. Maybe it’s difficult to get agents to directly improve accounting workflows outside of document extraction. But if it’s difficult for me and I’ve got OpenAI on my cap table and engineering talent out the wazoo, then it’s going to be nigh impossible for others.1 And that certainly feels like a moat.

Norm Law: Software then Services

Norm.ai recently announced the creation of Norm Law alongside a $50m investment from Blackstone. Norm.ai has gotten its start developing compliance agents for some of the most important financial institutions across the industry.

Now, they’re expanding from a more traditional SaaS play into providing legal services themselves. I’m very interested to see how this goes. In my view, it makes complete sense. Norm already internally has tons of legal talent, has a true software and AI advantage, and can likely monetize higher-value legal services than probably what they can obtain just from pure SaaS revenue (even though those SaaS contracts are in the seven figures).

We’ve seen firms go from services to software. Are we going to see an advent of software companies becoming service companies as they accumulate a true data advantage and people advantage?

Unlimited announces a $12m Seed to build a vertically-integrated EPC

Unlimited came out of stealth this week to announce a new construction company. Importantly, they are an AI-enabled service around engineering and design services.

With the boom in data center investment, there’s billions of dollars in infrastructure projects up for grabs. Unlimited is betting that they can use engineering and design services as a way to grab a share of this pie.

Can better AI-empowered design services lead to a large vertically-integrated construction company?

Funny you should ask! About a month ago, I asked the same question to a senior exec at a large EPC. My hunch was and still is that there’s a path from design services to do this sort of vertical integration.

His take? There’s certainly something there, but unless you’re solving the massive labor shortage in construction as part of it, it’s really tough to solve and may be the wrong starting point. Effectively, labor is the constraint on EPCs and projects. There’s billions in contracts that they can’t bid because there’s simply not enough construction labor.

But solve the labor issue? There’s billions in revenues up for grabs.

And I think there’s some really great strategies for an enterprising company that wanted to solve this sort of labor coordination around the EPC.

Around the Ecosystem:

• A must-read piece on Epic Systems.

• A vision for Healthcare AI from 8VC

That’s all for this week. Next week, a deeper dive into healthcare AI and a Verticalized report for paid subscribers on Recommerce Trends.

Thanks again for reading!

Subscribe now

Incredibly fun to sit down with Anna Berger, CEO of Trayd, to talk about construction payroll, job costing, and why there is still so much software left to build in 2025.

On paper, we live in an era of instant payments, modern payroll platforms, and a full stack of fintech primitives. In that world, you would assume payment delays are largely solved.

They are not.

The construction payment chain is still onerous, multi party, and frequently delayed. Roughly a quarter of GCs list payment delays as a major cause of project delays. A single late payment can add anywhere between one and three weeks to a job.

Underneath that sits the real control point: Labor.

Construction only happens if people get paid. As Anna likes to say, about seventy five percent of a contractor’s back office cost structure is payroll.

Once you hear that, payroll stops looking like a boring back office workflow. It starts to look like the singular control surface for most of the job’s economics. Whether people get paid on time, at the right rate, with the right benefits and the right compliance paperwork attached, determines whether the project keeps moving or quietly stalls.

The interesting question is what happens if you treat payroll not just as a cost to manage, but as the core data feed for job costing and bidding.

What if payroll is not only the gate on whether work continues, but the primary way to prevent late payments and blown jobs upstream? What if better labor analytics and real time tracking of labor spend can drive better bids and healthier margins for contractors?

Turns out that’s exactly what Trayd is doing.

The founding: starting at the hardest point

Anna did not wander into construction payroll due to a top-down market analysis.

She grew up in a construction family. She saw the stress of making payroll. She watched the guesswork around labor forecasting. She saw how hard it was to track the true cost of variable labor on real projects.

And despite it being fairly unsexy and plenty of naysayers, Anna and the team went to work.

For the first few years of Trayd, her job was not glamorous. She spent her time convincing contractors, unions, and back office teams that a startup could handle their most sensitive, heavily regulated workflow.

Two early product bets stand out.

First, Trayd built for workers and the complexity of making sure they were were paid fairly.

That meant embracing the full mess of construction payroll. Automated union deductions. Complex overtime and double time rules. Shift differentials. Benefit calculations that change by trade and geography. All the details that generic payroll pushes into a catchall “custom fields” bucket became first class objects in Trayd’s system.

Second, Trayd focused in on helping smaller and mid sized contractors compete with the back offices of national construction groups.

Federal and state infrastructure projects are booming. Bridges, tunnels, airports, offshore wind, stadiums, transit. These come with their own rulebook: prevailing wage requirements, certified payroll reports, and constant regulatory change.

For a lean back office, that is overwhelming. Trayd’s data model and workflows are tuned for this world. The system knows how to calculate prevailing wage, generate certified payroll, and keep up with an ever shifting compliance landscape. That lets contractors bid on projects that would previously feel out of reach or result in scaling back office complexity.

Those two bets are now compounding.

Trayd is growing roughly 500 percent year over year. And their customers see field teams grow on average around 265 percent after adopting Trayd while back office headcount stays flat.

A modern, construction specific payroll platform is already a solid business. The more interesting part is the wedge it creates. Once you own payroll, you are standing in the perfect spot to attack everything upstream of it: job costing, bidding, and project analytics.

Job costing: materials, labor, and where the real risk lives

If you think about construction naively, you can put a project into two buckets.

Materials.
Labor.

If you want to forecast the cost of a job or set a bid that does not blow you up later, you need to get both roughly right.

The industry has put a lot of energy into materials. Takeoffs, procurement, dynamic pricing, inventory. None of that is easy, but at least it feels countable.

Labor is where the real chaos lives.

Hourly rates per trade. Union contract components. Project type. Shift timing. Location. Benefits. Travel. Per diems. Hazard pay. A long list of rate modifiers that can all apply to a single shift.

Historically, nobody has modeled this cleanly at the level of an actual job. Horizontal payroll thinks in terms of a company, not a project. It runs biweekly or weekly cycles. The primary object is the employee, not the job site.

In construction, that is the wrong abstraction.

A contractor does not live inside “payroll for ACME Builders.” They live inside “this stadium expansion” or “this terminal roof” and whether those specific projects lose money or not.

Two bidders can assume nearly identical material costs and still have wildly different outcomes. The difference shows up in how labor actually plays out in the field. Did that night work run longer than expected. Did the mix of trades on a certain phase drive the effective hourly rate much higher than the estimate. Did overtime spiral out of control.

The firms that survive are usually not the ones squeezing an extra few percent out of materials pricing. They are the ones that understand the labor side of the equation in a grounded way.

That blind spot is the zone Trayd works in.

Trayd’s data model: labor as the core primitive

Matt Brown has a line I love. Data models are destiny.

Matt Brown's Notes

Your data model is your destiny

Product market fit is the startup holy grail. “Product” and “market” are essential, but a startup’s data model is the dark matter that holds them together…

Read more

7 months ago · Matt Brown

Vertical tools serve specific industries with deep domain complexity. Leverage comes from what you choose to emphasize. Toast elevated menu items—not transactions—with prep times and kitchen routing as first-class data. Klaviyo promoted order data to equal status alongside email metrics.

Vertical tools are really just strong opinions about what should be first class in a system.Trayd applies that same philosophy to construction payroll.

Instead of treating “hours” as a generic field to multiply by a base rate, Trayd makes the full complexity of hourly work the center of the data model and ties it directly to the job site. Every question around who worked on what project, under what contract, with which rate modifiers becomes answerable.

And at that point, Trayd became a live data feed for construction labor.

From there, the analytics are straightforward to imagine. That stream becomes the foundation for real time job costing and for better future bids.

Every facet of the job site becomes queryable with the upside that the true cost of the job becomes verifiable.

And if you can determine the cost of a job, you can now forecast with greater precision than ever before.

This is where Trayd’s vertical depth and the true insight around focusing in on construction payroll really shows. It is not obvious at first glance that solving construction payroll compliance would unlock genuine labor analytics. But once you see it, you cannot unsee it.

The same data model that keeps contractors compliant on certified payroll and union rules is the one that gives you a truthful view into the labor portion of every job.

Real projects and an ode to traditional software

Today, Trayd is running payroll for roofing crews at JFK Airport, crews on the new Buffalo Bills stadium, offshore windmill rigs, and plenty more.

People are clocking in and out on some of the most visible projects in the country and Trayd is the system that decides whether they get paid correctly and on time. The same system is quietly building a ledger of what those projects truly cost on the labor side. And again, prior to Trayd, there just wasn’t a truly great system of record and data model for the complexity here.

Which brings me to a broader point.

We are clearly in an AI era. Foundation models are getting better. Every week, there is a new demo of an autonomous workflow that looks and feels impressive. I am very long this shift.

At the same time, my thesis is that we are still going to see a large number of multibillion dollar vertical software vendors created in very traditional ways.

There is an enormous amount of software left to build.

There are industries where systems of record are weak, fragmented, or non existent. The database is still an Excel file, a filing cabinet, and a payroll export. In those environments, AI agents do not have anything solid to stand on. They cannot plan, price, or automate the way we want them to, because nobody has captured the underlying reality in a structured system.

Trayd is a reminder that the first job is often to build that system.

Once you do that, everything else becomes possible. Analytics. Better bidding. Smarter planning. Eventually, more automation. But it all starts with boring, serious software that earns the right to be in the loop.

I think of this as an ode to traditional software.

A focused product that picks an unglamorous but vital workflow in a huge industry. A data model that actually matches how that industry works. A founder who has been talking about the same problem for four years and is more excited now because the market has finally caught up.

Construction work stops when people do not get paid.

Trayd started there. It built upward from that simple insight into a real system of record for labor. Along the way, it turned payroll into a live readout of what is happening on job sites and gave contractors a clearer view of what their work costs in the real world.

In a world fixated on what agents might do someday, I think it is worth paying attention to the teams quietly rebuilding the core software those agents will eventually depend on. Trayd is one of them.

Subscribe now

Let’s get to it shall we?

Today, an overview of ideas in AI I’m thinking through in relation to vertical AI.

RL partnerships could be a defining strategy for vertical SaaS companies shortly

Vertical SaaS companies are some of the largest benefactors from Reinforcement Learning becoming the state of the art approach for applied agents.

There’s a whole litany of companies and strategies across the stack here, but generally everyone agrees on a couple core ideas:

1. Agents need to train in realistic environments with rewards and verifiable signals for their actions.

2. Regardless of whether RL fully generalizes, it presents a cost-effective path to vertical or function-specific agents. The labs are betting that RL does somewhat generalize but even in the worst case scenario, they end up with highly adept agents for particular knowledge work functions like financial analysis, healthcare work, etc.

