Hadley Wickham, writing on his Substack at substack.hadley.nz, has run the numbers on a question many AI buyers ask: are the prices on offer propped up by venture money? His answer, built on a back-of-envelope server budget, is that, for pay-per-use access, they mostly are not.

A quick definition first. A token is the small chunk of text a model reads or writes, usually a word or part of one, and providers bill by the million of them. Wickham says he keeps hearing that labs sell tokens below cost to get customers hooked, then raise prices. He set out to test that with data.

His test case is Kimi K3, an open model from Moonshot AI that anyone can host. He has used it heavily for a couple of months and kept using it after Opus 5.5 arrived, because it is cheaper even though Opus feels a bit smarter to him. Because the model is open, many providers sell it, and Wickham notes that six of them charge about the same at the high end. One is BaseTen, which supplies models for Posit AI Pass, a service his team runs. He suspects some of the cheapest sellers are burning investor cash, but points to six providers pricing about the same at the high end as a sign against deep subsidy.

Then he priced doing it himself. Moonshot’s recommended setup calls for at least 8x GB300 chips. Wickham picks a Supermicro server at $550,000, borrowed at 8 percent over three years, which works out to about $17,000 a month. Renting eight GPUs at around $4 an hour each lands in a similar place. Buying has one edge: the machine might resell for roughly $150,000 after three years.

Housing it adds more. Wickham relies on Claude’s estimate of about $3,750 a month for colocation, $600 for bandwidth, and around $700 for power, and says such prices are rarely published. With smaller extras, the monthly bill reaches about $23,000. He flags that he used Claude to help with the research and that it may have steered him wrong, though he checked individual facts as he wrote.

On output, he cites SemiAnalysis for roughly 6,000 tokens per second per GPU. Across eight GPUs that is 48,000 per second, or 124 billion a month. At 75 percent utilization, which he calls reasonable for an API provider, the server’s cost comes to $0.25 per million tokens. For typical coding-agent traffic (20 parts input to one part output, with 95 percent of input served from cache), the going rate at BaseTen and similar sellers is $1.13 per million.

Wickham is candid about the limits. A real company would get volume discounts but would also pay staff. His conclusion is that prices track true serving costs, with competition squeezing them and still leaving room for a healthy margin.

He does see subsidies elsewhere. Citing a recent SemiAnalysis post, he says a $200 monthly plan can consume anywhere from $2,000 to $10,000 of tokens at list prices, though only individuals and small businesses can buy those plans; enterprises pay per token. His math also leaves out training, which he says could run to hundreds of millions of dollars and is currently paid by Moonshot’s investors, the Chinese government among them. He calls that a fair argument for a subsidy, and hopes governments will fund open models directly.

The analysis comes from one practitioner with no inside access to provider books, so treat the $0.25 as a floor, not a quote. Even so, a gap of more than four to one gives anyone budgeting for AI coding tools a reason to doubt a sudden jump in per-token prices, while flat-rate subscriptions are where his own argument sees the subsidy.

Analysis by Hadley Wickham, published on his Substack at substack.hadley.nz (undated in the source).