NVIDIA B300 servers near $1M

AI servers are the boxes companies buy when they want serious model training or high-end inference at scale. Nvidia’s new B300 systems sit near the top of that stack. Now the weird part — in China, those machines are reportedly selling for close to $1 million, or about double what similar systems cost in the U.S. That is not a normal product-cycle markup. It is what a market looks like when policy, scarcity, and demand all pile onto the same rack. ### What is a B300 server? The B300 is part of Nvidia’s Blackwell Ultra generation. In Nvidia’s own DGX B300 configuration, the system uses eight B300 GPUs, comes with 2.1 TB of total GPU memory, and is built for both training and inference in dense AI workloads. Nvidia pitches it as “AI factory” gear — basically, not a gamer card, not a workstation, but a full enterprise server meant to run the expensive part of modern AI. (nvidia.com) ### Why is China paying so much? Because this is not just a product market anymore — it is a constrained access market. Reuters reported that B300 servers in China have climbed to nearly $1 million, versus about $500,000 in the U.S., with traders and distributors tying the spike to U.S. export controls and tighter supply. When th(nvidia.com)e hardware starts carrying a scarcity premium on top of the normal Nvidia premium. (thenextweb.com) ### What changed on the export side? The broader backdrop is that Washington kept tightening the rules around advanced AI chips headed to China. In April 2025, Nvidia disclosed that the U.S. government had imposed a license requirement on H20 shipments to China, and Nvidia said it expected a $5.5 b(thenextweb.com)a AI Diffusion Rule in May 2025 but paired that move with fresh guidance aimed at strengthening chip-related export controls. So the rulebook changed, but the pressure on China’s access to advanced Nvidia compute did not go away. (techcrunch.com) ### Why does gray-market flow matter so much? Because gray channels are what smooth out shortages when official channels get blocked. If those channels slow down — whether from enforcement, arrests, or just higher risk — prices jump fast. One recent account tied the B300 s(techcrunch.com)logic is straightforward: fewer paths in means every available rack gets repriced upward. (thenextweb.com) ### Is this just about one flashy server? Not really. The B300 is the headline because it is new, powerful, and scarce. But the real story is what happens when frontier AI hardware becomes segmented by geography. A company in one market can buy near-list, wait for delivery, and plan capex normally. (thenextweb.com) to keep a cluster build on schedule. Same silicon family — completely different economics. (nvidianews.nvidia.com) ### What does this do to Chinese AI buyers? It turns infrastructure planning into a budget problem first and a technical problem second. If a rack costs twice as much, teams either buy fewer systems, delay deployments, or shift to whatever domestic alternatives they can actually (nvidianews.nvidia.com)ing is possible, but it is not painless — it is more like swapping the engine after you already designed the car around it. (thenextweb.com) ### Does Nvidia still benefit? In the short run, high resale prices show how strong demand for Nvidia hardware still is. But Nvidia does not automatically capture the gray-market markup, and export limits can still shut off large chunks of direct revenue. That is why the company has kept building C(thenextweb.com)blem is who gets to serve it, and under what rules. (morningstar.com) ### Bottom line A near-$1 million B300 server is not just a crazy price tag. It is a signal. Advanced AI compute is no longer one global market with one rough clearing price — it is splitting into separate markets shaped by export law, enforcement risk, and local desperation for GPU capacity. (thenextweb.com)