NVIDIA designs the B30A, a trimmed version of the Blackwell Ultra to keep its business in China under US restrictions.

NVIDIA’s strategy with the Chinese market has been a puzzle in constant adjustment for months. The Export restrictions imposed by the United States have forced the company to reformulate its products so as not to lose ground that, due to volume, is impossible to ignore. First it was the H20, which obtained the necessary license to cross borders, and now a new actor appears on the scene: the B30A.

This chip, according to Reuters, is a cropped version of the Blackwell Ultra B300one of the firm’s star accelerators. The key is that The B30A uses a single chip instead of the dual-chip design of the original model. With this, NVIDIA manages to reduce performance by half and, in the process, fit within the limitations imposed by the US government regarding computing power and technology transfer to China.

What changes compared to the original B300

The starting point is the Blackwell B300 Ultra, a true computing monster capable of deploy up to 15 teraflops at FP4, 7.5 at FP6/FP8, 3.75 at FP16/BF16 and close to 1.88 teraflops with TF32 precision. In the case of the new B30A, those figures fall exactly in half, since it only integrates one of the two chips that the original B300 has.

The reduction is significant on paper, but it doesn’t mean the throttle falls short. In fact, what NVIDIA is offering with the B30A It is still more than enough to stand up to any own design that may arise in the Chinese ecosystem.. The great asset remains the same: its software stack and development tools, far ahead of what the local competition offers.

What the B30A does not lose

Although power has been cut, NVIDIA has not given up key functions. The B30A maintains compatibility with HBM memory and NVLinktechnologies that are decisive both in the model training phase and in inference. That is, with respect to architecture and connectivity, this chip continues to play in the same league as the company’s most ambitious products.

The move is clear: offer AI laboratories in China a product that, although limited by regulation, allows them to continue working with competitive hardware without having to make the leap to alternative developments.

A permit with a lot of political weight

The other piece of this puzzle is the export license that the United States has granted to NVIDIA. Without that approval, the B30A would not have a presence in the Chinese market. The green light shows the extent to which Washington seeks a balance: containing Chinese technological advance without completely suffocating North American companies that depend on that market.

NVIDIA, for its part, cannot miss the opportunity. China is one of the great drivers of demand in artificial intelligence and leaving free space would practically give away the market to local rivals. The B30A, therefore, functions as a kind of “bridge” between political limitations and the technological needs of clients.

A future still in the air

The movement makes it clear that NVIDIA does not want to let go of China’s hand, but it also reflects the complexity of operating in an increasingly monitored scenario. It remains to be seen how the Chinese industry responds and whether the B30A performance cut is enough to meet the demands of its most ambitious laboratories.

The truth is that, with this launch, NVIDIA maintains its presence in a key market without breaking the rules of the game. And, incidentally, it makes it clear that, even with the limitations on the table, it still has room to maneuver and continue to be a main player in the race for artificial intelligence.