The Inference Layer

The Thesis

Two records exist for the same question — who’s winning enterprise AI — and in the summer of 2026 they stopped agreeing.

The dollar record comes from Ramp, whose corporate cards sit inside tens of thousands of US businesses and see the subscription line item before anyone writes about it. In April 2025, Ramp’s own data had OpenAI at 32.4% adoption among its cardholders, widening its lead. By January 2026 new AI-tool purchases had settled into roughly a 50/50 split between Anthropic and OpenAI. By March, Anthropic was capturing about 73% of new-buyer spend, even as its share of all US businesses was still trailing OpenAI’s, 30.6% against roughly 35%. By July, the lagging number caught up: Anthropic’s overall adoption hit 43.5% against OpenAI’s 39.7%, and Ramp called it a widening lead, not a photo finish.

The ink record comes from TEXXR’s own archive, and it says something close to the opposite. Since 2025Q2, OpenAI has drawn 3,086 articles to Anthropic’s 1,381 — better than two to one — and the gap isn’t closing. Quarterly momentum has OpenAI accelerating (velocity +0.22, acceleration +0.22) into 2026Q3 while Anthropic’s coverage is decelerating (velocity -0.12, acceleration -0.33) off a first-quarter spike that hasn’t repeated. The company winning the wallet, by the record’s own count, is not the company winning the coverage.

Both records are real, and reconciling them is not this page’s job — the two are measuring different things, one in dollars committed and one in words written, and there is no rule that says they move together. What the gap does is point at a layer neither record looks at directly: the inference-serving companies that sit between the model labs and the enterprises paying for them — Fireworks, Baseten, SambaNova, OpenRouter — none of which show up as a headline in either the Anthropic-versus-OpenAI spend chart or the Anthropic-versus-OpenAI coverage count, and all of which have raised money faster this year than either lab’s valuation moved. If the model layer is a two-horse race the record can’t settle, the layer underneath it is where the more interesting bet sits.

The Evidence

Start with the number that trades on its own: TEXXR’s archive holds 3,086 OpenAI articles against 1,381 for Anthropic across the six most recent quarters, and the trend line is diverging rather than converging.

Now the money. Anthropic’s own disclosed revenue run rate went from $4B in July 2025 to $9B by December, $19B by early March 2026, $47B by May, and $65B by the end of July — sixteen times in thirteen months. In February, sources told the Financial Times that Anthropic’s own investor guidance put full-year 2026 revenue above $30B; the company blew past its own forecast with five months left on the calendar. That growth showed up in the private market before it showed up on a Ramp card: Anthropic’s $65B Series H in May priced the company at $965B, overtaking OpenAI’s own $852B mark from the same window.

The corporate-card data is the part that arrives with a lag built in, and the lag is informative. the Wall Street Journal reported in October 2025 that corporate accounts already made up roughly 80% of Anthropic’s revenue, with one contemporary estimate putting its enterprise API share at 32% against OpenAI’s 25% — Anthropic ahead on that specific measure nearly a year before Ramp’s broader adoption count agreed. A different measure disagreed in the other direction: an a16z survey of 100 CIOs in February 2026 found 78% running OpenAI models in production against 44% for Anthropic — usage breadth, not spend, and OpenAI still well ahead on it. Three datasets, three different things counted — API revenue share, production deployment breadth, card-spend share — and none of them agree on a single winner, because none of them are measuring the same thing.

What the Ramp numbers add that the others don’t is a split between the marginal dollar and the legacy budget line. In March 2026, Anthropic was already taking 73% of first-time AI buyers’ spend while its share of the existing base of AI-paying businesses sat at 30.6%, a full 12 points behind OpenAI’s ~35%. New budget moved to Anthropic well before the installed base did. It took another four months for the overall number to flip.

Beneath both labs sits the layer that routes and hosts the models neither of them run for the customer directly. OpenRouter went from 5 trillion to 25 trillion tokens a week in six months, raised at a $1.3B valuation in May, and by July was generating ~$140M in annualized revenue — up nearly 3x since April — with Stripe reportedly offering $10B to buy it. Fireworks passed $1B in annualized revenue on a $17.5B valuation. Baseten’s valuation went from $5B in February to $11B–$13B in June, more than doubling in four months. SambaNova raised $1B at $11B and signed JPMorgan as a customer to run its own chips in-house — a bank choosing to deploy inference hardware itself rather than rent a frontier lab’s API. None of these four companies has a ticker. All four grew their valuations faster in 2026 than Anthropic’s grew relative to OpenAI’s.

