NVIDIA B300 sells for $1 million

AI servers are the steel mills of the current tech boom. They are the machines companies buy when they want to train big models, run inference at scale, and stop renting every ounce of compute from someone else. Now one of Nvidia’s newest systems — the B300 — is reportedly selling in China for about $1 million. That is not just a rich-company flex. It is a sign that export controls are reshaping the market in real time. ### What is the B300, exactly? The B300 is part of Nvidia’s Blackwell Ultra generation — basically a top-end AI server platform built around eight B300 GPUs in a DGX system, or larger HGX-based configurations for data centers. Nvidia pitches it as a step up from the B200, with higher dense FP4 and attention performance, which matters for newer reasoning models and heavy inference workloads. (nvidia.com) ### Why is the China price so wild? Because this is not a normal open market. Reuters says B300 servers are trading in China for around 7 million yuan, roughly $960,000 to $1 million, and that this is close to double typical U.S. pricing for comparable systems. Once a product becomes hard to import legally, the price stops reflecting only manufacturing cost and starts reflecting risk, scarcity, and who can still get units across the border. (nvidia.com) ### Is this an official Nvidia sale? Probably not in the clean, ordinary sense. The whole point of the story is that reduced gray-market supply is helping push prices up. That means the relevant market is not just Nvidia’s list price or an OEM quote. It is the effective street price inside China after middlemen, logistics, regulatory risk, and limited inventory all get layered on top. The catc(nvidia.com)RP. (nvidia.com) ### Why did supply tighten now? Because U.S. restrictions on advanced AI chips have kept ratcheting tighter, especially for products aimed at China. Over the last few years, Nvidia has had to redesign offerings for the Chinese market more than once, and every new rule narrows the set of chips and systems that can move freely. Even when a specific server is not the headline target, the broader (nvidia.com)source. That is how you get shortages before a machine is even common. (trendforce.com) ### Why would anyone still pay it? Because the alternative may be worse. If a Chinese cloud provider, lab, or startup needs frontier-class compute now, waiting six months can cost more than overpaying for hardware today. In AI, being compute-constrained can mean slower model training, weaker products, and lost customers. A million-dollar server sounds absurd — but if it unlock(trendforce.com)Does this mean Nvidia wins anyway? In one sense, yes — scarcity keeps the product desirable. But it is not a clean win. Export controls can reduce official sales, fragment distribution, and push business into channels Nvidia does not fully control. They also encourage Chinese buyers to look harder at domestic alternatives and at software techniques that squeeze more work out of weaker hardware. High gray-market prices are a symptom of demand, but also of market distortion. ### What does this say about the AI race? Basically, compute geography now matters almost as much as compute itself. The same Nvidia box can carry one price in the U.S. and a radically different one in China because policy has become part of the bill of materials. That changes who can build, how fast they can scale, and where the next generation of AI infrastructure gets concentrated. ### Bottom line? This is the new shape of the AI hardware market — not just faster chips, but politically segmented ones. A $1 million B300 in China is what happens when the world’s most wanted server runs into borders.