The Discussion · AI on the Deal Desk · Unit Economics
What does a token really cost? Nine questions on the economics underneath AI.
Are AI labs selling tokens below cost to get us hooked? A widely shared analysis says mostly not — and the arithmetic, the comments underneath it and the margins it implies tell investors more about the AI business than most earnings calls.
BY ERIK HUDDLETON · EDITOR · NOVEMBER 2026 · 16 MIN READ
There is a story about AI pricing that almost everyone in the deal community has heard and many quietly believe. It goes like this: the big labs are burning investor money to sell tokens below cost. They are the Uber of compute — subsidise now, get everyone dependent, raise prices later. If that story is true, every AI-heavy business plan in a data room is built on a price that will not last, and every valuation built on those plans is wrong.
This week, the statistician and software developer Hadley Wickham tested that story in a post on his Substack, AI for data science, titled “How much do tokens really cost?”. He took an open model he uses daily, Kimi K3, priced it across every provider selling it, then costed out what it would take to buy a server and serve the model himself. His conclusion: per-token API prices look like honest prices, not subsidised ones. The subsidies, he argues, are somewhere else.
The comment thread underneath was almost as interesting as the post, with readers pushing the analysis into closed-model pricing and lab-level margins. We think this is one of the more useful pieces of public arithmetic on AI economics this year, so we treated it the way we would treat a management presentation: we asked nine questions, checked the maths, and added the view from the deal side.
How to read this piece
This is an interview-style discussion of a published analysis, not a conducted interview. Answers draw on Wickham’s post and on comments left publicly beneath it, with direct quotations attributed. Where we have re-run or extended the numbers, we say so. Our own view sits in the dark boxes marked Pivot’s take.
The analysis in four numbers
$0.25
Estimated cost to serve one million tokens on a self-hosted server at 75% utilisation
$1.13
What mainstream providers charge per million tokens on a typical agentic workload
~78%
Implied gross margin on serving at that price (our calculation)
10–50x
Token value a $200/month subscription can deliver, per SemiAnalysis estimates cited in the post
Part one · The market price
Start with what a token sells for.
01
Why use an open model to answer a question about frontier AI pricing?
Because an open model is the only place where you can see a real market. Anyone can download Kimi K3 and sell access to it, so dozens of providers do. That turns the question “what is the right price?” into something observable. Wickham chose Kimi K3 because he has used it heavily for months — and, tellingly, kept using it after Anthropic released Claude Opus 5.5. He finds Opus 5.5 slightly smarter, but keeps reaching for Kimi K3 because of what it costs.
“Kimi K3 is a reasonable proxy for a solid frontier model: maybe it’s not quite as good, but it’s pretty darn close.”Hadley Wickham, AI for data science
Pivot’s takeFor anyone valuing an AI business, this is the right instinct. When you want to know whether a price is sustainable, find the closest thing to a commodity version of the product and look at what a competitive market charges for it. Comparable-company analysis, applied to compute.
02
So what does the market actually charge?
Wickham pulled pricing for 21 providers selling Kimi K3 through OpenRouter. Because input, output and cached tokens are all priced differently, he built a “blended” price using a ratio typical of agentic coding: 20 input tokens for every output token, with 95% of input served from cache. The spread is wide — roughly 3x from cheapest to most expensive — but there is a cluster in the middle that matters.
| Provider | Input | Output | Cached input | Blended |
|---|---|---|---|---|
| Wafer (cheapest) | 0.28 | 9.60 | 0.13 | 0.59 |
| Relace | 0.21 | 9.33 | 0.21 | 0.64 |
| Fireworks | 2.10 | 10.50 | 0.21 | 0.79 |
| DeepInfra | 2.85 | 14.25 | 0.28 | 1.07 |
| Amazon Bedrock, Together, Modal, BaseTen, Parasail, Moonshot AI | 3.00 | 15.00 | 0.30 | 1.13 |
| Alibaba | 3.45 | 17.25 | 0.34 | 1.30 |
| AkashML (most expensive) | 1.20 | 14.00 | 1.20 | 1.81 |
Source: OpenRouter pricing as published in Wickham’s post, November 2026 (7 of 21 providers shown). Blended price assumes 20:1 input-to-output and 95% of input from cache. We re-ran the blend for the $3.00 / $15.00 / $0.30 tier: (19 × 0.30 + 1 × 3.00 + 1 × 15.00) ÷ 21 = $1.13 — confirmed.
03
Couldn’t the cheapest sellers just be burning venture money?
Some probably are, and Wickham says so. But the more revealing fact is the cluster: six providers, including Amazon Bedrock and the model’s own developer, all landing on exactly the same price at the top of the range. A subsidy war doesn’t usually produce six large, well-funded sellers converging on one number. A cost floor with a healthy margin on top does.
Pivot’s takeRead the cluster like a pricing table in a sector report. When the strongest balance sheets in a market all charge the same thing, that is usually the price at which the business makes sense for a serious operator. The discounters below it are fighting for share; the cluster is telling you where the economics settle.
Part two · The cost of production
Then work out what it costs to make one.
