Businesses used 36% more tokens and paid just 2% more
Ramp's latest index shows total token consumption across its businesses up more than a third in a month while spend barely moved. Usage growth is now clearly outpacing spend growth.
Three fresh enterprise AI datasets from Ramp, Databricks and Arcana Research point the same way: price competition has handed buyers real leverage, and the leading companies are using a lot more AI without paying proportionally more for it. Plus: Nvidia's board appears to shut the door on Jensen's return, Commerce finally moves on chip diversion, and OpenAI's embedded safety evaluators fall short of what was promised.
Business AI has become a buyer's market, and the smart buyers are using it.
Across Ramp, Databricks and Arcana Research, the convergent story is stronger than the divergent one. Model competition is pushing token prices down, enterprises are switching and mixing providers far more readily, and the leaders are growing usage much faster than spend. The open question is whether the top 1% of spenders, at hundreds of times the median, are outliers or a preview of where everyone else is heading. That answer will shape the size of the market and whether all this CapEx turns into ROI. For enterprises themselves, the practical takeaway is that you probably have more negotiating leverage than you are using, as long as you've built the infrastructure to switch.
Ramp's latest index shows total token consumption across its businesses up more than a third in a month while spend barely moved. Usage growth is now clearly outpacing spend growth.
Average price per million tokens across OpenAI and Anthropic fell 21% in October. That doesn't necessarily mean companies are moving to cheaper Chinese models. The mere presence of those models is pressuring the big providers, on top of their own push toward more efficient models.
In July, 64% of companies on Databricks' Unity Gateway used one model provider. By early November, 72% used two or more. Databricks argues raw token price is the wrong number to watch, and that the metric that matters is cost per successful task.
Arcana's Global 2000 survey finds 91% expect to spend more on AI in 2026, mostly 20-40% more. Yet two-thirds lack visibility into their model and API spend. That's a sign of the gap between vanguard companies and the middle of the pack, and likely something that changes fast.
In Arcana's survey, a third of OpenAI customers said they'd consider moving, frequently to Anthropic. 43% of GitHub Copilot customers were open to switching, generally naming Claude Code. Switching is becoming normal, and that is what gives buyers leverage.
On Ramp, the median firm spends $16.32 per employee per month, while top firms spend over $10,000, about 625 times as much. Ramp skews tech-forward, but Databricks sees the same pattern: a $2.49 median user against a 99th percentile of nearly $700 a week, a 281x gap.
72% of AI spend on Ramp comes from engineering teams. A third of Arcana's Global 2000 respondents are piloting AI coding tools, the largest pilot group of any category.
Just under 45% of US businesses on Ramp pay Anthropic, versus 40.6% for OpenAI. Anthropic takes 68.8% of proprietary model tokens on Unity Gateway, and 81% of Arcana respondents named it when asked for an exciting AI vendor.
Customer support teams' share of AI spend on Ramp jumped from 0.8% in June to 13.3% in November. 38% of Arcana's Global 2000 respondents already use AI customer service.
Ramp token spend grew 1.05% in August, 0.22% in September and 1.65% in October. If October's rate held, it would compound to about 22% a year. That's a real slowdown from past growth, but still within the 20-40% increase Arcana's respondents expect.
Frontier models' share of Ramp spend rose from 67.7% to 69.5% in October, yet Fable's token share on Databricks fell from 5.5% to 0.6%. Over six weeks, Opus 5.5 climbed from 7.2% to 37.9% of Databricks tokens. The divergence may be less about frontier models in general than Opus simply leaving Fable 5.2 without a place.
Year-end retrospectives will likely crown 2026 the year of agents. Yet Arcana ranks autonomous agents fourth on its heat index of what companies are excited about, and dead last on its adoption index of what's actually deployed.
Each dataset has caveats: gateway-only customers, a narrow tech-forward sample, a small Global 2000 pool. But the convergent story outweighs the divergent one. The big question is how far the median buyer drifts toward the 1% spending hundreds of times more. That will say a lot about market size and whether the CapEx turns into ROI.
Competition is vicious, and model providers are desperate not to become generic commodities. There are deals to be had, but only if you've built the infrastructure to actually switch.
A 180,000-signature petition, shareholder letters from Brad Gerstner and Gavin Baker, and even President Trump ("That certainly isn't a bad idea") backed Jensen's return. On Monday, sources told CNBC the board is focused on Wachtell Lipton's independent investigation into whether his conduct compromised national security or company secrets, and is not discussing bringing him back.
— Sam Altman, at Semafor's event, on AI without Jensen Huang. Altman's answer to fears about losing Nvidia as AI's de facto central bank. Even OpenAI relied on a $100 billion Nvidia guarantee for its Ohio facility, but he insists "people still trust their guarantees very much."
— Jensen Huang, to NeoCloud executives at a Berlin meeting last month, as reported by The Wall Street Journal. Per The Wall Street Journal, Jensen was already pulling back on commitments and demanding upfront payment from NeoCloud bosses before his resignation. Interim CEO Colette Kress is expected to rein in data center backstops further, and Needham warns of a more uncertain credit environment for NeoClouds.
A long-awaited rule targets Chinese labs reaching Nvidia chips through third-party countries. Data center operators in Singapore, Malaysia, Thailand, Indonesia and Brazil now need export licenses for advanced AI chips, and existing facilities have 90 days to get them, which lets Commerce investigate prior shipments.
The new rule says it was written "by and at the direction of" Treasury and the NSC. Sources say Bessent pushed it through with Rubio's support. Trump framed the split as one secretary who wants to be friendly with China and one who wants a blowup. A former Commerce official couldn't recall another agency distancing itself from its own rule, and warns the visible infighting invites workarounds.
Commerce is expected to name its first AI national champions next week: reportedly Nvidia, Amazon, Microsoft, OpenAI and Anthropic. The designation brings expedited export licenses, government help with international business development, and federal financing. A core goal is bundling chips, models and products for international sales.
Apollo Research, METR, Transluce and Goodfire get "deep access" to internal models. Apollo's six-person team will spend four months testing whether models subvert alignment and control mechanisms. But OpenAI can withhold information, limit staff access, keep evaluators off Slack, and censor reports for confidential details. Conclusions must still be published unchanged.
— Apollo Research, on its embedded evaluation agreement with OpenAI. The evaluator itself called the pilot only "an ambitious first step." DeepMind's Neel Nanda said it falls short of the employee-like access Altman promised in September, and ex-OpenAI researcher Steven Adler criticized what he called a "tattletale clause" routing employee concerns through OpenAI.
Even if Anthropic's version is better, labs are still choosing their own evaluators and deciding what access they get. That's reasonable for a business, but its incentives differ from the societal need auditors are supposed to serve. Embedded independent auditing is one of the rare safety ideas with broad consensus, so it deserves agreed-upon standards set by a genuinely independent force.
Logins, ATMs and transfers failed across JPMorgan Chase, Bank of America, Wells Fargo, Zelle and more, traced back to Chase's backend systems. Chase blamed a software bug in on-premises hardware, not a cyberattack. After Microsoft and FAA outages, public sentiment is growing more incredulous of the denials, and AI's role in that narrative is one to watch.