NVIDIA B300 servers sell for $1M
- Nvidia’s top-end B300 AI servers are now changing hands in China for about 7 million yuan each as U.S. curbs squeeze supply. - The striking detail is the premium — roughly $1 million per system, nearly double late-2025 levels and well above comparable U.S. pricing. - This matters because export controls now distort hardware markets, pushing China toward scarcer gray-market imports and domestic substitutes.
AI servers are the shovels in the generative-AI gold rush. And in China, one of Nvidia’s newest shovels now costs about as much as a mansion. That is the real story here — not just that B300 systems are expensive, but that the same box can carry radically different prices depending on which side of the export wall it lands. In late April, industry sources said Nvidia’s B300 servers in China were selling for around 7 million yuan, or roughly $1 million, after tighter U.S. controls and a crackdown on smuggling choked supply. (money.usnews.com) ### What is a B300 server, exactly? This is not a normal “server” in the old enterprise-IT sense. Nvidia’s GB300 NVL72 is a rack-scale AI system built around 72 Blackwell Ultra B300 GPUs and 36 Grace CPUs, tied together with high-speed NVLink and designed for giant training and inference jobs. Basically, it is closer to a mini supercomputer than a standard data-center box. (nvidia.com) ### Why is the price in China so weird? Because China is not buying from a normal, open market. The reported 7 million yuan price is nearly double late-2025 levels, which tells you this is a scarcity story as much as a demand story. When supply is constrained by policy, logistics, and enforcement, buyers stop paying “list price” and start paying whatever clears the market. (money.usnews.com) ### What changed on the supply side? Two things stacked on top of each other. First, U.S. export controls kept tightening around advanced AI hardware bound for China. Second, authorities also pressed harder on the gray channels that had been moving restricted chips and systems into the country. So the back door got narrower just as front-door access was already limited. That is how you end up with a million-dollar rack. (money.usnews.com) ### Why are buyers still paying up? Because top-end compute is now a strategic input. Chinese cloud companies, labs, and model developers still want the fastest systems they can get for training large models and, increasingly, for inference workloads that need h(money.usnews.com) models doing more work per query. (nvidia.com) ### Is this just Nvidia winning harder? Not really — at least not cleanly. Nvidia benefits from demand for its hardware, but export restrictions also block it from serving a huge market directly. The company already disclosed a $5.5 billion charge in April 2025 tied to new U.S. licensing requirements for its China-focused H20 chip. So the premium in China do(nvidia.com)d reflects friction, risk, and middlemen. (morningstar.com) ### Why does the gray market matter so much? Because it was acting like a pressure valve. If official channels are restricted, unofficial ones keep some price discipline in place by moving supply where demand is hottest. Once enforcement tightens, that pressure valv(morningstar.com)ed lanes. (money.usnews.com) ### What does this mean for China’s AI market? It means fragmentation. Some buyers will still pay up for imported Nvidia systems. Others will redesign workloads around weaker or older chips. And the longer this lasts, the more room domestic alternatives get to improve. That is the catch with export controls — they can deny access in the short term, but they also create a big incentive to build around the restriction. (money.usnews.com) ### So what is the bottom line? A $1 million B300 server is not just a price tag. It is a signal that advanced AI hardware is no longer one global market with one clearing price. It is becoming a split market shaped by geopolitics, enforcement, and scarcity — and that split is getting more expensive.
