According to Reuters, the new silicon will start rolling off production lines by June and should be all over the Chinese market by the end of the year.

Thanks to the US export tantrum, the Hopper-based H100 and H200 accelerators never really had a chance to shine in China. Nvidia attempted a workaround with the H800, which lasted until October 2023 before being added to the naughty list. Then came the H20, the AI stopgap that got binned last month, costing Nvidia a $5.5 billion inventory write-off.

Now the B40 is stepping in, and tipster Jukanlosreve reckons it’s just a rebadged RTX Pro 6000D. A report from China's GF Securities suggests the chip ditches expensive high-bandwidth memory (HBM) for good old-fashioned GDDR7, and steers clear of TSMC’s fancy CoWoS packaging. That strongly hints at a monolithic die.

While Nvidia’s keeping technical details close to its chest, there’s some reasonable guesswork to be done. The B40 could be built on the GB2XX consumer-grade silicon, specifically a variant of the GB202 found in the RTX 5090. That means no NVLink support and a narrower interconnect, but it does make the card cheaper and less risky for export approval.

Reuters suggests the B40 is server-class, but not in the same league as the H100. Price estimates range between $6,500 and $8,000, significantly less than what Nvidia was charging for the H20, and roughly on par with existing global RTX Pro 6000 models.

The previous HGX H20 could be packed into eight-way configurations, but without NVLink, the B40’s multi-GPU abilities will be limited. Nvidia is expected to fall back on its ConnectX-8 SuperNICs and PCIe 6.0 switches to link multiple B40s, with anything beyond eight GPUs relying on Spectrum-X networking wizardry.

It’s all speculative at this point, and Nvidia’s not saying much. But with Chinese customers locked out of the best AI silicon, the B40 might be their best bad option.