3. Enterprises are also placing bets here due to the reduced cost of RL training and the corresponding impact on their operations, often favoring partnerships with RLaaS vendors to train models on their unique operations.

SMBs are never going to stand up reinforcement learning on their business operations. They’re going to simply rely upon vertical vendors to perform this necessary function for them.

Likewise, its borderline uneconomical for the labs to build RL environments and recruit the experts for labeling data in every vertical.

In my view, these two insights position vertical SaaS companies as the RL power brokers in their verticals. They are the beneficiaries from foundation models needing RL environments possibly leading to partnerships with vertical SaaS vendors. And they are the distribution mechanism for RL-trained models specific for their vertical.

Which means the real leverage sits with the systems of record that already intermediate these workflows.

Incentives are more aligned with the labs than they are with most vertical AI startups. Systems of record benefit when their software becomes a core part of the benchmark, and when agents are trained to treat their application as the default surface for getting work done.

They have already assembled the distribution, and more importantly, the expert networks who actually know what “good” looks like in that vertical. Those same experts are the ones who can label and define proper evals.

Since state of the art agents are still sub optimal for nearly every industry’s operations, whoever first wires their workflows into repeatable RL loops will have real power. If your software is the environment, then your customer base becomes the training distribution. That is a very different position than simply “adding an AI copilot.”

I do not see any indicators that systems of record are under siege from foundational model providers. The pressure is hitting the vertical agent companies instead. Labs are actively searching for ways to produce evals and agents that can properly navigate traditional web applications and existing SaaS. That points them toward platforms that already have stable usage and structured workflows.

The cost of RL is nowhere near the cost of pre training. With potential subsidies from labs in exchange for access to RL environments, the economics can become even more attractive. You can imagine co funded programs where the lab underwrites runs and the SaaS vendor provides stable environments plus expert reviewers.

Even so, while RL training costs are lower than pre training, human data labeling costs are climbing, and there is every reason to believe this remains a bottleneck for model progress.

Daniel’s Substack

Human Data is (Probably) More Expensive Than Compute for Training Frontier LLMs

Post-training techniques (e.g., supervised fine-tuning and reinforcement learning with verifiable rewards) are crucial to recent advances in LLMs. Unlike pre-training, post-training relies heavily on annotated data provided by humans, often requiring expert input…

Read more

9 months ago · 17 likes · 2 comments · Daniel Kang

Vertical SaaS companies may be well positioned to have lower cost methods to produce human experts through existing distribution, real insight into which users take what sorts of actions inside of software. And of course they’ll be getting paid handsomely for enabling these sorts of expert marketplaces.

In short, the same insight that drove embedded initiatives inside vertical SaaS may end up driving the frontier of RL environments, training, and RL runs. If there is a lower cost way to procure the ingredients that agents require to improve, and if those ingredients sit inside of system of records already, it becomes rational to treat RL partnerships as both an R&D priority as well as a revenue line.

There are clear paths to revenue via end agent releases for customers and training.

Of course, this hinges on having the machine learning expertise and ability to recruit human annotators for the required vertical functions.

But doesn’t this as a result sort of look like other embedded opportunities? Vertical SaaS companies didn’t want to become payments, lending, or banking experts. They instead chose to partner with preferred experts across the stack to expose these products to their customers. I wonder if a similar insight will drive RL partnership initiatives with of vertical SaaS companies to deliver RL environments, expert networks, and more.

In fact, I think there’s clear reasons why this will end up being a company.

1. Labs don’t want to have to procure environments from tons of individual vendors. They want their procurement costs to be low, want little variability in training across RL environments, and likewise want Mercor-like simplicity in recruiting experts and knowing the end data that they’re getting back.

2. For vertical SaaS, the majority do not want to in-house a research group. They want to participate (greatly) in the upside for successful RL runs, but incur as little opex as possible.

All of these look very similar to the reasons why embedded partnerships took off across financial products inside vertical SaaS. We may be entering an era where the same holds true for frontier AI.

If you know any RL engineers exploring this, I would love to chat with them.

As an aside, I am thinking about hosting an “RL for vertical SaaS” meetup in San Francisco. If you are interested, let me know.

Systems of Action are judged by revenue growth, not cost optimization

The winners inside vertical AI so far are primarily centered around revenue uplift.

This will likely continue for the next few years, at minimum.

There is no need to make this overly elaborate. If AI usage still requires a human in the loop, and it clearly does, then the most successful AI implementations will enable revenue growth through more efficient use of human labor.

This logic has underwritten the success of EvenUp and related companies that convert labor bottlenecks around revenue into far more digestible services and automations.

AI enabled services are thus far more compelling narratives and more robust companies than AI enabled rollups. Rollups mostly have bought existing distribution and then concentrated on opex reduction. AI enabled services have focused on reimagining the service itself and when successful grow organically in correlation with their own end customer’s growth.

A couple examples:

Crosby AI lets customers move faster on MSAs, NDAs, and contract reviews; all bottlenecks inside of a deal.

Owner for restaurants is totally focused on the revenue impact from their AI and software.

And lastly Endeavor focuses almost exlusively on the revenue generating functions for manufacturers leading to more revenue and faster revenue recognition through sales order automation.

In short, the leading AI companies are synonymous with revenue growth for their clients. The more intimately tied an AI product or service is with revenue growth, the more likely, it’s going to find incredible product-market fit.

Subscribe now

Harvey’s CEO sat down for a chat around legal AI and the application layer more generally. The whole thing is valuable, but for our purposes you really only need the first 30 seconds. (the full discussion is at 46:00):

OpenAI is going to compete with vertical AI. They’ve hired investment bankers to build better finance models and there are certainly more verticals on the way.

If the safest posture is to assume foundation models are going to build into your vertical, what then is the proper posture for vertical companies innovating?

Weinberg seems to think it’s to anchor around problems that require systems vs. ultimately commodifiable LLM prompts.

But I also think it paints a different case for companies that already occupy the system of record for an industry. Systems of record are not disruptable. They’re in some sense the largest beneficiaries from AI becoming more performant for their domain. Epic benefits from OpenAI wanting to create better healthcare agents. Coupa and Ramp as well.

What are the second order effects from labs wanting to train better models? I think it’s beneficial to current systems of record. More soon.

Around the Ecosystem:

• Rebel raised a 25m Series B to build a recommerce network. Reverse logistics/recommerce has fascinated me lately. I’ll be releasing the first Verticalized report on the industry next Thursday for paid subscribers.

• Are MSOs finally entering the legal profession? MSOs in physician practices have been one of the big stories in PE over the past 10 years. If that happens in legal, what does that mean for legal AI tech or new business models that can be developed and commercialized?

• Prediction Markets are on the rise and in the past week have gained immense institutional respect from media and finance. Event Horizon has done a series of breakdowns on where volume lies, the partnerships, and more. The newsletter is well worth a read!

Event Horizon

Kalshi Hits No. 1 In Finance Apps, But Volume Is Down (Even With PrizePicks)

I’m going to give you a couple of facts, and then ask a question…

Read more

6 months ago · 1 like · Dustin Gouker

• Sean Cai writes about the history of RL and the current moment. More to come next week on this topic. I think it’s worth deeply considering the game theory involved around the need for RL environments in myriad verticals.

• Tidemark released the 2025 Benchmark Report for Vertical SaaS:

  

This is the big takeaway in my view. Vertical SaaS adoption of AI is accelerating. What does that mean for incumbent vs. new AI companies in these industries?

Subscribe now

But let’s not bury the lede, I’m deeply skeptical that to date real advances in end-to-end automation with agents have been created.

These are important productivity advancements that have resulted in time compression. But it remains unclear if these are enough to underwrite true transformative gains vs. begin to rely on more traditional rollup tactics.

So today, let’s explore the dynamics behind the AI-enabled rollups.

Rule of 60

One of the primary backers of AI-enabled rollup initiatives, General Catalyst, defines a new Rule of 60 for AI-driven M&A:

The SaaS industry’s Rule of 40 reflects the inevitable tradeoff between growth and profitability. We believe AI-enabled roll-ups can eliminate this tradeoff, achieving higher combined growth and profitability through efficiency gains that yield software-level margins, immediate inorganic growth from acquisitions, and expanded outcome capabilities that create new markets. The companies we back aim to take businesses growing in single digits to 10-20% growth through a combination of organic capability-led growth and acquisitions. They simultaneously aim to double profit margins, often targeting 30-40% margins. They are setting their sights on a new Rule of 60 standard.

If the Rule of 60 underlies the expected return and therefore the multiples of these businesses, are we on the path with current AI tech to deliver Rule of 60 AI-enabled rollups?

Well to know that, we would want some further clarity on Rule of 60 dynamics:

1. Is it inclusive of R&D costs?

2. If not, is there a market expectation that AI technology will somewhat take care of the devops, maintenance, and agentic automation build going forward?

3. Are we primarily looking at the margins of the individual acquisitions or the parent company?

4. Inorganic growth and organic growth have different implications for what needs to be accomplished with automation.

The Bull Case

You can imagine the bull case through a toy example like the following:

I buy an asset at 3m in revenue with 15% EBITDA. I buy $450,000 in profits

I subsequently double EBITDA to 30% and grow each business by 20% YoY, obtaining $1.08m in profits in the next year.

I’ve added $630,000 in profits, and will continue to grow bottom line every year.

In this example, I’m not even overly reliant on end-to-end automation.

If my operational costs are fixed, and I increase the throughput of the underlying business via AI and technology, I can keep the same headcount and support that extra 20% growth and achieve 30% EBITDA.1

The bull case on the Rule of 60 is essentially the following: these businesses can grow at 20% and keep the same labor force in place but with enhanced throughput to achieve higher EBITDA.

Theoretically, I could achieve this via inorganic growth as well. Once I have a core labor force, I could imagine utilizing that labor force on new acquisitions and part ways with personnel previously managing the new acquisition.

We will circle back to this, because this is where I think it gets a bit more fraught; especially because of the sensitivity here to acquisition multiples and cost of capital

In short, the Rule of 60 stems from either:

1. Organic growth with higher throughput of existing workforces

2. Inorganic growth with the replacement of workforces with a higher efficiency one

So then we have our assumptions and frankly our target sectors for AI-enabled rollups:

1. Higher throughput must be achievable

2. Once high throughput is achieved, organic growth or inorganic growth are vastly more important than further labor automation.

