The Substrate Inversion
What the labs are actually buying
The frame that held for three years
For three years the question that has organized AI commentary, capital allocation, and competitive positioning has been some version of which model wins. Whose model is at the frontier this quarter. Whose handles tools better. Whose is cheaper per token. Whose has the longest context. Each benchmark release was treated as a meaningful event. Each leaderboard shift was treated as a rebalancing of the competitive map. Underneath all of it, mostly unspoken, was the assumption that the model is the asset, model capability is the moat, and the rest of the AI economy is downstream of who has the best model.
The frame held up for a long time. Through 2023 and most of 2024 model capability moved fast enough and was hard enough to replicate that the lab with the best model could plausibly compound into the largest enterprise business. The model was scarce. Access to the model was the product. Customers paid for tokens. And for a while, the valuations matched the assumption.
Then sufficient capability for many enterprise workflows became cheap and abundant faster than the old frame expected. Stanford’s 2025 AI Index documented an approximately 280-fold drop in inference cost for fixed performance over an 18-month window. Sam Altman has written that the cost of a given level of intelligence falls roughly tenfold every twelve months. Open and open-weight models also narrowed the practical gap, making sufficient capability harder to treat as scarce. The question shifted. Who alone could supply the intelligence stopped being the binding constraint. Who could make the available intelligence work inside a specific organization started to be.
Harvey is the clearest single signal. The legal AI company at eleven billion dollars in valuation now describes itself as multi-model by design, routing across leading foundation models because no single model is best at every task. The model became swappable, and the business kept growing. If you sit with what that fact implies, the conclusion is uncomfortable. Whatever Harvey actually owns, whatever justifies its valuation, is not its model.
You can read this as a story about commodification. Most coverage has read it that way. Margins compress, the competitive set widens, the labs need to scale faster to compensate. Inside that frame, the rational response for a lab whose model is commoditizing is to push harder on capability, push harder on cost reduction, and find more enterprise customers willing to pay for incremental capability advantages while they still exist. Inside that frame, the labs should be doubling down on the model. They are not.
What the labs actually did on May 4
Now look at what the labs actually did on May 4.
Anthropic, with Blackstone, Hellman and Friedman, and Goldman Sachs, announced a one-and-a-half-billion-dollar joint venture to build a new enterprise AI services company. Bloomberg reported the same day that OpenAI had finalized a separate four-billion-dollar vehicle, internally called DeployCo, with a private equity syndicate of nineteen investors anchored by TPG, Brookfield, Bain, Advent, and Goanna. The first-order reading was that both deals were the labs moving up the stack into services revenue, or the labs becoming consultants. Both readings sit comfortably inside the old frame. Both assume the labs are doing whatever they are doing in service of monetizing model capability more effectively. Inside the old frame, the announcements are distribution channels for tokens. The announcements describe something else.
Anthropic’s own announcement gives a different picture if you read it slowly. A typical engagement starts with a small team of engineers from the joint venture working alongside Anthropic Applied AI staff inside a customer company. In the healthcare services example given in Anthropic’s own materials, this means engineers sitting with clinicians and IT staff to identify where time disappears in workflows like documentation, medical coding, prior authorization, and compliance reviews, then building tools around the answer. The clinicians know where time disappears and what good patient care requires. The engineers build around that knowledge. Reuters reported in early May that both ventures are pursuing acquisitions of AI services and consulting firms specifically to add hundreds of engineers and consultants who can tailor models to company-specific data, systems, and workflows.
That is the substrate.
The detail the old frame explains too thinly
There is one detail in DeployCo’s reported structure that the old frame explains only as expensive distribution. That explanation is not wrong. It is incomplete.
OpenAI reportedly guaranteed its private equity backers a 17.5% annualized return over a five-year window. The Financial Times reported it. Reuters and Bloomberg cite the FT figure but did not independently confirm. If the deal underperforms, OpenAI is reportedly the one who pays.
The reported 17.5% floor is the structural feature worth sitting with. It should not be read as a clean forecast of DeployCo’s expected return. It is better read as a price signal. OpenAI was reportedly willing to give PE investors a high fixed-return floor, while retaining control through super-voting shares, in order to secure capital, speed, and access to PE-controlled distribution.
Access is the point. Private equity does not just bring money. It brings operating companies where workflow changes can be pushed from the boardroom, measured in EBITDA, and rolled across portfolios. If the model were the whole asset, OpenAI’s cleanest strategy would be to sell access to the model at the highest possible price. A subordinated deployment vehicle makes much more sense if the position around the model captures value the model alone cannot.
The substrate inversion
This is the substrate inversion.
The model is no longer the only asset, and at the enterprise application layer it may no longer be the primary one. The scarce asset is the substrate that turns model capability into trusted operational change. Forward-deployed engineers translating tacit operational knowledge from senior practitioners into systems. Evaluation rubrics that capture what good output looks like in this firm in this vertical. Routing logic that picks which model handles which kind of task. Permissions, retrieval indices, exception handling, feedback loops where the senior practitioner reviews outputs and the system updates. Without it, capability is a benchmark number. With it, capability is enterprise value.
