Nvidia B300 Servers Hit $1M on China's Grey Market — 82% Premium as Export Curbs Meet AI Token Surge
Nvidia B300 servers are clearing China’s grey market at $1 million per unit. Reuters reported the figure Thursday, sourcing it to buyers and intermediaries operating outside official channels. The same hardware lists at $550,000 in the United States, making the grey market premium 82% — roughly $450,000 per box.
The gap is structural, not speculative. US export controls block direct sales of Nvidia’s H100 successors to Chinese buyers, which compresses legal supply to near zero. Demand has not compressed at all. Chinese firms’ share of global AI token consumption grew from 5% to 32% between January and March 2026, according to data cited in the Reuters report — a 6x jump in a single quarter that reflects the scale of domestic model deployment now running on whatever hardware can be obtained.
What a B300 Server Actually Buys
A single B300 server carries 8 GPUs performing 14 quadrillion FP4 operations per second and 288GB of high-bandwidth memory. At that throughput, even $1 million amortises quickly against token revenue at current Chinese inference pricing. Companies that cannot justify the purchase price are instead renting: the going rate is 190,000 yuan per month, roughly $26,000, for access to a single server.
The economics are distorted enough that legal actions have followed. US authorities have pursued criminal cases against individuals involved in diverting restricted chips into grey market channels, which has tightened rather than loosened supply — each enforcement action removes a node from the distribution network without reducing the underlying demand.
The Token Share Story Is the Real Number
The 5% to 32% figure is the harder-to-explain data point. Chinese AI workloads were a marginal fraction of global token traffic at the start of 2026. By the end of Q1 they represented nearly a third. That is a demand signal of a different order than anything the export control framework was designed around.
The shift reflects deployment, not experimentation: production inference for consumer products, enterprise tools, and developer APIs running at scale across Chinese cloud providers. Models like DeepSeek V4 Pro, Qwen3.6, and Hy3 are operationally competitive with frontier western models and are being served at volume. They need compute to run.
Enforcement Versus Economics
Export controls create a price floor on illegal supply, not a ceiling on demand. At $1 million per server, the grey market is functioning — just expensively and riskily. The controls have successfully raised the cost of accessing restricted hardware; they have not stopped Chinese AI infrastructure from scaling.
The question now is whether the 82% grey market premium represents a temporary arbitrage that will collapse as restrictions loosen, or a permanent tax on Chinese AI deployment that gradually widens the compute gap between US-supplied and restricted-supply model training. The token share data suggests the latter outcome has not materialised yet.