   1. Inorganic growth must be based around the core ability to use my higher throughput operations.

   2. Organic growth solves everything

If AI doesn’t substantially impact the throughput or the organic growth profile, I think we are essentially doing private equity.

One more question

I take it as given that gifted engineers and M&A types have been making real throughput gains. AI-assisted code has increased their own velocity. Internal tooling, data, retrieval has reduced the time spent on parsing deals.

But let’s be honest: these people cost a lot. Is Rule of 60 including the added costs of engineering and M&A talent?

Will you be able to keep these organizations lean if you have to serially acquire assets across a vertical? It’s too early to tell, but is an important part of the achievability of the Rule of 60.

Tells here are going to be:

• The amortization of the same engineering work across many entities

• Compounding savings

Organic growth once again cures a lot. In fact, it’s really the essential quality in making the inorganic growth and engineering components work.

What’s working, what isn’t

If there are tells that an AI-enabled rollup is going well according to Rule of 60, there’s corresponding tells when it isn’t.

If a business is struggling to achieve Rule of 60, we would expect to see:

• Labor substitution: If I can’t achieve enhanced throughput via AI, I have to likely obtain cheaper labor to achieve Rule of 60.

• Engineering and R&D rising as a percentage of revenue.

• Stalling organic growth and increased M&A activity2

If these are the primary levers at work, then the story looks more like a classic private equity rollup, not a Rule of 60 AI platform.

In many cases, they will be able to. But vertical selection is everything.

Long Lake in HOA management (and other verticals) is crushing. They have been able to substantially increase throughput and grow organically via improved HOA operations with AI being a large driver.

Buena in Germany is a clear winner in this model and have been able to centrally manage property management functions via AI.

But other verticals are trickier. For instance, accounting looks far more difficult.

1. Accounting work is bespoke. It’s difficult to get throughput advances across a number of localized assets. That combination makes it much harder to achieve the “same headcount, more throughput” version of the bull case.

2. Organic growth is tough.

3. And so it looks like some form of labor substitution is happening.

But there may still be reasons to do rollups in these verticals.

For one, the inability to automate is at least some form of a moat. If AI timelines elongate, you will still be able to automate on longer timelines without risking the value of the work itself being commoditized to zero.

But then ultimately the question is simply: will AI-enabled rollups look more like really good private equity rollups? Or more like Rule of 60 companies that change a vertical forever?

Of course, game changing AI agents may change the math altogether. But is there evidence that this is happening? I think not.

Subscribe now

Today’s newsletter:

The primary challenge for every single technology company now is simultaneously disrupting some core area of an industry while navigating the constant disruption risk they too face from AI accelerating competition.

Everyone has a different interpretation of the game theory here which is why every day you can read a new article on “why vertical AI is dead on arrival” or “why vertical AI is better positioned than foundation models” or “why vertical SaaS is the biggest beneficiary of AI.”

My take? The truth is incredibly vertical and business model dependent.

But I do want to briefly sketch the capital frameworks that more investors are beginning to group their bets with.

Kingmaking

Harvey at $8B, OpenEvidence at $6B, EvenUp at $2B.

Financially speaking, crowning a vertical winner prior to mass market saturation is designed to give companies sufficient capital to guarantee their dominance in the vertical.

To customers, this signals “this company is here to stay and is safe to partner with for the long run”

For competitors, it’s meant to signal “don’t even bother trying to compete.”

These are very valuable messages to communicate when the cost of code has precipitously fallen. Communicating stability matters to companies when there’s 20 different companies in the vertical competing for their attention. Credibility is simply at a premium.

And by no means should we discount the user growth and revenue underpinning these businesses. These are rocketships in their own right, simply trading at multiples that are unfamiliar in vertical tech.

Put succintly: The kingmaking strategy is not reducible to “put in cash at high multiples.” There is a strategic bet that buying companies at higher multiples is a way to protect the investment as a whole.

Of course, competing with crowned kings is still a viable strategy. In the words of Scott Stevenson, founder of Spellbook Legal. The whole thread is worth a read, but some excerpts below:

Is cash a moat? It’s unclear. But there’s a case to be made that the cash itself is not the most relevant reason why companies are taking kingmaking rounds.

If kingmaking dynamics are going to incentivize somebody to take the money, it might as well be you. It’s much better to be the one driving up CAC to unsustainable levels than be the business that can’t sustain it.

Despite a lot of attention on the mega-rounds inside of vertical AI, it’s somewhat surprising that most verticals have not seen a kingmaking moment.

Insurance, logistics, construction, and plenty of others have no obvious winner yet or are not nearly as winner-take-all in their construction as VCs seem to expect medical and legal to be.

Will we see a venture firm look to solidify this strategy in every vertical? Or are other verticals less susceptible to kingmaking dynamics?

When disruption is priced in, durability is mispriced:

Marc Andreessen has said that a16z is “investing on the thesis that everything will be rebuilt because of AI.”

Rebuilding everything implies massive opportunity in disruption. When this becomes the market expectation, disruption starts to get priced at a higher multiple than assets who are favored for durability.

This of course was the original Constellation thesis. But a new crop of firms are starting to bet heavily on the durability of assets and their reinvention in the AI age.

Two have been the talk of the town this week: Bending Spoons and Beacon Software.

There’s many variants of this thesis, but I think you can distill this into one core idea: companies that survive and improve in the AI era will likely endure for 30-50 years.

In other words, it’s business kintsugi. Take pottery showing some cracks and fill in the cracks with gold.

These companies will long term be judged by their capital efficiency, the durability of their assets, and the transformation of the underlying companies.

And I think the real bet here: if your underlying companies can survive the next 10 years in the AI age, they likely are positioned to survive the next 50. What disrupts them once the very means of software production and nature of software itself has become intelligent?

This trade in software is particularly powerful because of the current perception of software today. Beacon, Bending Spoons, and I’m sure many others are stomaching the current disruption risk to ultimately take advantage of a long term durability multiple as they help these businesses navigate the next 10 year stretch.

Subscribe now

This is some of the most fun I’ve had writing a piece.

We need far more analysis around the unique ways companies in industries are going to use their inherent capital and technology advantages to transform industries.

As we leave the pure-play SaaS era, some of the most fantastic opportunities will demand new balance-sheet constructions, asset-based financing, and creative partnership structures.

Dealmaking is destiny.

Kobold Metals is one of the greatest dealmaking companies around.

Subscribe now

The defining feature of Kobold is its ability to exercise leverage across partnership and capital structures—unlocking the value of its technology to discover rich mineral deposits around the globe.

You can tell one story about Kobold as an AI company, but you can tell an even better one about Kobold as a company that exercises this leverage to solve the political and partnership ramifications of drilling for billions of dollars worth of rare earth minerals.

A refresh

Kobold Metals is a minerals exploration company that explores in areas where rare earth metals can be potentially located. That means lithium, that means cobalt, that means copper. That means an assortment of potential mineral interests that are necessary for the future of electrification and that come with massive geopolitical ramifications.

A primary issue is simply identifying sites that can support large-scale mining operations. While these minerals aren’t rare, they are difficult to find for a number of reasons. Most of the best active mines are in the Chinese’s orbit of influence. That geopolitical reality means much of the rare metal space is effectively closed off to U.S. companies.

And simply put, there’s no low-hanging fruit anymore. Kobold Metals came onto the scene with one critical idea: what if we could leverage AI and machine learning to identify the highest propensity drill sites for exploration, discover large deposits of critical minerals, and subsequently capture a large part of the upside in bringing these sites to production?

To do that, they have a two-pronged approach around their technology: TerraShed now hosts somewhere between 3% to 5% of all geological records across the globe.

The idea here is to literally extract decades old written copies of geologic reports, satellite imagery, and maps from places like Zambia into a unified database. It is no small feat to simply obtain this data and normalize it into a unified map.

Once they have this data, they then leverage Machine Prospector, their ML platform, to identify the highest propensity sites for potential mining activity. And then the fun ensues.

The State of Mining

The mining industry can essentially be broken down into three core categories, with, of course, some overlap.

1. At Scale Mining Companies: These have assets in production and usually also have projects at the exploration phase. These are mature businesses and well managed.

2. Junior Miners/Exploration Companies: These companies are typically speculative. They possess the rights to mineral interests in defined areas, but quite often have minimal funding to properly explore their claims. This is a huge problem. Theoretically, there are a large swath of minerals that are underground, but we wouldn’t have any idea due to the inadequacies of companies sitting on these claims with nearly nothing to show for it.

3. Governments: The greatest single blocker or aid in mining. Governments may lack technical savvy, but if you find a deposit worth anything, you better be ready to cut deals in order to get these minerals out of the ground.

And so, if the biggest blockers to new mining projects are, quite frankly, land speculators or governments who have been burnt before, the entire success of the business does not depend upon necessarily your technology.

It depends on how you can structure partnerships and incentives that enable you to quickly establish whether there is potential in a region, but not overcommit if there isn’t. The entire ballgame for success relies upon structuring these deals.

And so how do you get deals to the table? That is all about technology and financial leverage.

Technology + Capital Leverage

Let’s say you have two things in mining exploration that nobody else has: technology and a balance sheet worth billions.

What do you do? The pain points are twofold.

First, much of the land worth exploring already has a mineral claim on it - you need to get access. It’s time to cut deals.

Second, you can’t afford when obtaining this access to be in for a penny, in for a pound. You’re dealing with mining companies hungry for capital; that see your balance sheet and see opportunity.1 You need early opt-outs. You need the right to walk away, and you need majority upside with minimal downside.

Check out this interview with the CEO of Libra Energy in Canada to get an instant perspective on how successful Kobold has been in achieving these objectives.

Skip to 40:03 if it doesn’t autoplay from there.

The dollars that [Kobold] is going to have to put into this [JV], exceeds what our company valuation is.

Ladies and gentlemen, that’s leverage.

So to explore what these joint ventures entail, let’s do a brief tour through what Kobold was able to extract in this Joint Venture.

The deal between Libra Energy and Kobold is structured around an earn-in interest. Essentially, Kobold has to commit to spending enough money on mining exploration activities to obtain a Joint Venture on the underlying mineral claim.

In this case, Kobold committed to spending $33m in exploration activities on 3 properties in Libra’s portfolio.2

Some things to note about the deal:

1. They got the best mining claim sites. Kobold was only interested in three mining sites for exploration: Flanders North, Flanders South, and Soules Bay-Carron.

