Watching the Wrong Race

The model wars are loud right now. Claude releases something, Codex answers within weeks, Gemini quietly leapfrogs both on a benchmark nobody had heard of last quarter. DeepSeek and Qwen ship open weights that sit roughly three to six months behind the frontier at something like a tenth of the cost. DeepSeek itself acknowledged the three to six month trajectory gap in its V4 release notes, which is a refreshingly honest thing for a lab to say out loud.

It’s genuinely fun to watch. It also matters less than people think.

The interesting story isn’t the leapfrog. It’s that the model layer is quietly turning into a two-tier commodity market while the real enterprise dollars are migrating somewhere else entirely.

The Two-Tier Market

The frontier labs are converging on a similar shape. A handful of closed players trading the lead every few months, with the gap between them narrowing to single digits on most benchmarks anyone cares about. Underneath, a Chinese-led open source tier that’s structurally three to six months behind and an order of magnitude cheaper. That’s not a stable equilibrium where one side wins, that’s a permanent split. The frontier captures the workloads where the last 5% of capability is worth paying 10x for. Open source captures everything else.

For founders building application-layer startups, the thin wrappers sitting on top of any single model, this is bad news compounding from both directions. Your underlying capability gets cheaper every quarter, which sounds great until you remember your competitors get the same discount. And the moment your wrapper actually demonstrates a viable workflow, the model lab ships it as a native feature. The wrapper layer is being squeezed from both sides… costs falling and capabilities being absorbed upward. Neither force is something a startup can outrun.

This is roughly where my hesitation on AI application investments comes from. I keep looking at the AI application layer and seeing a vertical with no obvious place for durable defensibility. Speed to market doesn’t compound. UX doesn’t compound. The model underneath you compounds, but you don’t own it.

What the Frontier Just Told Us

This month, both Anthropic and OpenAI announced the same thing. New enterprise services companies, majority-owned by the labs, backed by private equity consortiums, designed to embed forward-deployed engineers inside mid-market enterprises to deploy frontier models into core operations. Anthropic went first on May 4 with a $1.5 billion venture alongside Blackstone, Hellman & Friedman, Goldman Sachs, Apollo, General Atlantic, GIC, Sequoia, and Leonard Green. OpenAI followed on May 11 with $4 billion and 19 backers led by TPG, Advent, Bain Capital, and Brookfield, plus an immediate acquisition of Tomoro to bring 150 forward-deployed engineers in on day one. Bloomberg actually reported OpenAI’s fundraising hours before the Anthropic announcement broke, which means both labs were racing to publish the same structural conclusion at the same time.

Now look at those backer lists together. Blackstone, H&F, Apollo, General Atlantic, GIC, Sequoia, Leonard Green, TPG, Advent, Bain Capital, Brookfield, SoftBank, Warburg Pincus, BBVA. That’s not a capital stack, that’s a distribution list. The combined commitments are rounding error for firms managing trillions in aggregate. What they actually bought is captive AI integrators with privileged access to their portfolios, and they locked competitors out of those same accounts in the process. The labs get the deployments. The PE firms get to inflate portfolio company multiples through “AI transformation” before exit. The forward-deployed engineers are the cover story. The distribution is the asset.

The other detail that should make every founder pause is the consulting side. Anthropic’s announcement names Accenture, Deloitte, and PwC as existing Claude Partner Network members. OpenAI’s launch brings Bain & Company, Capgemini, and McKinsey in as founding investors. Fortune framed Anthropic’s move as a shot at the consulting industry. That reading didn’t hold up for a week. The consulting industry didn’t get disrupted, it got invited inside before the door closed. The forward-deployed engineers, the PE distribution, the consulting reach… all of it is being assembled into a single stack that owns the model, the implementation, and the customer relationship.

Two frontier labs reaching the same structural conclusion within a week of each other isn’t a coincidence. It’s the new consensus. The labs have the best enterprise models on the market by most measures, and they still concluded that the model alone isn’t enough to capture the opportunity. The model is no longer the product. The deployment is the product, and the model is the lever that makes it work. That’s a direct shot at every startup currently positioned between a frontier model and a Fortune 1000 buyer.

Why This Is Structural, Not Cyclical

The interesting question is why mid-market enterprises can’t just buy AI off the shelf. The answer is that legacy fintech, insurance, healthcare, and industrials are sitting on something they’ve been undervaluing for thirty years, which is their proprietary data. Underwriting histories, claims data, customer interaction patterns, decades of records that nobody outside their four walls has ever seen. That data is the actual asset. The model is just the tool that makes it productive.

Connecting that data to AI safely is genuinely hard. You can’t put it on an open source model without losing the proprietary edge. You can’t ship it to a startup without legal and competitive risk you can’t fully underwrite. You can’t build it in-house because every quarter the frontier moves and your internal team falls another generation behind. So you hire forward-deployed engineers from the frontier lab itself, backed by capital partners who can write nine-figure checks, with the big consultancies already inside the tent. The new ventures aren’t competing with Accenture and McKinsey. They’ve absorbed them. What they’re competing with is the in-house AI team you keep failing to build, and now they have the consultants and the PE network handing them the customer list.

The Markdown Nobody Wants to Take

If both frontier labs are vertically integrating into the customer relationship, every Series A wrapper backed in the last twenty-four months is now sitting upstream of two competitors with structurally lower CAC, native model access, and PE-channel distribution. That’s not a thesis correction. That’s a markdown event a lot of funds haven’t taken yet.

The 2023 and 2024 application-layer venture vintage was underwritten on the assumption that the model layer would stay neutral plumbing. The model layer just announced, twice in one week, that it isn’t going to. Marks will lag because nobody wants to be the first to write down a hot AI position, but the structural deterioration is already there. You can either price it in now or wait for the next round to do it for you.

Where This Leaves Me

I’m not investing in AI applications right now. AI isn’t overhyped, but the layer on top of it is. The startups building thin tools on top of someone else’s model are getting squeezed from both sides. The better bets sit underneath. Proprietary data. Infrastructure the models actually need. Domain expertise that survives the next model release. Those compound. Wrappers don’t.

The filter I’m running on every AI deal right now is a single question. Does Anthropic’s or OpenAI’s services arm make this company more valuable or less? If the answer is less, pass. That screen kills most application-layer wrappers in one pass and surfaces the real opportunities. Companies that own data the frontier labs can’t get without paying for it. Infrastructure that sits underneath the deployment. Knowledge of a specific business or industry that no model release can shortcut. The deals worth doing are the ones where the frontier lab becomes a customer or a partner, not a competitor.

The leapfrog will keep being entertaining. Claude will pass Codex. Codex will pass Claude. DeepSeek will ship something at a third of the price and force everyone to recalibrate. None of that is the real game.

The real game is who becomes indispensable to the companies actually doing the work. Anthropic and OpenAI just told us they want that seat, and they brought $5.5 billion to take it.