The substrate is also what compounds. Forward-deployed engineers leave behind deployment patterns the next acquisition can reuse. Evaluation frameworks calibrate themselves over time as more outputs flow through them. The data trace from running the substrate at scale becomes a moat that does not depend on which underlying model produced any particular inference. Models change every few months. The substrate persists.
None of this means the model race is irrelevant. Better models still expand the surface area of what can be deployed, and the labs still hold real advantages in roadmap access, reliability, integration, and trust. The inversion is not that models stop mattering. It is that model advantage increasingly has to be captured through a substrate position. Capability creates the option. The substrate captures the economics.
If you accept the inversion, the financial topology of the May 4 structures becomes easier to read. The expensive part is not the token call. The expensive part is the engineering bench and everything around it. Integration. Security. Evaluation frameworks. Permissions. Workflow redesign. Customer success. Ongoing maintenance. Forward-deployed engineers at lab-JV scale run several hundred thousand dollars per year fully loaded, and a small deployment pod assigned to a customer or portfolio company can become a seven-figure annual cost before anything else is counted. Scale that across dozens of companies in active ramp, and the main cost center becomes obvious.
Reuters reported that most of the capital raised through the OpenAI and Anthropic ventures is expected to fund acquisitions of engineering services and consulting firms. The labs are not organizing billions primarily around inference resale. They are organizing billions around deployment capacity.
Anthropic’s structure does the same thing without the floor. Anthropic’s CFO Krishna Rao framed it in the announcement. Enterprise demand for Claude is significantly outpacing any single delivery model. The bottleneck, at least in this part of the market, is not only model capability. It is delivery capacity, which is substrate capacity by another name. No comparable floor has been publicly reported for the Anthropic vehicle, which may reflect a different structure, stronger negotiating leverage, a different investor mix, or simply less disclosure. But the broader move is the same. Both labs are organizing capital around deployment capacity rather than treating model access alone as the full capture mechanism.
What follows
Four things follow from the inversion. The first three are market-structure consequences. The fourth is the constraint the whole structure runs on.
The first is that the consulting industry is being re-ranked rather than disrupted. Anthropic’s announcement preserved the Claude Partner Network with the major systems integrators. OpenAI launched Frontier Alliances with McKinsey, BCG, Accenture, and Capgemini in early 2026. The integrators are still in the picture. They may even do much of the labor. But the center of gravity changes if the lab-controlled vehicle owns the architecture, the model roadmap, the deployment patterns, and the feedback loops. Consulting does not disappear. It becomes one component in a lab-shaped distribution system.
The second is that the lab business has structurally changed even if the labs are reluctant to say so publicly. The May 4 structures suggest that pure model access is not enough to capture the enterprise surplus the labs need to justify their training capital expenditure. Research remains essential. Better models still create advantage. But commercially, model capability has to be realized through deployment, through the systems, workflows, permissions, evaluations, and organizational change that let AI become part of how companies operate. Anthropic and OpenAI are still frontier research labs. They are no longer commercially legible as pure model companies.
The third is the question every operator and investor downstream of these structures should now be holding. The substrate race is not empty territory. Sierra is already trying to own the customer-experience substrate. Its agents power billions of customer interactions, and it raised nine hundred and fifty million dollars at a valuation above fifteen billion dollars on May 4 itself. Harvey is already trying to own the legal substrate. Cursor and Cognition are already fighting for the coding substrate. The labs are not moving into a blank market. They are moving into vertical positions that application-layer companies have been occupying for two years. The question is no longer whether the substrate is the new asset. The labs answered that with multi-billion-dollar deployment vehicles. The question is whether they can occupy it faster than the application-layer companies that started occupying it before the labs noticed.
The fourth is a level up. The substrate is drawn from senior judgment. The clinician who knows which prior authorizations fail. The lawyer who knows which clause matters. The auditor who knows which anomaly is signal. But senior judgment is produced by years of junior repetition. PwC quietly scrapped its 2021 pledge to add a hundred thousand to its worldwide headcount by mid-2026, and is reducing entry-level hiring across multiple regions. The broader labor-market evidence points in the same direction. Brynjolfsson, Chandar, and Chen found that more experienced workers in AI-exposed occupations remained stable or continued growing. Workers aged 22 to 25 in the most exposed occupations experienced a 16% relative employment decline.
If AI compresses the junior work through which future experts are trained, the first wave of substrate creation may be drawing on a stock of expertise that the next wave is not replenishing. The structure does not only have a moral problem. It has an input-supply problem. The first-wave equity is real. The second wave depends on a pipeline that the same technology is quietly disrupting.
What this leaves us with
I put the receipts in a separate reference document. The deal structures, the dollar flow, the unit-economics assumptions, and the case studies across legal, coding, customer experience, and accounting. It is in the open if you want to see what is actually inside these structures.
Reference document: The May 2026 AI Lab–Private Equity Joint Ventures
The May 4 deals are not just the labs becoming consultants. They are the labs acting as if model access alone is no longer where the enterprise moat sits, and racing capital and engineering bench scale into the position above it before the position closes around someone else.
For three years the question has been which model wins. The answer does not matter the way it used to. The model race still matters. It just no longer answers the commercial question by itself.
Capability creates the option. The substrate captures the economics. The model is the input. The substrate is what compounds.
Concepts Compilation thesis · The linker