It turns out those are 3 of the 4 best potential sites in Libra’s portfolio.

2. They maximize their partnership interest. If Kobold performs $33m in exploration activities over 6 years, they will own 75% of the joint venture. And you better believe (as we will see later), that they are only committing that $33m if there’s going to be a huge outcome.

So in short, Kobold can bring in proprietary technology into land with active mineral claims, determine relatively quickly if there’s potential, get vast upside if there is, and nearly no downside if there isn’t.3

And so it’s no surprise that Kobold has structured a large swath of JVs.

Kobold’s Joint Ventures

I wanted to give you a brief glimpse into the breadth of these JVs. If there’s a region in the world where there’s a viable path to a mine, my guess is Kobold already has or is structuring a JV with partners to enable exploration.

And there’s one other important thing we can glean here. You’ll note that there’s two projects that have been subsequently abandoned by Kobold during either an earn-in period or during the early process of a JV.

In both cases, what’s remarkable is how quickly Kobold entered and then left.

The Greenland project was one of their earliest announced JVs. The partnership was formalized in 2022 and officially ended this year. But Kobold? Kobold was in and out. According to the press release, Kobold only did exploration activities during 2022.

We can surmise that they didn’t like what they saw and quickly bounced. Now, their expenditures here were $13.4 million. My guess is that in every subsequent JV, (because again, this was their earliest) the initial mining exploration projects to determine mining viability are going to decrease. There’s simply too many gains to be had in one company from technology improvement and human talent learning curves.

A similar situation occurred with Midnight Sun mining in the Dumbwa JV. Kobold was in and out within a year.

If Kobold doesn’t reach a positive determination quickly or if there’s complications with their JV counterparty, their team has shown a willingness to get out fast.

The crowning achievement thus far of Kobold is on the Mingomba mining joint venture. Kobold initially bought a $150 million stake in the region in 2022, subsequently discovered positive results, and then ended up discovering that this could be one of the largest mining projects for copper in all of Africa.

This too is a JV construct. Except this time it’s with a pseudo-government-aligned organization. It used to be a fully public government enterprise until some international banks forced the mining company to go private. This is essentially the biggest provable milestone thus far that Kobold has incredible advantages, not only on the JV side, (they were able to extract 80% of this stake in the mine), but also in their technological ability to go further within two to three years to determine the viability of a mining operation. The net revenue from this project is going to reach billions of dollars. And that opens up an entirely new dilemma. Because after all, once you’ve uncovered something worth billions, statecraft comes into play.

Statecraft

Building a mine after exploration is a multi-year process with constant negotiation and renegotiation and that increasingly involves geopolitical negotiation. It requires deal-making acumen across financial markets and political actors. Kobold CEO Kurt House has had to spend myriad amounts of time across the Biden administration, Trump administration, and across the Zambian government, ensuring that there are no key blockers to building the mine and subsequently being able to export mineral resources into other countries.

For instance, the Biden administration at the tail end of 2024 negotiated the construction of the Lobito Corridor, which is designed to connect the CopperBelt region to a port.4 It’s does you no good if you mine a resource but not get it shipped. And as always, if there’s a large enough underlying mineral value to be extracted, everyone’s elbows come out. Building the mine on expedited time horizons are not only matters of capital efficacy but of geopolitical viability. If you don’t extract the resources fast enough for the state to be made whole and to fund often their entire government, you risk losing control of the project altogether.

Democratic Republic of the Congo

Perhaps this is most clear in the DRC, where Kobold has also purchased 100% mineral claims throughout the Monono region.

Importantly, here, this is being litigated! Now, this is not legal analysis, but here’s my interpretation of what happened. A prior mining company named AVZ Minerals had the right to mine lithium in this region. The lithium deposits were documented and yet there’s still no lithium being extracted. As a result, the DRC terminated AVZ’s licenses, saw Kobold Metals success in Zambia, and went all in on enabling Kobold to find and extract lithium in the region.

Suffice to say, this did not go over well. And after international hearings, legal determinations, the AVZ should not have been kicked out, tentative agreements between Kobold and AVZ for Kobold to acquire the interests, and subsequent scuttlebutt since then, Kobold has the permits and is ready to commence activities.

Again, leverage is everything. Elbows out and cut deals to unlock rare earth minerals. The critical bit of success for Kobold is not solely its technology, it’s their success at maneuvering these issues of statecraft and coming up on top.

Asymmetric Returns from Leverage

All of this amounts to a company capable of creating asymmetric bets at a scale that nobody else can.

So once more let’s document the core drivers here of this success.

1. Their balance sheet allows them to walk away.

Having a larger balance sheet enables them to take asymmetric bets and walk away the instant they don’t see upside. That’s a fundamentally different approach from what many of these other companies can tolerate.

2. Their technology and team enable them to reach these determinations far faster than a traditional explorer ever could.

3. Because of these unfair advantages, if minerals are discovered, they obtain nearly all the economics. These economics will be the critical driver of the next 30 years at Kobold. With measurable revenue comes new tradeoffs. Does it make sense for incremental dollars to go towards new exploration or towards mining efficiency?

Future

As Kobold begins to develop mineral deposits into fully operational mines, it will be fascinating to watch how their JV structures morph.

Right now, Kobold is willing to tolerate exploration expenses of areas that display potential for large scale mining projects.

But as they identify these assets and own mines that produce billions in copper and lithium a year, will they narrow down their strategic exploration projects?

Let’s remember that minerals are commodity markets. A glut of copper supply or cobalt supply will reduce the price of the minerals you’re extracting. And so an interesting dilemma may settle in. Remember, timelines are everything in mining. Your joint venture partners may not be interested in tolerating a 15 year build out timeline in order to protect your other interests from a glut of supply that renders it uneconomical. And so will we see joint ventures morph into full acquisition of the underlying mineral claims? The jury’s out. But as Kobold brings mines into production, the subsequent phase of their deal making will become fascinating to watch.

Thanks for reading. More to come this Friday.

Today we are doing a tour through what’s working in vertical AI architectures.

Agentic Workflows Work

The vast majority of vertical AI platforms right now are modeled off of a traditional workflow with AI components primarily built in for document processing and communications.

The key idea here: leverage AI where traditional software doesn’t work. But don’t make everything agents.

Nonlinear just announced a fundraise to do exactly this in the AEC industry. And of course a substantial portion of Harvey’s product is built around AI workflows.

Typically, vertical AI workflows are designed to leverage the least amount of generation necessary to solve a business problem.

And nearly every fast growing vertical AI company is primarily built around AI workflows. Given OpenAI is creating SDKs to help developers create these workflows faster, it would stand to reason this is an enduring architecture going forward.

And in fact, many of these platforms may have the best claim on becoming neo-systems of record. Vertical AI workflows platforms are often going after substantial and industry-wide pain points. In turn, that naturally starts to look like accumulating a data moat.

OpenEvidence and Agentic Search

In certain verticals, professionals have to digest a breadth of information that is simply impossible.

OpenEvidence recognized this problem in healthcare and set about creating an agentic search engine across the medical corpus. The architecture here is LLMs + RAG.

The success OpenEvidence has had is somewhat breathtaking:

OpenEvidence is actively used across more than 10,000 hospitals and medical centers nationwide and by more than 40% of physicians in the United States who log in daily to make high-stakes clinical decisions at the point of care. OpenEvidence continues to grow by over 65,000 new verified U.S. clinician registrations each month. In July of 2024, OpenEvidence supported approximately 358,000 logged-in, verified U.S. physician consultations in one month. One year later, OpenEvidence now handles that many each workday—and supports over 8,500,000 clinical consultations by logged-in, verified U.S. physicians per month—a 2,000%+ year-over-year growth rate. More than 100 million Americans this year will be treated by a doctor who used OpenEvidence.1

To date, OpenEvidence has given their search product to physicians for free. Can search platforms become the basis for a system of action for physicians and hospital systems in order to reap incredible revenue payoff?

That’s starting to crystallize with their new Visits release that looks a lot like a scribing product. But the twist here is that when you have access to the full medical corpus + specific patient visit data, you can likely begin to provide a litany of extremely valuable insights for patient care. And that’s can be monetized.

Agents work(?)

Are agents working? In my view, this is the biggest architectural question for AI services over the next 10 years.

Here’s why:

Selling work and labor can not be reduced to a series of workflows. Work requires intelligence. To some degree, this is the enduring moat of many AI services companies. Through leveraging AI, they are able to create more throughput in the specialized human labor inside their business.

And so the core question: can you enable even more throughput in specialized labor via taking over more and more of the “intelligence” required for individual tasks?

A tractable example: when does an architect no longer need be generating the actual CAD and design plans in a BIM solution?

When can a game designer generate assets in their game engine with near effortless ability?

When can other domains start to experience the Claude Code moment for their work?

I’m not sure. But it’s coming. And across the agent landscape, several patterns are starting to emerge to bring this to life.

The core ideas: vertical AI agent companies are going to spend a lot of their time thinking deeply about the right mix of custom tools, libraries, and context to give their agent. Don’t reinvent the agent devops. Concentrate time and attention on the right harness for agents.

Harnessing as a Service:

I really love the idea expressed in this graphic:

The author here makes a very simple point as it applies to agents:

1. Don’t do devops. Use Claude Code and off the shelf SDKs

2. Your value comes from the “harness” you give your agent. This is the set of tools, custom prompts, context, and sub-agents you give.

Here’s the money quote:

We’re moving rapidly towards a world where builders create custom harnesses and users plug into them to edit further or use as a product. We’re already starting to see this movement from companies like bolt who helped kick off the vibe coding revolution. They’re using Codex and Claude Code in their app building product directly and probably did a ton of harness customization to get the product to work well. There’s a massive opportunity for companies to start using existing harnesses as application primitives to build their product experiences around. My bet is that in the next 6 months, the majority of user facing AI products will be using an existing agent harness as their core user interaction pattern.

I think this is directionally right and there’s widespread vertical applicability.

In short, I think the agent world is quickly moving into a space where developers spend their time creating the right tools for agents to wield in the right harness.

The structure and development of these tools is perhaps the now the greatest open question for vertical agents.

Less is More: Smaller models as tools?

Last week, a paper from Samsung came out detailing a new model that obtained fairly impressive accuracy scores on some benchmarks.

The shocker? It only had about 7m parameters. LLMs? They typically have hundreds of billions of parameters.

This new model, called a Tiny Recursive Model (TRM), is based off another model paradigm that came out of China a few months back.

