Tokens vs Bare Metal: GPU Economics

Since March, the rent on a B200 rose about 18% while the price of the tokens it produces fell by a third to four fifths, depending on the model. That spread decides how GPU fleets get financed.

A GPU owner has two ways to sell the same hour of silicon. Rent the machine: a bare-metal or reserved contract at a fixed rate per GPU-hour. Or sell what the machine makes: inference tokens on open-weight models, listed on an aggregator next to a dozen competitors. In mid-2026 the two businesses are moving in opposite directions, and the gap between them is the most instructive price signal in the compute economy.

The token side looks irresistible at first. OpenRouter alone routes on the order of 22 trillion tokens a week to its top six open-weight models, models anyone with GPUs can serve without permission. DeepSeek carries about a fifth of all requests on the platform. The sales channel is a listing, not a sales force, and the demand churns violently toward whatever model is best this month; the number three model by volume grew roughly 500% month over month, and a top-ten model did not exist five weeks ago.

Then the price record ruins the romance. Within a model generation, competition destroys price almost immediately. Xiaomi's MiMo-V2.5 fell about 84% in 66 days on OpenRouter's own price history. MiniMax's M2.5 lost about 60% in four months. DeepSeek V4 Flash dropped about 36% in the same 66-day window. GLM 5.2, released in mid-June, already has 27 providers quoting it. Each new generation resets prices upward, and each reset is competed away in weeks. A late entrant earns the cheapest-provider floor, not the median, and the largest serving operators are reported to realise a blended rate that is a small fraction of sticker prices once caching and batching are counted.

Meanwhile the machine itself got more expensive to rent. Reserved B200 rates on public listings rose from about $3.80 to $4.47 per GPU-hour between March and July. RTX PRO 6000 reserved rates more than doubled. B300 on-demand went from around $7 to near $11. The same hardware is being paid more as a rental while its output gets cheaper, which means the market is currently pricing scarcity of machines, not scarcity of tokens.

Two prices for the same machine. Indexed to March 2026 = 100. Reserved B200 rent on public listings against blended token prices for three open-weight models on OpenRouter. Sources: archived OpenRouter price history; public GPU rental listings. Paths between observations are interpolated; indicative.

Put the two prices on one GPU and the arithmetic is uncomfortable. Using published production benchmarks for DeepSeek-class serving on B200, a fleet running flat out at today's floor token prices grosses about $4.84 per GPU-hour. The reserved rental rate for the same GPU is about $4.47, and the rental carries no utilisation risk, no price-decay risk within the term, and no serving-platform cost. Run the same arithmetic as a utilisation question and the regimes separate cleanly: at floor prices, token revenue matches the rent only around 92% utilisation; at median provider prices, around 59%; at the premium prices of a model's first weeks, the bar drops to roughly 19%. Tokens only clearly beat the rental in one regime: the first weeks of a new model's life, when prices are three to ten times the eventual floor and even modest utilisation clears the bar. The token business, examined closely, is not a yield product. It is a speed contest in operations, won by whoever stands up each new model first.

The floor arithmetic, $/GPU-hr. Token gross at cheapest-provider floor prices, from published DeepSeek-class serving benchmarks on B200 at the benchmark traffic mix, versus the reserved rental rate for the same GPU. The token bar assumes 100% utilisation and is before channel fees and serving costs. Indicative.

Utilisation needed for tokens to match the rent. The utilisation at which B200 token revenue equals the $4.47 reserved rate, at each pricing regime. If floor prices also keep falling at recent rates, the floor case moves above 100% — unreachable. Derived from public benchmarks and listed prices; indicative.

For this desk, the interesting part is what the spread does to credit. A reserved rental contract is collateral: it has a counterparty, a tenor, a payment schedule, and, if papered properly, assignment language that lets a lender step into the revenue. The standard neocloud capital stack — debt advanced against signed offtake plus the machines — exists because that contract exists. Merchant token revenue has none of those properties. We are not aware of any lender, in India or globally, that advances meaningfully against uncontracted token sales. A fleet bought to sell tokens is, in practice, an equity-funded fleet, and equity that must outrun a price curve falling tens of percent per quarter.

The market's own behaviour confirms the ranking. Token-first operators raise equity to buy hardware and data centres; infrastructure owners add token layers on capacity they already control; nobody at scale stays pure-play on the merchant side. The stable configuration across DeepInfra, Fireworks, Together and Nebius is the hybrid: contracted metal as the financed core, tokens as an opportunistic layer on capacity that would otherwise idle, where the marginal cost is roughly power.

Two things would change this analysis, and both are worth watching from India. The first is token offtake: enterprises pre-buying committed token volumes on contract, which would give inference revenue the tenor and assignability that debt requires. The second is vendor revenue backstops of the kind Nvidia has begun extending to mid-tier cloud operators, which substitute a rated counterparty for merchant risk. Either would make token cash flows bankable. Until one of them arrives, the rule for a financed fleet is simple: the machine's contract is the asset, and what the machine makes is a trading business run on the side.

The machine, or what the machine makes