NVIDIA B300 servers cost $1M
- Nvidia’s top-end B300 AI servers are now changing hands in China for about 7 million yuan, or roughly $1 million, after supply tightened sharply. - That is nearly double late-2025 pricing and far above similar U.S. systems near $550,000, with smuggling crackdowns choking the gray-market supply chain. - The squeeze matters because Chinese buyers still need frontier AI capacity, so scarcity is now inflating hardware prices and cloud access.
AI servers are the physical bottleneck behind the whole generative-AI boom. They are the racks full of Nvidia chips that train models, run inference, and basically turn money into compute. Now the newest version of that bottleneck is getting much more expensive in China. Nvidia’s B300 servers are reportedly selling there for about 7 million yuan — roughly $1 million each — because demand is still intense while legal and gray-market supply both got tighter. (money.usnews.com) ### What is a B300 server? The B300 is part of Nvidia’s Blackwell family — the company’s newest high-end AI hardware. This is not a consumer PC part and not even a normal data-center GPU in the old sense. It is a full AI server platform built for giant training jobs and heavy inference workloads, which is why buyers talk about whole systems and racks, not just single chips. China wants these because they are the fastest path to more model capacity. (money.usnews.com) ### Why did the price jump now? The immediate trigger seems to be a supply crunch. Industry sources told Reuters that prices in China have nearly doubled as a crackdown on chip smuggling dried up black-market supply. That matters because a meaningful share of advanced Nvidia gear has been reac(money.usnews.com)higher fast. (money.usnews.com) ### Why is China paying such a premium? Because China is trying to buy frontier AI compute in a market where the legal route is restricted and the unofficial route got riskier. Reuters says a comparable B300 server in the U.S. is around $550,000, so the China price implies a huge scarcity premium. That premium is not just for the hardware. It is also payment for procurement risk, logistics complexity, and the chance that the next shipment never arrives. (money.usnews.com) ### Where do U.S. export controls fit in? They are the backdrop for the whole story. Nvidia’s H20 — the China-focused chip designed to comply with earlier rules — was hit with a U.S. license requirement in April 2025, and Nvidia said it expected up to $5.5 billion in related charges. Even thou(money.usnews.com)tighten with little warning. (techcrunch.com) ### Is this just about one expensive server? Not really. A $1 million sticker is the visible symptom. The deeper issue is that compute in China is getting harder to source predictably. If you are a cloud provider, model lab, or enterprise trying to build AI products, you care less about one headline price than about whether you can sec(techcrunch.com)uipment remains strong, which means the squeeze can spill into rentals and cloud capacity too. (money.usnews.com) ### Does this help Nvidia? In a narrow sense, scarcity reinforces how central Nvidia still is. But the catch is that extreme shortages and export barriers also create openings for substitutes. The more painful procurement gets, the more incentive Chinese buyers have to use domestic alternative(money.usnews.com)r stress. (money.usnews.com) ### So what is the real takeaway? This is what an AI infrastructure war looks like when it reaches the hardware aisle. The important number is not just $1 million. It is the gap between what the machine costs in an open market and what it costs in a constrained one. That gap tells you compute is no longer just a product — it is a geopolitical asset. (money.usnews.com)