04
What does it actually take to serve a frontier-class model yourself?
A great deal of hardware. The model’s developer recommends at least eight Nvidia GB300 GPUs. Wickham priced one suitable machine, an 8U Supermicro GPU server, at about $550,000. Financed over three years at 8%, that is roughly $17,000 a month. Renting the eight GPUs at around $4 an hour each lands in the same place once hosting is included. Then the server needs a home: colocation, bandwidth and power add about $5,000 a month, and smaller items take the total to about $23,000.
He is upfront that he used Claude to help assemble the hosting estimates and checked the individual facts himself — an unusually honest disclosure, and a model for how this kind of back-of-envelope work should be presented.
| Line item | Assumption | $ / month |
|---|---|---|
| Server financing | $550,000 server, 8% interest, 36 months | ~17,200 |
| Colocation | Rack space in a data centre | ~3,750 |
| Power | Eight high-end GPUs running continuously | ~700 |
| Bandwidth | Serving API traffic | ~600 |
| Other | Smaller items | ~750 |
| Total | ~23,000 |
Source: Wickham’s estimates. We checked the financing line with a standard loan formula: $550,000 at 8% over 36 months gives a payment of about $17,240. The “other” line is our balancing figure to reach his ~$23,000 total. Excludes staff.
05
And how many tokens does $23,000 a month buy?
Using SemiAnalysis’s estimate of about 6,000 tokens per second per GPU, eight GPUs produce 48,000 tokens a second — around 124 billion tokens in a 30-day month. At 75% utilisation, a reasonable figure for a busy API provider, that works out to about $0.25 per million tokens. Compare that with the $1.13 the mainstream cluster charges and the gap is large.
“Assuming we have high capacity, it seems like the cost of running the server is significantly cheaper than what providers are charging.”Hadley Wickham
That “assuming” carries a lot of weight, so we stress-tested it.
Our stress test
Utilisation is the whole business.
Pivot Capital Review calculation: $23,000 ÷ (124.4 billion tokens × utilisation). Break-even against the $1.13 cluster price is about 16% utilisation; against the cheapest $0.59 seller it is about 31%. Ignores staff, resale value of hardware and economies of scale.
Pivot’s takeThis chart explains the entire inference industry. A serving business is a fixed-cost asset that only becomes profitable once it is busy. Above roughly a sixth of capacity, a $1.13 price is very profitable; below it, the same hardware loses money. That is why providers fight so hard for volume, why the cheapest ones can afford to undercut, and why a single company self-hosting for its own occasional use almost never beats buying from a provider. It is also the first thing to diligence in any company that owns GPUs: not the hardware list, the utilisation curve.
Part three · Where the subsidy really is
The money is being lost somewhere. Just not where people think.
06
If per-token prices are honest, is anything being subsidised at all?
Yes — subscriptions. Wickham points to SemiAnalysis estimates that a $200-a-month plan can deliver somewhere between $2,000 and $10,000 worth of tokens at API prices. Those plans are sold to individuals and small businesses. Enterprises pay per token.
“I just don’t think they’re on per-token API costs. Instead, they’re happening on subscription plans.”Hadley Wickham
He adds a second, larger subsidy his own numbers leave out: training. A model like Kimi K3 may have cost several hundred million dollars to train, paid for by its developer’s investors. The serving price recovers the cost of running the model, not necessarily the cost of making it.
| Buyer | How they pay | Subsidised? | What it means |
|---|---|---|---|
| Individual / small team | Flat monthly subscription | Often heavily | Heavy users can get 10–50x the plan price in tokens |
| Developer on an open model | Per token, competitive providers | Rarely | Price close to cost plus a margin; shop around |
| Enterprise on a closed model | Per token, list price or committed spend | No | Pays full price — procurement is the main lever |
| Model developer’s investors | Fund training runs | They are the subsidy | Training cost sits outside the token price |
Pivot Capital Review summary of the arguments in Wickham’s post.
07
Then why are closed models so much more expensive than open ones?
This was the sharpest challenge in the comments. Carl Boettiger argued that the open-model result is close to economics 101 — anyone can host the model, so competition pushes price toward cost. Closed models face no such pressure. In his comparison, a closed frontier model at around $25 per million output tokens sits against Kimi at roughly $10 and another open model of broadly comparable quality at under $0.70.
“The real question then is not why the prices are so low, but why closed model prices are so high, no?”Carl Boettiger, commenting on the post
He offers three possible answers: closed models may be much larger and costlier to run, labs may be earning arbitrage profits on locked-in customers, or both. He also notes that subscription pricing for the same closed models has been estimated to be dramatically cheaper per token than the API at full utilisation.
08
Can we estimate what a frontier lab actually earns on a token?
