The most important price signal of 2026 is not on a rate card. It is in a capacity lease: OpenAI's deal with Cerebras, roughly 750 megawatts of dedicated inference capacity, reported at about 10 billion dollars of value, priced at roughly 0.47 per megatoken of reserved throughput. Lease the plant. Sign the term. Subsequent to it, Cerebras went public (CBRS, market cap in the ~23 billion range at its listing) - the first large public company whose underlying asset, in the earnings call's own framing, is a megatoken. Infrastructure that used to be a secret operating cost is now a listed thing.
Reading the number: 0.47 per megatoken
The 0.47 per megatoken figure is a capacity-clearing price: what a buyer pays for the right to draw throughput at stated utilization over a multi-year term, with the full stack - power, silicon, site, financing, amortization - bundled into the number. Two things follow that are worth sitting with:
- It is the industry's full-cost floor, in writing. The wholesale reserve is the sum of the parts: power at contract rate, HBM and accelerator amortized over the term, land and cooling, working capital. The public market's spot prices - 1 to 10 per million tokens for common classes - are, in effect, the spread above this reserve. When the reserve is written in a contract, "below cost" stops being a feeling and starts being a delta you can compute against a published number.
- It tells you the direction of the spread. The reserve is priced for a utilization assumption that earns the buyer a return at the stated throughput. The marginal cost of serving one more token on the leased plant is a sliver of that - mostly power - which is exactly why the spot market keeps landing below the reserve and keeps landing there faster each year. The reserve is a ceiling on cost inflation; the spot is a contest about how fast you converge to the power floor, per the power-contract math.
Doing the energy math on 750MW
The megawatt number is the only part of the deal you can recompute yourself, and it is worth doing once:
- 750 MW at 100% load is 6,580 MWh per hour - about 57.5 TWh at a full year. At a realistic sustained utilization you land in the low tens of TWh.
- At a combined-rate power contract in the 4-6 cent per kWh range, the pure electricity content of that plant runs to hundreds of millions of dollars a year. Call it a quarter to a third of the deal in steady state.
- What the rest of the 10 is: the silicon, the site, the financing, the term. That remainder is exactly the part the spot market is racing to amortize away, and it is the part that used to be invisible.
The listed-token-asset consequence
A public company whose capacity is reserved by contract, whose backlog is megatokens, whose 10-K names inference throughput as a line item: this is the shape of an energy company, circa 1990. The consequences for the rest of the market are slow but real:
- Analysts will start valuing inference capacity the way they valuate enrichment capacity - megawatt-to-throughput ratios, utilization, reserve pricing. The "cost of inference" arrives in earnings calls, and it comes with a number.
- Credit markets learn to underwrite GPU-heavy sites. The cost of building the floor falls, and the floor falls.
- Your provider comes to its repricing decision with a public cost structure - the era of repricing as mystery is shortening.
What a buyer should do with this
- Put the 0.47-per-megatoken reserve in your cost model as the "full stack" reference and compute your current blended rate against it - the delta is your margin cushion, and it is a number per model class now, not a mood.
- Watch the 2026-2027 capacity contracts as the industry's price discovery. Each new lease is another data point on the reserve curve, and the curve is the only thing that sets the long-run floor.
- Treat the spot-versus-reserve spread as a timing signal: it is wide because the amortization is still running, and it contracts with every utilization leap, per the per-token pricing explainer.