The open-weight angle behind that growth has its own data point, also from a company in this layer: OpenRouter’s own usage logs, reported by the Financial Times in March, showed lower-cost Chinese models from DeepSeek and MiniMax overtaking their US rivals in token consumption since February — evidence, from inside the routing layer itself, that enterprise buyers route to cost and capability across labs rather than staying loyal to one, which is the condition The China AI Stack tracks from the supply side.

The Companies

Salesforce is the cleanest public proxy for Anthropic’s win, holding a stake reported at roughly $5B that started as a $50 million bet in an early 2023 round. Its exposure isn’t only financial: CoreWeave signed a multiyear deal to supply Anthropic with Nvidia-chip data-center capacity in April, tying the same buildout this page’s AI capex supercycle already tracks to Anthropic’s specific growth curve, not the industry’s in general.

SpaceX carries the other side. Its xAI secondary, rebranded SpaceXAI in July after SpaceX absorbed xAI in an all-stock deal valuing the combined company at $1.25T, showed up on Ramp’s August ranking of the fastest-growing vendors by market share — a name that isn’t new capital reaching the AI trade so much as an existing capital base changing its label and getting counted again. TEXXR’s own record shows the xAI-SpaceX relationship logged first as rumored partnership talk in January 2026, then rumored acquisition, then completed acquisition, then merger — a company’s status in the graph moving through four labels in five weeks before the market and the coverage record agreed on what to call it.

Nvidia sits underneath the inference-serving layer as its chip supplier, and its own relationship with one member of that layer shows what happens when the incumbent decides not to compete for a company, but to hollow it out instead. TEXXR’s record has Nvidia rumored to be acquiring Groq from December 2025 through March 2026, then that acquisition denied, then, by June, a different shape: Nvidia licensed Groq’s technology and hired much of its talent. Groq’s next funding round, in August, priced the company at $3.5B — down from $6.9B eleven months earlier. Not every name in the inference layer is compounding. The one that got closest to Nvidia’s own turf lost half its value to the deal that was supposed to save it.

The Lenses

Hamilton Helmer’s test for a moat is whether a Barrier survives someone actually trying to break it, and the inference-serving layer is built to break exactly the kind of Barrier a model lab would want to claim. If OpenAI or Anthropic held real Switching Costs, a router that moves 25 trillion tokens a week across 400-plus models on a customer’s behalf wouldn’t have a market — a buyer already locked to one lab’s API and prompt scaffolding would find leaving expensive. OpenRouter’s growth argues the opposite: switching, for a meaningful and growing share of enterprise spend, costs close to nothing, because a layer of infrastructure now exists to do it for you. That reframes what Anthropic’s wallet-share lead actually is. Nothing in the record shows a durable Barrier behind it — no Cornered Resource, no Process Power built over years. It reads more like a Benefit without a Barrier yet: real, priced correctly for now, and open to being copied the moment OpenAI’s own product closes the gap that opened it. Nvidia’s move on Groq is the sharper Helmer case in this group — not an acquisition, which would have priced Groq’s technology at what a buyer competing for it might pay, but a licensing-and-hiring deal that got the capability at roughly half the valuation a real bidding contest would have set. That is Cornered Resource logic applied to people and code instead of a physical asset, and the coverage record shows it working exactly as the framework predicts: the target’s next round priced in the damage.

Richard Thaler’s mental accounting explains the gap between the 73% and the 30.6% better than any capability story does. A business buying its first AI subscription in March 2026 wasn’t reallocating existing budget — it was opening a new account, with no anchor set by a prior vendor relationship, and Anthropic won that account at a rate more than double its share of the installed base. The installed base moved slower because it had somewhere to stay: a renewed contract, a procurement relationship, a reference price set by whoever won the deal first. That is Thaler’s inertia finding, the same force that keeps a retirement plan’s default allocation in place long after a better option appears, showing up in enterprise software procurement instead. It also explains the four-month gap between the money and the ink: a newsroom’s sense of “who is winning” is itself a reference price, set years earlier by which company shipped ChatGPT first, and a reference price is exactly the kind of number Thaler’s research says people defend past the point the original reason for it has faded.

What Moved

Sources