NVIDIA B30A

The NVIDIA B30A is a China-market Blackwell accelerator that has been reported in detail but never released, and it is best understood as an artefact of United States export policy rather than as a product. Reuters first described it on 19 August 2025, naming it tentatively as the B30A and reporting that it would use a single-die design taking one of the two dies from the dual-die package of the flagship B300, which would likely give it about half the raw computing power of that part. Two features were reported as retained, and both matter: high-bandwidth memory, and NVLink for fast chip-to-chip transfer. NVLink is the contentious one, because it is what allows accelerators to be clustered into a machine large enough to train a frontier model, and earlier China-specific Blackwell proposals had reportedly stripped both HBM and multi-GPU interconnect out for exactly that reason. Positioned above the H20 that NVIDIA was then permitted to sell in China and well below the B300, the chip was expected to cost roughly double the H20 and to reach Chinese customers as samples in September 2025. It never did. In November 2025 the White House informed federal agencies that it would not allow the B30A to be sold to China, and NVIDIA was reported to be revising the design in the hope that the decision would be revisited. When the Commerce Department eased its licence review policy in January 2026, moving the H200 and AMD's MI325X to case-by-case review, Blackwell-class parts were not included, and in May 2026 Jensen Huang told CNBC that NVIDIA had largely conceded China's AI chip market to Huawei and had told investors to expect nothing on approvals. NVIDIA's own public position throughout has been to neither confirm nor deny: it says it evaluates a variety of products for its roadmap so it can compete to the extent governments allow, and its chief executive has said only that a follow-on to the H20 is not the company's decision to make. No performance figure on this page is derived from the widely circulated estimates, because those estimates do not agree: within one week of the original report, one outlet published 7.5 petaFLOPS of FP4 by halving the B300 and another published 3.5 petaFLOPS by halving the B100 instead, and neither matches half of the B300's actual dense FP4 rate. The naming is also unresolved, with the same effort reported variously as the RTX Pro 6000D, the B40, the B30 and the B30A, and a separate Wall Street Journal account described a B30 at around eighty per cent of full Blackwell performance, which cannot be reconciled with half of a B300.