Trend pillar
The Inference Layer
Ramp's card data says Anthropic passed OpenAI in April. TEXXR's own coverage record says OpenAI is still ahead and pulling away. Both are true, and the layer underneath both of them is growing faster than either.
The Thesis
Anthropic Beat OpenAI on Ramp, Not in Coverage
Ramp’s July 2026 US cardholder index puts Anthropic ahead of OpenAI in paid business adoption. TEXXR’s 2025Q2–2026Q3 archive counts 1.50 times as many OpenAI-tagged articles as Anthropic-tagged articles, derived from the counts below. Ramp counts businesses paying for tools; TEXXR counts tagged articles. Neither measures market share.
The divergence is the finding. It does not establish why businesses chose one model provider over another. Inference providers may make it easier to use several models, but the cited records do not connect their customers to Ramp’s cardholders. They also lack comparable revenue, margin, retention, and capital-burden data. Routing is a testable explanation, not this page’s conclusion. The token ceiling tracks a separate question: how much work users hand to machines, rather than how many businesses pay for a subscription.
The Evidence
TEXXR’s archive count is its quarterly series of articles tagged to each company, 2025Q2 through 2026Q3, read 27 October. It measures attention, not customers, usage, revenue, or the frequency of real-world events. The component’s coverage-momentum label describes the direction of quarterly article counts only.
Ramp’s path to the April crossover was not a straight comparison of like measures. OpenAI adoption reached 32.4% of businesses in Ramp’s sample in April 2025, versus 28% in March. Among businesses buying AI tools for the first time, spending was split roughly evenly between OpenAI and Anthropic in January 2026; Anthropic took about 73% of that first-time buyer spend in March. Across all US businesses in Ramp’s sample, Anthropic adoption reached 30.6% in March, versus 24.4% in February and roughly 35% for OpenAI in March. Ramp first put Anthropic ahead on overall adoption in April, 34.4% to 32.3%. In Ramp’s July 2026 index, 43.5% of US businesses using Ramp cards paid for Anthropic, versus 39.7% for OpenAI. The July gap was 3.8 percentage points, derived, versus a derived 2.1 percentage points in April. First-purchase spending and adoption across the cardholder base have different denominators.
Other corporate records count still other things. The Wall Street Journal reported in October 2025 that corporate accounts supplied roughly 80% of Anthropic’s revenue and cited a July estimate of 32% enterprise API share for Anthropic versus 25% for OpenAI. An a16z survey of 100 CIOs in January 2026 found 78% running OpenAI models in production versus 44% running Anthropic models. API share, production deployment, and Ramp adoption cannot be combined into one share figure. OpenAI told investors in April that its earlier compute buildout gave it an advantage; that is the company’s claim, not a measure of the Ramp gap.
Bloomberg reported Anthropic’s annualized revenue run rate at $4B in July 2025, $9B that December, $19B by early March 2026, $47B in May, and $65B by the end of July. The July-to-July change is 16.25 times, derived from Bloomberg’s endpoints. Investor guidance reported by the Financial Times in February put annualized revenue above $30B by year-end 2026; Bloomberg’s reported July run rate was already $35B above that threshold, a derived comparison across different dates. Run rate is not recognized annual revenue. Anthropic’s $65B Series H in May valued it at $965B, versus OpenAI’s $852B valuation in the same report.
What would prove routing mattered. OpenRouter’s weekly token volume rose from 5 trillion to 25 trillion across more than 400 models in six months, a derived fivefold increase; it raised funding at a reported $1.3B valuation in May. The Information reported about $140M annualized revenue by July, nearly three times its April level. Fireworks reported more than $1B annualized revenue in July and raised $1.5B at a $17.5B valuation. Baseten’s June financing used $11B–$13B valuations, versus $5B in February; SambaNova raised $1B at an $11B valuation and signed JPMorgan to deploy its chips in-house. Those records mix volume, run rates, and financing terms. None identifies Ramp cardholders switching labs through a provider.