The big idea here: for problems that require heavy logic-based reasoning, how a model thinks is far more important than its size.

I’m going to go ahead and quote from Mostly Borrowed Ideas on the comparison between LLMs and this TRM here:2

Large Language Models (LLMs) like Gemini or GPT-4 typically generate answers… one word or token at a time. When solving a complex problem, they try to produce the solution in a single forward pass. If they make an error early on, they struggle to recover, often rendering the entire answer invalid.

To compensate, LLMs use techniques like Chain-of-Thought (CoT)… and Test-Time Compute (TTC), where they generate multiple answers and select the best one. However, these methods are computationally expensive and still fail on tasks requiring deep, multi-step logic, such as complex Sudoku puzzles, maze solving, and abstract reasoning benchmarks (i.e. ARC-AGI)

TRM [instead]… operates much like a human solving a difficult puzzle: iteratively making guesses, evaluating the consequences, and revising the guesses until a solution is found...

It keeps two things in memory: y: its current best guess for the answer, and z: an internal “scratchpad” of reasoning. At each step it updates z a few times, then updates y once, and repeats. That’s it. The loop lets a 2‑layer, ~7M‑parameter network iteratively fix its own mistakes (“less is more”).

Okay, so why should you care?

Well for one, this specific training run cost less than $500. Two, the data required to gain this performance was also incredibly small. We are talking about a couple thousand input/output pairs vs. the entire corpus of the internet.3

But here’s the business reason:

Agents can make great harnesses. But state of the art is quickly pointing to a world where smaller models can be more performant on subdomains than LLMs.

Is there a future where harnesses and foundation model-based agents leverage smaller models for specific domain tasks? Almost undoubtedly so. And that opens up some new potential moats for vertical AI companies against foundation models.

AI has a last mile problem in delivering the requisite performance to wholesale change a profession and render specialist labor more effective.

But perhaps smaller models aren’t the answer. There’s some bright people that are fairly convinced that current models are sufficient for nearly every domain.

Software Libraries as Tools

What are models extremely effective at in agent contexts? Code development. And in particular code development in Python and JavaScript.

In short, every domain has a list of software libraries that allow developers to wield massive amounts of leverage over work inside the domain:

That looks like a list of traditional software that is instead built and exposed specifically for agents:

• Canonical data models for the vertical

• Deterministic primitives: validators, calculators, normalizers, parsers, scorers, matchers for work product in the domain.

• Regulatory and policy logic embedded as code (eligibility rules, compliance checks, audit trails).

• Integrations to systems of record

• Safety rails: permission checks, PII/PHI tagging, redaction, and guard policies that the agent cannot bypass.

In theory, the goal here is rather than have an LLM live inside a workflow, turn many of the safeguards built into the workflow into software libraries and harnesses that agents are built to use natively.

These domain packages can start basic (an insurance policy parser) and quickly grow into potentially fully featured environments for both agents and humans to collaborate in (a new, agent-native BIM solution).

But whether through new software library packages developed entirely for agents to be able to wield effectively or through smaller models used as a tool for tasks, it seems clear that there are tractable paths forming for widespread agent adoption in myriad industries.

The future of vertical AI may very well belong to those who build the libraries that make intelligence usable.

Subscribe now

Today: a tour through capital structures.

The Capital Supercycle

Some anecdata:

• There’s an expected $2T entering the private credit markets over the next 5 years. This would double the current private credit market’s AUM. The raises are staggering.1

• There’s interest and growing demand for novel capital structures from founders, investors, and more. Interest and businesses performing rollups, buyouts, neo-franchise models, and more have only grown.

• Hard-tech too is fueling a capital renaissance. To reach industry transformation, novel hard-tech strategies are necessitating novel financing structures much earlier in company lifecycles. (Base Power is the example that most readily comes to mind).

• Great power dynamics are increasingly creating new wedge points for innovators. The interplay between global infrastructure, China’s Belt and Road Initiative, construction and manufacturing sectors, and the Trump administration’s “dealmaking as policy” are creating a new crop of winners and new opportunities across the equity and debt stack2

• Will Manidis’ takes on merchant banking seem prudent and exemplified at the upper echelon by Thrive Holdings and at the lower end that are rethinking company incubation altogether.

In short, the smartest founders I know are starting their businesses not simply with a product idea, but they are increasingly spending their time plotting the capital constructs that will enable them to do wholesale industry transformation.

We can maybe riff on an existing term and call this the capital structure idea maze.

I don’t want to be misunderstood: this isn’t operators asking “can I be more capital efficient via venture debt?” It’s founders increasingly asking “what mix of capital is critical for the success of industry transformation?”

Or in Will’s words:

Throw in “company” instead of “professional life” and I think we have a good sense for what is required.

Let’s take a brief tour through the new capital structures and realities in each part of the market.

Product-market-capital-fit.

It’s worth considering a potential emerging trend that likely will refactor some of the capital stack in years to come. One of the direct impacts for all sorts of companies is the new ways in which their businesses are becoming more “underwritable” for debt. My favorite example is General Catalyst’s Customer Value Fund. Scaling businesses can often determine their CAC/LTV mix. Why fund growth with equity? That’s the question that GC asked too and the result is a fund specifically for companies like Ladder to scale their businesses with less dilution.

Once a business has components that are “structured”, like S&M and CAC, those components deserve a solution that is precisely matched to their output – just like a company with inventory gets financing tied to that inventory, or a company in need of a machine takes out a financing tied to the value of that machine. Using equity to fund S&M is a poor use of that capital, which is even more pronounced now given the capital constrained environment.

Which leads to a new hypothesis one could have:

AI Services are a hypothetical new asset class

Here, let me paint a brief picture of asset-based finance.

Asset-based finance lends against an asset. In fintech, this has mostly looked like receivables purchases.

Pipe was one of the first to collateralize SaaS contracts, take a haircut on your monthly ARR to convert that into a lump sum annual pay out.

We have also seen the collateralization of sales via far better payments data from companies like Stripe, Square, and Toast.

These companies could underwrite the purchase of future sales and solve cash flow and lending challenges with certainty of execution for SMBs who previously had slow or no options to do this.

How does this apply to AI services you ask?

AI services are neatly tied with new arrangements that often move services from time-based billing models towards outcome-based project models.3

Crosby just announced their latest funding rounds. By no means, am I suggesting they should do this, but it’s a simple enough example to show how radical of a shift certain AI-based services are and how that enables different asset-based finance compositions.

Law firms historically bill hours with tons of variability in the number of hours required, the future variability of how many hours will be billed for a project, and the length of the project. Underwriting is slow. Lending facilities are imprecise. There’s no asset to really collateralize here. And so instead, it’s necessary to extend lines of credit.

Crosby? Crosby just says, “here is your price for an outcome on an MSA.”

And that simplification and corresponding ability to forecast the asset (the outcome) enables a service that was once too complex, costly, and onerous to underwrite to become a series of assets that can be collaterized individually.

In short, AI services simply by nature of their existence reduce the cost footprint to underwrite. Depending on the industry and context, these can now be modeled as future receivables to be purchased, measured around their outcome-linked components (if you want a General Catalyst-inspired flavor) and because AI services are de facto digitally native, they’re suddenly rendering transparent hundreds of different services that previously were invisible as an asset class.

That in turn reduces the cost to service, the cost of debt, the velocity of lending, the velocity of payback, and more.

You can imagine the net result:

A whole slew of receivables just became ripe for asset-backed lending.

How much does this matter? It might matter greatly in the future for AI services companies also performing buyouts, rollups, or other structures with longer cash cycles. It certainly may matter for future platforms that potentially are built in a Shopify-esque manner to allow SMBs to create AI-based services. The certainty of pricing and outcome with AI services could enable these companies to acquire an asset, convert customer agreements to outcome-based AI service models, collateralize these new receivable assets for quicker cash conversion, and grow faster. In turn, this could give rise to a new source of capital efficiency for services companies scaling.4

Does this happen and does it pencil out how I outlined? Jury is out, but it’s now theoretically possible and I find it extremely interesting.

Time to Innovate vs. Time to Scale - Rollups and Buyouts

If AI services represent capital efficiency at the micro level, rollups and buyouts represent that same logic at the corporate level. I won’t belabor this point too much right now:

Many of the well known rollups right now are on record as not spending much time thinking about debt. These are primarily equity-infused deals designed to give teams time to fully understand the P&L impact that AI can have on a firm prior to scaling. It is quite hard to imagine that at a certain point, many of these companies won’t also use acquisition financing to scale.

But equity is going to continue to be a part of the stack and my guess is a meaningful part of accelerating business acquistions. And I’m very curious to see how startups utilize equity-incentives in their acquisition structure. After all, this is a massive component of what led to Acrisure becoming one of the largest insurance brokerage in the US.

Industry Disruption through Capital Structures

Base Power is 50% battery company and 50% “we found an incredible structure to unlock a crazy revenue pool.”

The short of it: manufacturing is capex intensive but worth doing for all sorts of reasons. If you want to manufacture batteries5, what if you initially operated as both a battery leasing company6 and a Retail Electric Provider to enable another revenue pool through buying and selling power into customer storage?

Private credit markets love leasing arrangements because they generate consistent yield over time. Base’s unlock was partly realizing that you could sell batteries, not to end customers, but instead sorta to private credit funds.7

Shinkei Systems can be somewhat modeled the same way.

Great Powers and Capital Structures

Lastly, new horizons are opening. I have no take on whether companies should do this nor wish to speculate on whether its a good development or not, but it is now a viable(?) avenue.

Government equity stakes in exchange for industry approval is happening.

In fact, maybe we leave a more extended analysis here for a different time. Because after all, new capital realities are where the alpha lies both for investors and founders.

Subscribe now

1

Armchair Quarterbacking

The Data Tapes Edition #31

October 02, 2025 | Read Online…

Read more

7 months ago · Conor Witt

Today’s newsletter: getting more specific around underwriting AI in buyout contexts.

EA gets LBO’d

2 weeks ago, the largest LBO on record took place, resulting in EA being taken private.

$55B purchase price, with $18-20B in debt underwritten by JP Morgan.

Reportedly Silver Lake’s co-CEO, Egon Durban has been a long term admirer of EA’s management team and believed the business was undervalued by the public markets.

Durban said EA was an attractive target because of its “premier sports franchise, with accelerating revenue growth and strong and scaling free cash flow”. He added that Wilson had “doubled revenue . . . and driven a fivefold increase in market cap during his tenure” since becoming chief in 2013.1

And there are certainly intrinsic reasons to believe the business with or without AI can sustain performance

But also reportedly, with the amount of debt being taken out here, the majority of EA’s free cash flow will be going towards debt service.

