Two different questions, one blurred answer
"Is the company profitable" and "is a served token profitable" are different questions. Most of the public argument about AI economics answers one while sounding like it answered the other.
Company P&L
- Training runs for the next model, often costing more than the last one
- Research salaries and equity for the teams chasing that next model
- A capital buildout race for GPU capacity ahead of confirmed demand
- Free tiers, promotional credits, and sales & marketing spend
Unit economics
- The electricity and GPU-hours consumed while generating one reply
- Measured per million tokens, in and out
- Compared against what that reply was sold for, by the API price list
- Nothing about R&D, headcount, or next year's cluster
A company can lose money overall while every token it serves clears a healthy margin, the same way a retailer can run stores at a profit while opening too many new ones to show a profit that quarter. This site is only about the right-hand column: what a served token costs, and what it sells for.
The strongest argument against, made fairly
Before the receipts, the case against. Two arguments get cited constantly. Neither one is nonsense, and neither one actually claims what people think it claims.
OpenAI's cost of doing business is two or three times what they are making in revenue… that's just the inference, that's before training.
Zitron's sharpest specific claim is that heavy users on flat-rate subscriptions (Anthropic's consumer plans, in his framing) burn somewhere between $8 and $13.50 in token cost for every dollar of subscription revenue disputed figure. Reasoning models make this worse: a query that used to cost a few cents can now spend far more tokens thinking before it answers.
Separately, Sequoia's David Cahn has tracked a widening gap between what the industry needs to earn to justify its GPU spend and what it actually earns. By his math, that gap is roughly $840 billion by the end of 2025.
Notice what's missing from both. Nobody here is saying it costs more to generate a token than that token sells for on a usage-based API. Zitron's argument is that flat, all-you-can-eat pricing is mismatched to how much heavy users actually consume, a pricing problem. Cahn's argument is that the industry is building GPU capacity faster than revenue is arriving to fill it, a capital-spending problem. Both can be entirely true while a served token, at the price it's actually sold for, still turns a profit.
What it costs to run the model yourself
Open-weight models are the cleanest test case. Anyone can rent the same GPUs the model's own maker uses and do this math themselves. Several people have.
DeepSeek, by its own disclosure self-disclosed
In March 2025, DeepSeek published its own numbers: running V3 and R1 inference on H800 GPUs leased at $2/GPU-hour cost roughly $87,072 a day, serving 608 billion input and 168 billion output tokens. At list-price API revenue, that same day of traffic would have earned about $562,027, a 545% markup over compute cost.
Kimi K3, three independent estimates cross-checked
Kimi K3 (Moonshot AI's frontier open-weight model) sells through DigitalOcean's serverless inference at $3 per million input tokens and $15 per million output tokens. Three separate parties have each modeled what it costs to self-host it instead: DigitalOcean itself, an independent engineer, and the cost-tracking site TokenCost. They don't agree on the number, because they don't agree on the hardware tier or the utilization rate, but every single one lands profitable at realistic scale.
| Source | Hardware | Monthly cost | Breakeven volume |
|---|---|---|---|
| DigitalOceanmanaged dedicated inference | 8× H100 | ~$22,100/mo$30.32/hr | not modeled |
| Market self-host estimaterented GPU capacity | 8× H100-class | ~$27,800/mo$38/hr | ~23.6B tok/movs. $9/M blended |
| cefboud.comindependent engineer analysis | 8× B300 | $1.37/M tokcompute cost, Pareto-optimal | not modeled |
| TokenCost.appindependent cost tracker | 8× B200/B300 | ~$42,000/mo1.4TB VRAM node | ~2B tok/movs. $15/M output |
Run the math yourself
The harder case: labs that don't publish their costs
OpenAI, Anthropic, and Google don't disclose per-token compute cost the way DeepSeek did, or the way an open-weight self-host estimate can be independently reproduced. Everything in this section is inferred, not confirmed. Treat it as directional.
We're profitable on inference. If we didn't pay for training, we'd be a very profitable company.
Dario Amodei has made a similar distinction for Anthropic: the marginal cost of serving a request is low relative to price, but the company spends roughly half its compute budget on research for the next model, which is what eats the margin at the company level rather than at the token level.
The most useful sanity check here is the crudest one. Independent analyst Martin Alderson ran napkin math on GPT-4o-class serving: input processing at roughly $0.001 per million tokens, output generation around $3 per million tokens, against list prices several times higher. His own conclusion: even if that estimate is off by 3×, inference still comes out comfortably profitable.
So, is inference profitable?
Yes, by every reconstruction of the numbers that exists, from a company that showed its own books to three independent analysts who each modeled the same open-weight model differently and still landed on the same side of zero. The skeptics' strongest specific claims, read closely, aren't actually about this. They're about subscription pricing that doesn't account for heavy users, and about a company spending faster on training and GPU capacity than revenue is arriving to match it. Both of those are real problems. Neither one means a served token is sold at a loss.