OpenRouter’s usage logs provide a narrower portability signal: the Financial Times reported in March that lower-cost DeepSeek and MiniMax models had overtaken US rivals in token consumption there since February. The logs do not disclose why buyers moved, whether they appear in Ramp’s sample, or OpenRouter’s margin. The China AI Stack tracks that supply-side contest.
The Companies
Salesforce has financial exposure to Anthropic through a stake reported near $5B, against an initial investment of about $50M in early 2023. It is not a pure Anthropic proxy: Salesforce announced Koa, a CRM reasoning model built on Nvidia’s Nemotron architecture. CoreWeave signed a multiyear deal in April to supply Anthropic with Nvidia-chip data-center capacity. That ties one customer to the buildout tracked by the AI capex supercycle; it does not disclose Anthropic’s inference margin.
SpaceX complicates the model map. It absorbed xAI in an all-stock deal that sources valued at $1.25T for the combined company; xAI was rebranded SpaceXAI in July. TEXXR had logged that relationship as rumored partnership talk in January 2026, then rumored acquisition, completed acquisition, and merger over five weeks. A changing graph label is not fresh capital. Reuters later reported a SpaceXSI rename plan; that was a plan, not a new buyer.
The product and capacity records are more concrete, but have different status. The Information reported OpenAI’s planned counter-features for Grok Bot “teammates”; Bloomberg reported, citing sources, that SpaceX discussed buying startup data for training. The Information also reported Microsoft compute-leasing talks, not a signed contract. By contrast, the Wall Street Journal reported an Anthropic–SpaceX compute deal brokered by Anthropic co-founder Tom Brown at $1.25B per month. That is a derived $15B pace over 12 months if the monthly amount persists. The report does not establish recognized revenue, duration, utilization, or margin. In another direct routing example, The Information reported Musk’s statement that Grok Bot would choose a back-end for each task, including Claude Opus 5.5 and other APIs. Neither example identifies a business behind Ramp’s July lead.
Nvidia supplies chips to this stack, but its Groq record warns against treating every provider valuation as cumulative value. Bloomberg placed Groq’s $350M round at a $3.5B valuation, versus $6.9B in September 2025, before Nvidia later licensed its technology and hired much of its talent. Bloomberg did not attribute the decline to that deal. The Financial Times reported a Delaware lawsuit in which former engineers alleged a $20B non-exclusive Nvidia acqui-hire deal in 2025 short-changed them. CNBC puts Nvidia’s December 2025 payment for Groq assets at $20B; that is not a financing valuation.
The Lenses
Hamilton Helmer would ask whether a model lab’s Barrier survives easier routing. OpenRouter’s volume is consistent with demand for model choice, but does not measure enterprise switching costs. Prompts, workflows, security reviews, and procurement can remain costly to move even when an API offers several models. Ramp establishes Anthropic’s current adoption lead in its sample, not a durable Cornered Resource or Process Power. Groq’s separate financing, licensing, hiring, and disputed deal records likewise show why corporate value cannot be read straight from technical utility.
Richard Thaler offers a different hypothesis for Ramp’s first-buyer and installed-base measures. A new buyer may open a budget line while an existing customer keeps its contract, procurement route, and reference price. That fits Thaler’s account of inertia, but Ramp does not isolate it from capability, price, or contract timing. Nor can TEXXR’s article counts explain why newsrooms covered OpenAI more often. Both hypotheses need buyer-level evidence before they explain the observed gap.
What Moved
As of November 6, Ramp’s July 2026 paid-adoption reading remains the latest here; TEXXR’s comparable OpenAI–Anthropic article series ends in 2026Q3. New reports address capability, routing, and delivery costs, not the size of the Ramp gap.
Nvidia’s balance sheet is part of that distinction. CNBC reported on September 4 that its equity investments were worth $99B, up from about $7B a year earlier. Bloomberg reported, citing sources, that Nscale sought up to $3.5B, including $2B from Nvidia; those were talks. The DeadRisk market layer records Nvidia FY2026 revenue of $215.9B for the year ended January 31, 2026, versus $130.5B in FY2025, and free cash flow of $96.7B versus $60.9B. Bloomberg reported a $150B increase in Nvidia’s buyback authorization in September, from a derived $85B to $235B remaining, with completion planned through FY2028. Capacity to finance the stack does not establish returns from routing.
