Ramp Says 1% of Businesses Drive 80% of OpenAI and Anthropic Revenue in Its Data. Build Like You Are Not the Customer They Optimize For.
Ramp's November AI Index, published November 11 by lead economist Ara Kharazian, has one sentence I keep rereading: "1% of businesses drive 80% of the revenues to OpenAI and Anthropic." In the same data, the median business spent $16.32 per employee per month on AI in October, while the top 1% spent a median of $10,199.20. If you are a solo builder, you are the $16 customer. The two companies you probably depend on most earn most of their money, at least in Ramp's sample, from someone else.
The numbers, without the spin
Ramp's index is built on card and bill-pay data from more than 85,000 US businesses. Here is what it says about October.
AI adoption reached 56.8% of businesses on Ramp, up 1.0 percentage point from September. Total AI spend hit a new high but rose only 1.65% month over month. Anthropic's adoption share was 44.8%, up 1.1 points. OpenAI's was 40.6%, roughly flat. Token spend grew 2.1% for OpenAI and 0.5% for Anthropic. Ramp also says blended token prices fell 21% over the last month.
Ramp adds that the top 1% is "highly correlated, largely made up of VC-backed companies in tech, finance, and technical services." Its top 1% spent about 9% more per employee than in September. The top 10% was roughly flat at $801. The median moved by less than a dollar.
I want to be careful with the 80% figure. Ramp doesn't publish how it was computed. The most sensible reading is that it describes spending by businesses in Ramp's sample, not OpenAI's or Anthropic's total revenue. It is still the clearest picture I have seen of how lopsided paid AI usage is.
Why concentration matters to someone paying $16 a month
Companies build for the customers who pay them. When most revenue comes from a small group of heavy, mostly VC-backed users, a few things tend to follow, and I've watched most of them happen this year.
Rate limits and quotas get shaped around high-volume API accounts and enterprise contracts. Consumer and prosumer plans get renegotiated in public, sometimes by press release.
New features target agent-heavy, high-token workloads: long context, long-running agents, frontier models. Ramp says frontier models (Opus, Fable, Sol, Astra) were 69.5% of spend in October. That's where the roadmap goes.
Price cuts land on the tokens heavy users burn most. That can help you, as the 21% blended drop shows, but it isn't aimed at you, and it can reverse.
None of this is a conspiracy. It is just arithmetic. A vendor deciding between a feature that keeps a $10,000-per-employee customer and one that delights a $16-per-employee customer will pick the first one almost every time.
How I'd build given that
I'm not going to stop using frontier APIs. They are the best tools I have. But I'd design around the fact that I'm not who they're optimizing for.
Keep a swap layer. Every model call in my products goes through one small function with the provider and model name in config. When a price, a quota or a model name changes, I change one line. I've written about this before, and the Ramp data is another reason to keep it.
Use the API, not the subscription, for anything a customer depends on. Subscriptions and plan quotas are where consumer-tier changes show up first. Pay-per-token APIs have clearer contracts and published prices.
Track cost per useful outcome, not per token. If prices drop 21% and my output per dollar doesn't change, the drop went to someone else's workload shape.
Have a cheap fallback that is good enough. For most of my classification and extraction work, a smaller model or an open-weight model is fine. I keep one wired up and tested, even if it's off by default.
Where I could be wrong
Ramp's caveats cut both ways. Ramp itself says its numbers likely underestimate adoption and spend, because free tools and personal accounts don't show up as business payments. That means lots of small users are invisible in this data, so the "1%" might look different if every solo developer paying on a personal card were counted.
There is also an argument that concentration is good for small users. Heavy customers fund the infrastructure that makes the $20 plan possible, and price cuts driven by big buyers still reach everyone. I believe that's partly true. I just don't want my business to depend on it staying true.
The honest take
Read the Ramp index as a description of who these companies work for. It is a small group of heavy spenders, and you are probably not in it. That's fine. Build like it: one swap layer, API over subscription for anything customers touch, and a tested fallback model. Then the next pricing change is an afternoon of config, not a crisis.
Author
Lukas
@lukcombinator