And while there are certainly other factors you could believe will lead to more free cash flow inside the business, the value creation case in large part must stem from AI.

The investors are betting that AI-based cost cuts will significantly boost EA’s profits in the coming years, people involved in the transaction told the Financial Times.

And so it’s worth considering where exactly the AI boost can come from and what will be required to make it happen. After all, as founders and investors in every market become more interested in inorganic growth and value creation, it’s incredibly important to pinpoint what exactly will lead to higher multiples post-acquisition, revenue growth, and increased profits.

Because the simple truth is this: if AI value creation works, the arms race for equity and private credit financing will speed up in every segment of the economy.

First, lets briefly touch on what underlying principles of the business enable the investment group to feel confident in performing a takeout

The gaming industry is an incredibly difficult place to perform M&A.

First, most gaming franchises only release new titles every 6-7 years. This in many respects mirrors revenue graphs that look a lot like depreciation curves in oil and gas. Tons of “capex” is required to bring a new “well” or game to life, and then corresponding revenue from that well (or game) dries up over time.

This is why gaming much like oil and gas has consolidated so aggressively. You want many different franchises, management efficiencies, and a ton of different production sites in order to create durable value.

Second, there was a theory that live service games which resemble annual recurring revenue (gamers log in consistently, buy annual battle passes, and show up year over year) would defy these dynamics and result in companies that look more like software revenue metrics. Good in theory, extremely difficult to implement practically.

Sony in particular bet the company on this with a string of acquisitions and new title bets. The majority of these have been failures.

The glut of potential live service models has resulted in the majority of these never recovering their development costs and not being the boon to revenue that investors expected.

So what does work consistently from a shareholder perspective?

Mobile Games:

Mobile games are immensely profitable due to lower development costs, potential install bases in the billions, and because of micro-transaction models.

Take Two, one of the last independent gaming studios trading publicly, reports 52% ($2.94B) of their revenue from mobile with 79% of their total revenue coming from in-game transactions.

EA’s mobile presence is in comparison far more limited comprising 15% of revenue:2

But what EA does have is arguably the most durable revenue in gaming via sports licenses to almost every major sport globally.

Licensed Properties

The outsized successes in gaming that have proved durable are live-service or pseudo-subscription business models from licensed IP.3

Year after year, gamers show up to buy the latest FC or Madden or College Football game.

Revenue continues to grow, but more importantly, net income is steady - EA isn’t susceptible to massive swings in R&D costs for their business.

In short, the business model is durable, resistant to gaming’s ills, and through its licensing agreements with every major sport is not as susceptible to revenue and cash flow swings. Add in a stellar management team who has renegotiated these licenses (or walked away in the case of FIFA) as they manage the business, and you have a truly great business that can theoretically get even better via AI.

You’ll also note one other thing: revenue is stable, but not necessarily growing. The sports audience is likely close to tapped out. Growth is going to need to come from other segments whether mobile, new IP, new licensed opportunities, or more.

But with the pending debt servicing, it’s going to be quite tough to develop new IP unless there’s some breakthroughs. So the win condition for EA broadly is going to be contingent on doing more with less. But unlike nearly every other business in gaming, they have the flexibility to innovate here without

So how does that happen?

The AI Case

The simple case:

1. AI is arguably really great at one thing: coding.

2. Games are bundles of code.

3. EA spends roughly ~$2.8B a year on R&D.4

4. If AI can deliver reduced R&D costs or increased speed in the development cycle, the result should be incredibly lucrative for shareholders.

But there’s one expected problem:

Gaming engines are incredibly complex environments. Off-the-shelf coding agents thus far do not excel in complex codebase environments.

They in particular don’t excel when the gaming engine comprises C++, proprietary scripting languages, and myriad other factors.

There is one advantage that EA has in this regard: they own their engine, Frostbite.

Originally, built for Battlefield, the majority of EA releases are now built on top of Frostbite. If EA can enable faster development around all aspects of Frostbite, they will not only save costs on their current gaming portfolio, but be able to bring new properties to life in reduced dev cycles.

And so all AI initiatives at EA in my view are going to be concentrated on turning Frostbite into an RL environment that enables custom coding agents across asset creation, game development, engine development, and more.

Let’s suppose the cost here per year to train is on the order of $20m-$100m with an expected 30%-40% efficiency gains in their games development pipelines.

Well two things then hold true:

1. An extra 700 million or so in cash flow gets shaken loose from current R&D budgets. There’s your increased budget for debt service and more.

2. New IP gets developed in less time or with less budget. Battlefield 6 reportedly has cost over $400m over the last 4 years to develop. It’s been 7 years since the last title release. If new titles can be accelerated, there’s your return to growth.

In short, the success of the EA acquisition is going to primarily come from turning the entire company into an RL feedback loop concentrated around the Frostbite engine.

There are various checkpoints in this endeavor I’d imagine. For instance, microtransactions through selling weapon or player skins is an incredibly lucrative strategy.

Can you create a faster design to implementation cycle using AI throughout the entire process?

And One last thing: Is it any surprise that Kushner is involved fresh off announcing a round for his AI implementation company, Brain Co?

Things to monitor

So the list of questions, AI companies should be monitoring:

1. How much does it cost to actually convert Frostbite into an RL environment and bring coding agents to market?

2. Does this work or are there other factors that create bottlenecks that are not reducible to AI?

3. If this works (and for the record, my bet is that it will), does AI here solely impact profits or does it allow EA to reaccelerate growth in a much more capital efficient manner?

I’m eager to watch this play out as to a large degree, the initial wave of these transactions will underwrite the entire AI inorganic playbooks that are beginning to develop over the next 20 years.

Subscribe now

Today’s newsletter: Durability in vertical AI.

Subscribe now

As the onslaught of vertical AI continues, I think it’s worth exploring where true moats can be found. As revenue figures skyrocket and everyone celebrates hefty ARR milestones in short time periods, I wonder if the durability of some of these businesses is present.

“We care deeply about organic revenue growth, but only if it’s profitable and sustainable. We will not trade profitability for topline.”
— Mark Leonard

I’m not suggesting we have to go full Mark Leonard, but I do want a deeper reason for why certain vertical AI businesses are sustainable beyond “the revenue numbers are high.”

One of the earliest pieces that I wrote used Moneyball as a way to explore the defensibility of vertical software over and against horizontal software.

At the time I wrote that vertical SaaS’s value above horizontal replacement stems from:

I want to revisit that now because these reasons are still true, but the focus here for how vertical AI will succeed are quite different. I take it as a given that vertical AI wins if there is clear value above:

1. Horizontal agents

2. Foundation model product offerings

So far, that should be an obvious insight, no?

Identifying where the true value above replacement lies is potentially harder:

What’s defensible?

Many will point to the distribution angle in vertical AI. This is a valid point; you could certainly argue that focused teams that crush on distribution will win even if horizontal products are at feature parity.But the durability here is an open question.

For instance, if you provide a RAG layer across a document set is this durable as models improve and myriad players continue to focus on this horizontally?

If you are solely a voice agent platform, does this endure?

Part of the problem in AI generally now is the widespread gap between what you have to do to coerce a model for work in a particular business setting and then expected capability of a model at some future date. This dichotomy pinpoints the complexity in articulating the long term durability of solutions that are codifying AI workflows and prompts inside of an industry.

The fact that I can argue it both ways maybe underlies the central point:

Defensibility ideally should not be questionable vis a vis horizontal companies. If there is a defensibility question, the long term value above replacement is in question. This does not mean that you can still be outcompeted by a fellow vertical company, but that is an altogether different question.

So, that would seem to mean we should be seeking durability beyond pure model capabilities.

Why did the last wave of vertical market software succeed? It was impossible to rip out of company operations due to system of record dynamics. If we assume that system of record dynamics might be harder to capture for pure vertical AI companies, where then does the moat come from?

Vertical value then must lie in a couple different things where horizontal software can simply not go.

That list in my view looks like the following:

1. Network effects

2. Compounding expertise that is not evenly distributed

3. Services

4. Industry business model reconfigurations

Let’s pull out some case studies to see what I mean:

Network Effects:

By network effects, we simply mean: the more customers, the more each individual customer benefits. Here AI is being leveraged as a way to unmask discoveries and analysis across an entire industry in order to provide insights that were previously illegible.

Perhaps there’s no greater company to talk about than Loop:

The true value of Loop is not in its financial LLM, albeit this is obviously a core and necessary driver of their success.

The true value lies in the compounding value of financial analysis for every invoice and cost metric in the industry. Every one wants access to this network. Every financial officer wants to know if their costs are out of whack vs. the industry norm.

In essence, signing up for Loop is signing up for a network for analyzing all financial information across the industry.

Compounding expertise that is not evenly distributed

If you’ve got a better way to say this, let me know! Here’s what I mean: because AI can automate drudgery, it unlocks true experts to focus in on the few things that are not commodity. Businesses that help experts focus here and alleviate the drudgery can distribute that expertise and talents in ways that were previously bottlenecked by the unit economics of running these sorts of businesses.

My favorite example here is OffDeal:

How do you enable SMBs to experience Goldman Sachs level investment banking when their business goes up for sale?

Well, you need to compress the unit economics of a transaction - bankers can’t be spending time on the minutia of decks, CIMs, and other things, and then you need to enable the best talent to share in the economics of the deal in order to give them a reason to join your firm.

In other words, the value that OffDeal accrues is in relation to their success at using AI internally to unlock better deal economics, give SMBs better access to expertise in selling their business, and in incentivizing said talent to spend their time with a market that they previously wouldn’t.1

Services:

This one is basic: offer a service to customers that don’t just want to use a platform to manage themselves. EvenUp offers a service and intelligence platform that enables their clients to generate more revenue at lower time expenditures. Demand letters, medical chronologies, and more.

If I’m a busy personal injury lawyer, and I’m confident that EvenUp is continuously incentivized to improve every aspect of their offering by utilizing the latest AI capabilities to make their service better to retain me as a customer, why would I bother managing this myself?

Industry Business Model Reconfigurations:

Lastly, I’ll group this as one category, but there’s an argument that this is nebulous. What I’m really getting at here: have a service so good and so cohesive across the core operations of a business, that companies are willing to contractually make you part of their continued life as a business.

Right now, you could loosely group all the MSO structures in healthcare into this category.

The moats here are contractual. Once you’re part of the MSO, it’s nigh impossible to leave.