Nvidia also moved into software and potential model supply. It launched Open Agent Safety in September, a reference design using CPU sandboxes and Nvidia DPUs; no revenue or adoption figure accompanied it. Nvidia agreed in September to buy Hugging Face for $12.9B; CNBC later reported, citing sources, that OpenAI had first offered to invest about $100M, and that AMD and Salesforce also held talks. The Financial Times reported October talks for Nvidia to acquire or invest further in open-weights startup Reflection AI; its sources said an acqui-hire was possible. Talks are not a closed transaction. The Register reported a $1B commitment over five years for US scientific-computing research, not inference-provider revenue.
Financing appetite has a visible limit. Bloomberg reported Firmus shelved its IPO in October and considered a private round. Its account of fund-manager objections put the proposed valuation at $30B against $51M of FY2026 revenue, about 588 times revenue, derived. That is a rejected IPO valuation, not a traded multiple. At the smaller hardware edge, Ghost’s $3,499 agent PC uses an Nvidia GPU; TechCrunch reported an $11M seed round. Neither record supplies deployed inference volume.
SpaceX has competing calls on capital. Bloomberg reported a low-band spectrum agreement for Starlink’s proposed US mobile service, not an AI compute contract. The DeadRisk market layer records FY2025 SpaceX revenue of $18.7B versus $14.0B in FY2024, while free cash flow was negative $14.1B versus negative $5.4B. Its October 12 price was $163.42, versus $171.09 on October 5, a derived 4.5% decline across those dates. Price and a spectrum purchase do not measure demand for Anthropic’s models or the economics of the reported compute deal.
According to The Information’s METR account, METR found Anthropic’s models improving about twice as fast as they did between March 2024 and October 2025. METR attributed roughly 80% of that acceleration to AI use inside Anthropic. But METR said Anthropic redacted much of the most informative evidence from its 100-page report. Readers must rely on METR’s judgment more than it considers ideal. That finding makes capability worth testing as a cause of Ramp’s lead; Anthropic’s redactions limit an outside check.
Anthropic launched Personal Finance in Claude with read-only Plaid connections to bank and credit card accounts for spending analysis. The product adds a workflow but discloses neither model-routing behavior nor a new Ramp business-adoption reading.
Capacity plans still need separating from delivered service. The New York Times reports, citing sources, that Anthropic expects more than 3.5GW of capacity in 2026 and is diversifying beyond Nvidia chips to Amazon, Google, and AMD; the report supplies no prior capacity figure for a growth comparison. Reuters reports a Google credit backstop for TeraWulf’s Kentucky project with Anthropic, scheduled to come online in the second half of 2027. Neither expected capacity nor a future project measures utilization or the cost of serving routed requests.
The provider-side earnings check is narrower. The Wall Street Journal reports Cloudflare’s Q3 2026 revenue at $782.5M, up 39% year over year from approximately $563M, derived, and $45.3M above the $737.2M estimate, derived. Cloudflare reported a $34.1M net loss and raised its full-year 2026 revenue guidance to around $2.96B, versus FY2025 revenue of $2.2B in the DeadRisk market layer. Those consolidated figures do not break out inference revenue or routing margin.
Capital pricing is moving on a separate record. The Wall Street Journal reports investor skepticism that Anthropic can reach a $1.5T IPO valuation after a market selloff; one investor put it closer to $1T. Both are reported views, compared with the completed May Series H valuation of $965B, not new transaction prices.
- 2026-08-17Corrected 27 October: Bloomberg dates Groq's round after Nvidia's licensing deal and talent hires. The Sources-card wording below wrongly implies the reverse order.
- 2026-08-12Corrected 27 October: an earlier version dated Anthropic's first overall Ramp adoption lead to July. Ramp's May index shows it first for April.
Sources
Across 5 member names, 2026Q3 drew 799 articles against 790 in 2026Q2. The largest single move was Cloudflare, -52%.
Coverage data as of 2026-11-22 · the essay above was last revised 2026-11-06