Another commenter, Mark Edmondson, tried, using the revenue and compute figures he took from Anthropic’s IPO documents for the first half of 2026: roughly $16.3 billion of revenue against about $8.9 billion of compute. The answer depends entirely on how much of that compute goes to serving customers versus training new models — which is exactly the number labs don’t disclose.
| Scenario | Revenue | Inference cost | Implied margin |
|---|---|---|---|
| All compute is inference (worst case) | 16,286 | 8,945 | ~45% |
| Half of compute is inference | 16,286 | 4,473 | ~73% |
| 30% inference / 70% training (as reported) | 16,286 | 2,684 | ~84% |
Source: calculation by commenter Mark Edmondson, based on figures he attributes to Anthropic’s IPO filings. Not independently verified by Pivot Capital Review; we checked only that the arithmetic is internally consistent. Even the worst case shows serving revenue well above serving cost.
Pivot’s takePut Tables 2 and 4 side by side and the picture is consistent: whether you cost a single server or reverse-engineer a frontier lab’s filings, serving tokens looks like a high-margin business. The capital intensity lives in training and in long-term compute commitments. For investors, that splits the AI stack into two very different risk profiles — an inference business that looks like a utility with excellent margins, and a training business that looks like a pharmaceutical R&D pipeline with enormous, lumpy bets.
“It’s plainly not the Uber model that the ‘tokens are subsidised’ position implies — different economics entirely: supply constrained, not demand.”
Mark Edmondson · commenting on Wickham’s post
Part four · The deal-side view
If tokens are fairly priced, procurement becomes strategy.
09
What should CFOs, founders and investors actually do with this?
Stop waiting for a price collapse to fix the AI line in the budget. If per-token prices broadly reflect real costs, they will drift down as hardware improves — Boettiger expects serving costs to keep falling as chip supply expands — but they are not a teaser rate about to snap back, and nor are they a bubble about to burst in the buyer’s favour. Enterprise buyers, who pay per token and get none of the subscription subsidy, are paying something close to a fair market price. That shifts the question from “will prices change?” to “are we buying well?”.
For most businesses that means four levers, in roughly this order: route each workload to the cheapest model that does the job, cache aggressively, compare providers the way Wickham did, and treat committed cloud capacity and credits as an asset to be managed rather than a sunk cost. That last point is the one finance teams most often miss. A large share of AI start-ups run on Azure, frequently funded by six-figure credit grants from Microsoft’s startup programmes. Because those credits buy tokens at full per-token prices, they carry real economic value — and because they expire on a fixed schedule, value that isn’t used is simply lost. Teams that pivot, switch to an open model or move workloads to another cloud are increasingly choosing to sell Azure credits they will not consume through a secondary marketplace, rather than letting the balance lapse — subject, as always, to the terms of the programme that issued them.
Pivot’s takeIn diligence, ask for three numbers on any AI-heavy target: cost per task by model, the share of inference running on committed or credit-funded capacity, and when those credits or commitments expire. A company burning through a credit grant can look far more profitable than it will be in eighteen months. A company sitting on unused credits it can’t consume has a small, recoverable asset hiding on its balance sheet.
The procurement playbook
1
Right-size the model
Send each workload to the cheapest model that meets the quality bar. Wickham’s own habit — reaching for a cheaper model even when a smarter one exists — is the discipline.
2
Cache the repeats
In Wickham’s blend, 95% of input is cached at a tenth of the full price. Workloads without caching pay several times more for the same output.
3
Shop the providers
For open models, the same tokens cost from $0.59 to $1.81 per million depending on who serves them — a 3x spread for an identical model.
4
Manage credits as assets
Track committed spend and credit balances against expiry dates, and recover value from capacity you won’t use before it lapses.
Reader questions
Quick answers on AI token economics.
Are AI companies selling tokens below cost?
For per-token API pricing, the evidence in this analysis says mostly not. Open-model providers appear to charge well above the cost of serving. The clearest subsidies are on flat-rate subscriptions and on the cost of training models.
What is a blended token price?
A single price per million tokens that weights input, output and cached tokens by how a real workload uses them. It is more useful than list prices, which can mislead because output tokens often cost many times more than input.
Is it cheaper to self-host a large open model?
Only at high, sustained utilisation. On the numbers above, a server needs to be busy roughly a sixth of the time just to match mainstream API prices, before paying any staff. Most companies never get there.
Why do subscriptions look so much cheaper than the API?
Because heavy subscribers can consume far more in tokens than they pay for. Providers accept that on individual and small-team plans; enterprises pay per token instead.
Will token prices keep falling?
Serving costs are likely to fall as more efficient chips arrive and supply expands. Whether closed-model prices follow depends on competition from open models, not just on cost.
How should a buyer value an AI company’s compute position?
Look at cost per task, how much inference runs on credits or committed capacity, and when those expire. Credit-funded margins can flatter results until the credits run out.
Source: Hadley Wickham, “How much do tokens really cost?”, AI for data science (Substack), 3 November 2026, and public comments beneath it by Carl Boettiger and Mark Edmondson. Provider prices are OpenRouter figures as reported in that post; throughput estimates are SemiAnalysis figures as cited there. This is an interview-style discussion of published material, not a conducted interview, and none of the people quoted has reviewed it. Utilisation, break-even and margin calculations marked as ours are Pivot Capital Review’s own and are illustrative.
Pivot Capital Review is an editorial publication. Nothing in this article is investment, legal or tax advice.