← Noteseconomics
Three numbers for the AI market, and the fine print on each
The AI market produces a record every week, so I find it more useful to pick a few numbers and read the footnotes than to collect headlines. Three tell most of the story right now: who is putting money in, how much the labs are taking out, and what intelligence costs per unit.
1. The money going in
Global venture funding hit a record $510 billion in the first half of 2026, more than all of 2025 (Crunchbase). In the US, PitchBook counts 86% of venture dollars going to AI companies (SiliconANGLE).
The fine print: OpenAI and Anthropic alone took $217 billion, 43% of all startup funding in that half (Crunchbase). A record that concentrated says less about “startups” than about two companies. And trackers disagree on what counts as an “AI company”, which is why the AI share runs from about 70% to 86% depending on who you ask.
- All startups510
- OpenAI + Anthropic217
- Everyone else293
Table
| Item | US$ billions |
|---|---|
| All startups | 510 |
| OpenAI + Anthropic | 217 |
| Everyone else | 293 |
Source: Crunchbase
2. The money coming out
In early October the Financial Times reported that OpenAI told investors its annualised revenue was approaching $50 billion. A week before, a figure close to $70 billion had been circulating (TechCrunch).
Nothing changed in the business between those two numbers. The higher one was investors’ attempt to compare like with like: Anthropic counts sales made through its cloud partners as its own revenue, OpenAI doesn’t (same source). Anthropic’s run rate had passed $65 billion by the end of July. The market still took the smaller number badly. AI infrastructure stocks fell on the day (CNBC).
Two more things to keep in mind. These are run rates, a month multiplied by twelve, not audited annual revenue. And the company that sells the shovels is the one with audited numbers: Nvidia closed its fiscal 2026 with $215.9 billion in revenue, $193.7 billion of it from data centres, at a gross margin above 70% (Nvidia).
3. The price of intelligence
Epoch AI estimates that since 2023 the cost of reaching a fixed level of performance on a benchmark has fallen about 47% per quarter, roughly 13× per year (Epoch AI). Right after a capability first appears, the price falls even faster.
The fine print, from the authors themselves: it assumes everyone switches to the cheapest model that is good enough, which real users rarely do. Benchmark scores aren’t useful work, and the data cover about three years. Cheaper per token also doesn’t mean a smaller bill, because agents use far more tokens per task than a chat does.
How I put them together
Money in is concentrated in two labs. Money out is real but reported in ways that flatter it. And the unit price falls so fast that today’s revenue buys a capability that will be nearly free in two years. That last point is the one I’d watch: if intelligence keeps getting 13× cheaper, the durable value is less in the model and more in the data, the evaluation and the product around it.
Figures as of 30 October 2026 from the linked sources. Run rates and private valuations change quickly. Check before quoting.