And because the moat is the contract, this is the area where I have the most questions. If the vast majority of businesses on your platform are 20-30 year businesses, then this concern goes away. But is this a crucial reason we haven’t seen these businesses scaled outside of health?

Tentatively, Reinforcement Learning:

I want to add one more here that certainly could be a source of compounding value above horizontals: Reinforcement Learning and RL Environments

RL is currently the best method for coercing models into new performance thresholds necessary for certain aspects of industry work.

If RL does not generalize, then compounding value can be found in the companies that own the RL environments and model finetunes for entire industries.

Subscribe now

Trading Cards

The economy is in a transitional state where the best place to make money is either in AI, sports betting, healthcare, or in buying Labubus and Pokemon cards.

The first two seem far less fun day to day than Pokemon cards and so we are starting to get asks from Shkreli if there’s anyone that wants to buy $200m of Pokemon cards.

But this isn’t pure ironic tweeting, the hobby markets are at least theoretically becoming an investment case.

We’ve already spent some time looking into the industry, mainly the card shop side. But there’s a case being made for how to transition trading cards and hobbyist markets generally from a pure hobby into one that Wall Street could do an asset-based finance play on in order to put a portion of the trillions in private credit to work.1 After all, what teacher’s pension fund doesn’t want exposure to Gen 1 Charizards?

And if your main reaction is “this is the stupidest thing, I’ve ever heard,” that’s mostly my reaction as well. But others are starting to classify the potential in the trading card market as something halfway between sports gambling or investing.

And in that sense, this industry is starting to mirror the same consumer trends driving Polymarket, Draftkings, Robinhood, and more.

Or you can forget the trends and simply bet that it’s likely to be attempted because the market is worth $24B today and could go up to $50B by 2030.

The logic goes like this:

High level:

There’s likely about $20B in Pokemon cards in value. We can quantify this fairly well. The industry has robust and pseudo-liquid markets, professional graders like PSA (which is an incredible business) for assessing individual card quality, and lastly consumer appetite just keeps growing for hobbyist assets whether its Pokemon cards or Labubus.

My guess is the PE funds thinking about this (and likely some founders) are grouping the opportunity into a couple different categories:

The AUM play

This one is simple enough. This likely looks a lot like wine investing: capture inventory that looks likely to appreciate, charge AUM fees, and keep your operating costs incredibly low.

In fact, that same strategy has been done on the wine investing side by Vinovest.

The benefit here vis a vis wine investing is that storing Pokemon cards and hobbyist assets is not as challenging. Any secure warehouse could do. Buy cards, grade cards, store cards. And because again, the consumer markets are frothing for speculative assets, it’s not hard to imagine the opportunity in a single tech AUM play being far larger than the wine investing market.

If we just use Vinovest numbers as indicative, could you get AUM to roughly 3.5B to have a $100m annual management fee? And could you do that on less than 50 headcount internally? It doesn’t sound too far-fetched in my book from this perspective.

Trading Platforms:

This is where things start to get really wonky in my view:

If there are firms performing AUM plays and consumer appetites continue to grow, then likely appetite for financially trading hobbyist assets is going to grow. Right now, if you’re going to trade Pokemon cards, you have to take physical receipt of the card itself. And so markets like Ebay, local card shops and more act as trading platforms. But if Pokemon cards becomes more about owning the card vs. having physical ownership at all times, could you see trading platforms for hobbyist goods come about that just issue you a certification of ownership, maintain storage of the card, and settle transactions? Heck, maybe they’ll even let you loan your card out for other traders to speculate on the price…

After all, is there that much of a difference between trading NFTs and Pokemon cards from an SEC perspective?

As one of the tweets mentions, this naturally opens up opportunities for all sorts of financial products that you need in order to enhance the trader experience… did I mention private credit markets are approaching 4T?

Anyways, should this get off the ground? I’m unconvinced. Will it? If interest rates come down, all bets are off.

Information Blocking and the Cures Act

I was catching up with the brilliant Brendan Keeler of Health API Guy and he got my wheels turning on some of the coming blows between systems of record and upstarts and he mentioned some of the things in healthcare that have been done to maintain some balance.

Healthcare famously has some of the most entrenched systems of record. There’s just no path to disrupting Epic anytime soon. But regulators also know this and want to encourage healthcare innovation. An industry monopoly is great, until patient outcomes suffer and physician practices are bloated because innovation can’t get useful data out of EHRs. As a result, the Information Blocking provision of the Cures Act mandates that EHRs cannot block information. The data must be able to leave systems (with of course several caveats).

I have two thoughts on this:

1. Is there a world where the Cures Act is used as a blueprint for regulations on systems of record in other industries?

   1. This is currently being litigated by Tekion against CDK Global in the automotive space.2

2. In industries with information blocking regulations, systems of record are still not necessarily designing their APIs to promote information leaving the system. But wait a minute: now we have computer use agents that can export the data simply via using the software in a defined way. If information blocking rules start to regulate other system of record environments, the data is getting out even if the APIs are rate limited, poorly designed, or simply not present.

Given that the majority of venture returns are going to be tied up in AI platforms that necessitate data from current systems of record, I am going to be very curious if this becomes a more public issue and policy demand over the next 4-5 years.

Mark Leonard

I did want to make a brief note here: A giant of vertical software announced his abrupt retirement this week. Given its abruptness, my thoughts and prayers are with Leonard and his family as there’s some speculation of health issues.

Constellation Software is one of the best managed and best performing capital deployment vehicles of the last 30 years. And Mark’s writing will continue to hold up over the next 30 years.

Subscribe now

The Conditions Have Changed Materially

I’ve been mulling over the game theory between systems of record and AI. About a year back, there was heightened fear that systems of record were going to be displaced by AI systems entirely. And while I remain less than convinced that systems of record are best positioned for the wholesale transformation of work stemming from some superintelligence, we don’t seem anywhere close to that reality right now.

In fact, the new paradigmatic approach in AI implicitly seems to acknowledge that fast takeoff scenarios are out the window. Instead, current approaches are leveraging RL techniques to upregulate desirable behavior + reasoning in models. And these RL techniques are being leveraged via environments that essentially equip a model with a playground for some sort of activity; this could be an M&A negotiation playground, a Wordle simulator, or literally anything else. These environments come with verifiable rewards, grading mechanisms which the labs then utilize to essentially pick the correct reasoning traces that result in the most favorable negotiation, the smallest number of Wordle guesses, or again literally anything else.

Suffice to say, we don’t know a) if this generalizes as a technique, or b) the number of environments that are necessary to assemble the AI god. But we do know that it works. If there is an eval, the labs are going to ensure their model crushes the eval.

But the reason I bring this up is that it illuminates the new conditions that underly vertical strategy. Fast takeoffs seem improbable, model progress is contingent on RL post-training, and these techniques are contingent on having properly constructed environments where models can make meaningful progression.

This is dramatically different from a timeline where training across a corpus of internet data is the only thing that matters, generalized models become more and more capable with pre-training compute, and AI progress blows past nearly every single software company.

Now? It looks more and more like vertical expertise is going to become a defining feature of the path to better AI models.

Game Theory

Here’s the new game theory gist:

Let’s take it as a given that vertical companies win in perpetuity if a couple conditions hold true: 1

1. Better AI models do not ultimately lead to system of record obsolescence.2

2. AI interaction for a business is mediated through their vertical platform and as a result vertical companies are fast adopters of new AI capabilities.3

This second portion is where things get interesting in my view. Because in a world where model progress is dictated through eval and environment construction, the companies that get the most rapid progress in their domains are the companies that not only operate as wrappers for AI models but instead as advocates for model improvement on the core deliverables that their industry cares about.

I use the work “deliverables”… well deliberately: Every model release grows the number of discrete productized “deliverables” that companies can theoretically build and support for their customers. AI systems are concerned with action. Every useful action for an organization must end with something being handed over. The caliber of this deliverable is ultimately what the entire AI experience will be judged by.4

Round 1 of these deliverables has mostly been structured around customer communications. Voice and text models made this feasible and it became a relatively easy addition to justify.

Round 2 (you can make the case this was also a part of Round 1) of these deliverables has been something like “RAG systems” for internal company search.

Things now get tricky after Round 2:

Now we want agents that can accomplish long time horizon tasks that require tons of context, various decisions, and multimodal use of tools, data artifacts, and more. These capabilities are now highly industry contingent for the quality of the deliverable.

From here on out5, deliverable delivery from agents is contingent on a) high quality data, b) expert feedback on evaluation and deliverable quality, and c) simulated environments that mirror the industry deliverable closely enough to result in high agent quality.

Every party - end customers foundation labs, and vertical platforms should want this to happen.

After all, for the labs this means better models on the path to AGI, for vertical companies this means increased revenue and customer capture from these more capable models delivering work inside of their platform, and for customers it means productivity gains.

So vertical companies sit between two groups who both are ironically willing to pay for increased model capability. The labs and vertical company’s customers.

What’s the opportunity here?

This piece is already getting too long for a Wednesday piece. But is this some sort of opportunity for vertical companies to monetize the creation of RL environments for their industry that then they are positioned to sell to their customer nearly immediately upon model capability improvement?

I’d imagine some sort of the following scenario:

1. Foundation model companies need RL data

2. Vertical companies have a great view into industry deliverables and the simulated environment that is required to evaluate and train agent capability on this task.

3. Likewise, vertical companies have distribution into customers that have expertise in evaluating deliverable and data quality.6

4. The labs are essentially willing to subsidize this for the foreseeable future as the race for the model to rule them all continues. This largely functions as covered R&D expense for vertical companies. I want to highlight just how unusual this is: the labs are willing to pay vertical experts to build models that vertical software companies will likely end up bringing to market as fully packaged products for an industry with revenue payoffs easily eclipsing tens of millions in ARR per year.7

5. The sheer number of deliverables where we likely need verifiable rewards and training data is insane. If we were to just take the construction sector, we are immediately looking at dozens if not hundreds of potential model capabilities that need to be developed. Mercor’s sentiment is not entirely wrong: The economy will become an RL Environment Machine.

6. Once the model is newly capable, the environment can be repackaged from a simulation into a product in the vertical. This easily can result in a large revenue uplift. And you can likewise imagine the labs prioritizing vertical companies that give them the best RL environments for the first look at new model capabilities.

Now there are various challenges with this ranging from time vs. effort, partnership negotiations, environment creation8, legal and compliance challenges, and more.

