New data dropped: AI spend per token has jumped 13-14x year over year for companies already paying for APIs. And yet, AI is still less than 1% of total business spend excluding payroll. That gap is where the next reckoning lives.
I was glued to my screen when I caught this conversation with Ara Khazarian, chief economist at Ramp, breaking down what their 50,000-business, $100B-spend dataset actually shows. The mind behind this research sees every receipt, every model choice, every API call going through Ramp cards. That is a kind of visibility no public dataset offers.
Here is the headline finding the original poster shared with the Wall Street Journal: Anthropic overtook OpenAI as the default choice for businesses buying AI for the first time. The flip happened inside three weeks in January. Before that, OpenAI was winning new buyers roughly 60/40. Now Anthropic leads.
Why did the switch happen?
- 🔹 Claude Code locked in the technical, high-intensity users first, then Claude expanded to non-technical workflows.
- 🔹 OpenAI spent years optimizing for consumers and was slower to refine enterprise use cases.
- 🔹 The Department of Defense labeling Anthropic a security threat did not slow adoption. If anything, growth accelerated.
Three practical applications from the data
- Budget for blowups. Uber burned through its annual AI budget in a single quarter. If you are forecasting AI spend on last year’s curve, you are already wrong. Build elastic budgets and review monthly.
- Route smarter, not richer. Most business spend still flows to the most performant models, but the savvy professional running this analysis expects that to shift hard. Tools like Open Router and Cursor route prompts to cheaper, more appropriate models. Haiku and Gemini Flash handle plenty of tasks Opus does not need to touch.
- Treat models as products, not commodities. The expert pushed back on the commodity framing. Switching costs are near zero technically, but stickiness is real. Developers stay on Claude Code even when it crashes and rate-limits them, which tells you product experience matters more than raw price.
Tips and pitfalls
- Tip: Track adoption intensity, not just adoption. Are teams renewing? Using multiple models? Spending more on tokens month over month? Those are the leading indicators.
- Tip: Watch the non-model vendors. Harvey for legal, ElevenLabs for voice, vertical tools layered on top of the labs. The economist flagged renewal rates here as the real signal of market maturity.
- Pitfall: Assuming Google is losing. Ramp only tracks paid spend. Gemini ships free through Workspace, so adoption is significantly underrated in the chart. Google is quietly powering tons of AI-native product experiences via Flash.
- Pitfall: Betting on local open-source hosting long term. The contributor doubts companies will keep up with frontier development by self-hosting. Hosted platforms running open models is the more realistic path.
One stat to sit with: adoption is on track to hit 90% within 1 to 3 years, conservatively. After that, the question shifts from who is using AI to how intensely, on which tasks, and whether productivity stats finally move.
Check out the full conversation for the deeper breakdown on Google’s quiet position, why Figma is still crushing it despite Claude Design, and what Ara expects from the next round of pricing pressure.