AI-Chip Wars 2.0: Nvidia’s Tailored Blackwell for China

date
02/07/2025
avatar
GMT Eight
Nvidia is rushing a cut-down Blackwell-architecture GPU - informally dubbed the RTX Pro 6000D or “B40” - into Chinese mass production after Washington’s April tightening of export rules banned the H20 accelerator.

Washington’s latest curbs, unveiled in April, introduced for the first time a quantitative limit on bandwidth as well as raw computing power. The ban sidelined Nvidia’s H20 overnight, forcing a US $5.5 billion write-off and an estimated US $15 billion hit to forward sales. Rather than abandon the world’s second-largest data-centre market, Nvidia chose to design a new Blackwell derivative whose throughput and memory subsystem sit just inside the legal threshold. Sources tell Reuters that engineering samples were shown to cloud customers in Beijing in May, with volume production pencilled in for late June and a second, even leaner Blackwell SKU tentatively scheduled for September.

Technically the 6000D steps back from the H20 on three fronts: it drops advanced HBM3 stacks in favour of GDDR7; it foregoes CoWoS 2.5D packaging, easing pressure on TSMC’s chronically scarce assembly lines; and it omits NVLink, relying instead on Spectrum-X Ethernet fabric for scale-out connectivity. Analysts at Jefferies calculate that these changes will slash training performance to well under half that of the H20, but keep inference numbers respectable for language-model serving - an acceptable compromise for Chinese hyperscalers that cannot legally import H100-class silicon.

Competitive dynamics are shifting in parallel. Huawei, already the main beneficiary of earlier export rounds, began shipping the dual-die Ascend 910C in May; local AI firms describe its inference speed as approaching that of an H100 at lower power but admit that the software ecosystem still trails CUDA by at least a generation. Meanwhile, domestic GPU start-ups such as Moore Threads and MetaX are preparing multibillion-yuan IPOs on Shanghai’s STAR Market, explicitly pitching themselves as national champions able to fill any future vacuum if curbs tighten again.

For investors the episode crystallises three take-aways. First, policy volatility is now a material line-item for earnings models: every incremental restriction rewrites the addressable market overnight. Second, supply-chain beneficiaries will be uneven - GDDR7 suppliers and conventional substrate makers gain share, while CoWoS capacity and HBM vendors lose China-related upside. Third, the long-run moat may migrate from hardware margins to software and platform subscriptions: Nvidia is already piloting CUDA-as-a-service pricing to preserve economics even if silicon ASPs fall.

Barring fresh sanctions, Blackwell 6000D shipments should anchor Nvidia’s China revenue near 2024 levels through year-end, albeit at thinner gross margins. Yet the strategic message is clear: the AI-chip race is no longer about absolute performance but about staying one regulatory step ahead while rivals at home race to close the gap.