But I think there’s something here. And I even wonder if the challenges that come to mind are solvable via a company who essentially operates as an embedded environment creator and middleman between the labs and vertical companies.

The opportunity for that company seems quite large and there’s some really interesting strategies to be run if the market for something like this is there. And because everyone right now is incentivized towards enhanced model capabilities with large revenue payoffs for all involved, I’m eager to explore this more. If this piques your interest even in this abstract piece, I’d love to hear from you.

Everyone wants models to get better. Everyone needs increased capabilities in their specific sector, the labs also want this, the labs are willing to pay to make this happen, and finally vertical companies likely become the biggest beneficiaries: if models advance, their moats get stronger as long as they have proprietary access. - which may partly stem from being the advocates and expeditors for model progress through creating the next thousand RL environments on the path to AGI.

Subscribe now

Today’s newsletter:

An amalgamation of thoughts on how atoms-based opportunities are getting financed at a high level.

Bits-based businesses are what we all know and love: software and data. High margins, relatively uncomplicated playbooks post-initial R&D, and large markets all yield a profile that has come to attract a wide range of investors. Appetite for risk still dictates the earliest stages and the eventual payoff, but at the other end of the spectrum in the publics: technology companies are attractive to shareholders that prioritize cash flow, dividends, and maturity.

Whatever your risk profile, you can find a bits-based business.

Atoms? That’s where it gets complicated.

Perhaps though, it’s better to distinguish between asset-light and asset-based businesses. Investors still love asset-light strategies. And they’ve parlayed Uber, Airbnb, and plenty more into massive equity growth. These companies are all coordinating atoms but leaving the atoms in someone else’s hands.

All else equal, asset-light is preferred. There’s less inherent complications, you take on no operating risk, and you can make the world more efficient without as much asset risk.

There’s always an inherent unease amongst builders around these strategies. After all, your business is still dependent upon somebody owning and operating the assets. And while it is obvious that Airbnb’s business model is the best fit for aggregating supply of previously underutilized assets , there are plenty of others that are contingent on stronger ties between the asset-base and the asset-light operation layer.

Tech operators are increasingly looking at legacy industries and seeing asset ownership as a necessary part of transforming the industry.

However, asset ownership is incredibly difficult. Look no further than public companies operating in asset-based industries. At the management level, most of these companies look alike: they’re entirely focused on managing the inherent risk in dealing with their portfolio. To do this, they increasingly rely upon debt financing to remove working capital requirements and look “asset-lighter” on the balance sheet.

1. In practice, these companies tend to act as a management vehicle which also happens to perform some discrete end goal, whether its building railroads, airplanes, or widgets.1

2. Because financing structures reign supreme, there’s often very little startup scene that has a shot at large outcomes. Startups quickly find that they are not capable of financing the working capital.2 Large financiers simply gravitate towards opportunities that they can forecast more easily.

3. The majority of new entrants into atoms-based businesses are private equity firms who can apply better management techniques and financial engineering methods to smaller firms. Many of these firms have not applied novel strategies around their operations, but instead have applied paradigmatic management techniques that enable larger firms to gobble up their own rollup.

Even with breakthrough strategy or new products, any startup ecosystem must account for these factors. Managing assets is often a huge and risky hassle.

The conflicting forces become evident:

1. Asset-light strategies are perhaps the best way to focus upon value creation in legacy industries.

2. Asset-light opportunities may be reaching end of life due to the proximity required between managing assets and operating assets inherent in most other industries.

3. The minute you switch into asset-ownership, forces compel you to spend far more energy on managing assets vs. innovation.

   It will take plenty of independent developments to reverse these conflicts. But they may already be reversing.

Let’s focus on the two biggest reasons:

1. The growth of private credit

2. The evolution of the propco/opco.

Private Credit is Growing

What was the biggest reason fintech as a category took off in the mid to late-2010s?

There’s various things you could point to, but I think the one that matters the most is the rise of asset-based private credit funds. A new pool of debt facilities came into the market looking for yield and were often willing to partner early with fintechs to create new financial asset classes.

The growth in private credit itself is pretty astonishing.

It took nearly 15 years for private credit to add $2T to its reserves. It will only take roughly 5 more to add another $2.5T.3 At the same time, that private credit is skyrocketing, bank lending has stagnated.

There are many reasons for this but here are a few:

1. Novel credit strategies are not a good fit for bank lenders. These novel credit strategies are paying off. This is the fintech story in a nutshell.4

2. Private equity has decisively shifted to private credit as their debt source.

3. The world has simply become more financially engineered with the banks being relatively tapped out due to a mix of regulations and internal speed.

So what to make of all this?

The demand for private credit has never been higher and the appetite amongst investors for private credit exposure has also never been higher.

The growth in private credit assets is quickly leading to growth in the number of opportunities private credit will look at.

At the same time that builders are looking for asset-based innovation, there is now an investor class looking for yield wherever they can find it.

How will technologists and private credit partner together to capitalize upon this?

An Interlude on Opcos/Propcos

One potential method is through the Opco-Propco model. I won’t belabor the point but, opcos-propcos are a relatively well-known method for tranching investors into different risk profiles. In short, some investors want exposure to asset-light parts of the business: the operation/management company and others want exposure to asset-based aspects: the property company.

There’s various ways to structure these companies. Red Lobster is a bit infamous for their Opco/Propco structure. Sell (red) lobsters through your OpsCo, pay rent to your Propco. But even within REIT-type efforts with some nuance around the exact opco-propco structure, there is more and more reliance upon technology as the operational and management advantage.5

My personal favorite is Lineage Logistics. Their key insight was simple: there was an inadequate cold storage supply chain. You had to both catalyze the real estate footprint for cold storage and operate the assets. The rest is a story of incredible management and structuring revolving around a opco-propco structure to build one of the largest cold storage REITs out there. Alongside their real estate playbook, they’ve spent more and more time building out an integrated automation and services layer.

There are theoretically no limit to the number of companies that can be structured as a REIT, opco-propco arrangement, or draw lessons from the overall structure. So long as there are assets with a known yield profile, the private credit markets are increasingly jumping at these opportunities.

Opcos/Propcos for tech-driven strategies have not fully taken off yet. But my hunch is that you will see more interest in this segment. Why? Because tech-driven strategies are increasingly the basis for novel underwriting and for novel returns on assets. This too is a massive component of the fintech story; fintechs underwrote deals that were impossible to do without software and ML enabled underwriting.

As this plays out, we would expect private credit to increasingly recognize that technology can create leverage both in a) the underwriting of assets and in b) the operation of assets outside of fintech related industries. Utilizing technology to perform unique underwriting and operations will increasingly be the lever that opco/propco structures build their business around.

New Financing and Underwriting Arrangements:

The lesson of fintech and vertical SaaS is that to the extent there’s a new line of payments volume, financing, or structuring that can be captured by a disruptor alongside the asset class they are financing, the juicier the opportunity.

Several models have been tried over the past few years that have utilized innovative financing and underwriting methods to build ownership positions in new financial instruments.

Two of the better known ones are:

1. Rent-to-own models:

2. ESOP Structuring:

Key to these models is an insight: increasing liquidity in assets via new instruments is often the best way to build a new opco-propco model altogether.

In these cases, companies focus on an opco that is concerned with new data and underwriting strategies with an ancillary focus on operating the assets that result, while the propco provides the capital for acquiring the underlying assets.

And in fact even in traditional industries, we can see opco/propco models working to quickly accumulate novel assets.

Opportunities abound.

Minerals + Oil and Gas:

Kobold Metals is in essence a land underwriting company. They use incredible ML models to predict where rare earth metals lie and then acquire land prior to competition. It’s worked in a big way. They’re currently spending their time on about 60 different exploration sites after discovering one of the largest copper deposits in Zambia.

My hunch is that there are plenty of other opportunities for novel propco-opco models involving land and mineral rights.6

Real Estate:

Within real estate, propco-opcos hold plenty of potential avenues for new opco-propco models.

Last week, we talked about Blue Jeans Golf. Opco/Propco structures give the ability to tranche investors into groups who want land value appreciation and rent income and investors who want direct exposure to the Golf Ranch brand.

Perhaps merely the business of structuring REITs to help enable innovative real-world companies becomes valuable.

"Surprisingly, while OpCo money is most at risk and should theoretically be harder to access, the real funding gap in the market lies on the PropCo side." - Thibault Weston Smith (Co-Founder, Crayon Partners)

- taken from Cherry Lawn’s Transitional Capital

Structurally, most acquisition based businesses are not a fit for opco/propco models. Or at least, they aren’t pure play opco/propcos. However, many business owners are structurally incentivized to sell any real estate holdings as part of their acquisition but while acquirers are often willing to take on the real estate, it’s often not mission critical. There are many reasons why owners wish to exit the real estate as part of the business sale:

1. The real estate has intrinsic value that may actually give rise to a higher exit price than the underlying business.

2. There is ultimately no point in managing the real estate if the underlying business is sold. Most owners have no compounding advantage in owning the real estate.

3. The value creation of a rollup has little to do with real estate operations. It is not a focus nor do most rollup strategies reach a scale where deploying sophisticated real estate strategies creates much value creation above and beyond operational synergies.

So why not utilize a novel opco-propco strategy to potentially solve this? I’m in the early days of exploration here, but will write more in the future.7

Energy

The energy transition has been bottlenecked by deployment. Here too, private credit is looking to bridge the gap. As Packy notes, when we talk about batteries, we are talking about a deployment problem that somewhat stems from a gap in financing.

For one thing, batteries are expensive. Customers understandably don’t want to pony up $15,000 for something that pays itself back in 12 years even after the 30% IRA tax credit. 

Private equity-backed battery farms like Broad Reach (Apollo), Jupiter (BlackRock), and aypa (Blackstone) on the other hand, are happy underwriting that investment, signing offtake agreements to sell their power, and levering the hell out of it at 70-80% loan-to-value to generate high levered returns.

In that space, Base Power has entered with a bit of a novel model but with the same basic idea: private credit will be fundamental to the company’s success.

For now, Base is financing batteries off of its own balance sheet. In the months ahead, it needs to grow the portfolio and prove out the economics of its batteries so that it can secure asset-backed loans to fuel its growth. Building a portfolio that appeals to lenders means that Base can unlock a lower cost of capital, which means more batteries and more profits, and ultimately even lower costs of capital. The moat widens. 

Across the economy, the question is morphing from “can software transform this industry?” and into a new question: “what is the correct capital structure that enables my team to maximize the leverage ML/AI/hardtech in this industry will give me?'“

Subscribe now