NVIDIA B300 selling for $1M

A B300 server selling for around $1 million in China sounds like Nvidia gouging the market. It mostly isn’t. The real story is that export controls have turned top-end AI hardware into a geography problem. The same box can have one price in the U.S. and a radically higher one in China because legal supply, service access, and delivery certainty are now part of the product. DIGITIMES pushed that story on May 1, and it fits a pattern that has been building for more than a year. ### What is a B300, exactly? The B300 is Nvidia’s Blackwell Ultra generation for data centers — the really expensive end of AI infrastructure, not a gaming card and not even a normal enterprise GPU. Nvidia pitches DGX B300 systems as “AI factory” hardware for training and inference, with performance gains over the prior B200 generation. In other words, this is the kind of machine bought by cloud providers, model labs, and companies trying to build serious AI capacity fast. ### So why would China pay so much more? Because the price is not just for silicon. It is for scarce, hard-to-replace, hard-to-import compute that may not be legally available through normal channels. Once a market is cut off from direct supply, every remaining route gets more expensive — distributors charge more, integrators charge more, and buyers pay up for certainty. A $1 million quote is basically the all-in scarcity price for access, not a clean list price comparison with a U.S. buyer ordering normally. ### Didn’t export controls reduce demand? Turns out they often do the opposite at the top end. They reduce legal supply faster than they reduce demand. China’s AI companies still want frontier-class compute, especially for training and large-scale inference, so restrictions create scarcity rents instead of making the market disappear. That dynamic already showed up with Nvidia’s H20 — the most advanced chip Nvidia could legally sell into China before Washington tightened rules again — when H3C warned in March 2025 about looming shortages amid strong customer demand. ### What changed with H20? The U.S. government told Nvidia on April 9, 2025 that H20 exports to China would require a license, and Nvidia disclosed that change in an SEC filing. Nvidia later said it expected a $5.5 billion hit tied to H20 inventory and purchase obligations. That matters because H20 had been the compliant release valve. Once even that path narrowed, the incentive to chase higher-end hardware through indirect channels got stronger, not weaker. ### Is there really a gray market for this stuff? Yes — and that is the part people keep underestimating. By July 2025, reporting tied to Financial Times and Reuters said at least $1 billion of Nvidia AI chips had entered China despite tighter U.S. curbs, with banned B200 processors described as widely available on a black market. That does not mean every quoted B300 server is smuggled. But it does show the mechanism: when official supply closes, parallel supply networks appear. ### Why doesn’t Nvidia just raise prices everywhere? Nvidia has been saying allocation is first-come, first-served rather than pure highest-bidder wins. If that is true, then Nvidia’s own pricing is only part of the story. The big premium gets created downstream — by resellers, brokers, import risk, and customers bidding against deadlines. Basically, scarcity is being monetized outside Nvidia’s list price. ### Why does this matter beyond one expensive server? Because it shows export controls are reshaping where compute lands, not simply whether it gets bought. The premium in China is a signal that frontier AI hardware still clears the market there — just at a much worse price and with worse certainty. That means longer lead times, more arbitrage, and more incentive for Chinese buyers to fund domestic alternatives from Huawei and others. DIGITIMES already expects China’s cloud AI accelerator market in 2026 to tilt heavily toward localization. ### Bottom line The $1 million B300 story is really about scarcity economics. Export controls did not kill demand for Nvidia’s best systems in China. They made access itself expensive — and turned geography, compliance, and supply-chain workarounds into part of the price tag.