Moonshot’s Blackwell Chase Shows China’s AI Ambitions Won’t Wait for Permission

Chinese AI startup Moonshot is reportedly scrambling to secure more Nvidia Blackwell GPUs for its next foundation model, and the move tells you everything about where the global compute race actually stands. While Washington draws up new export-control spreadsheets, Beijing’s best-funded labs are already treating next-gen silicon as a scheduling problem, not a geopolitical one.

The chip that hasn’t shipped yet

Blackwell — Nvidia’s post-Hopper architecture — exists mostly in roadmap slides and analyst notes. Yet Moonshot, backed by Alibaba and HongShan Capital, is already negotiating allocation. That’s not optimism. That’s a bet that the first teams to tape out Blackwell clusters will own the next generation of Chinese-language reasoning, coding, and multimodal models. Everyone else will be fine-tuning on borrowed time.

Export controls are a speed bump, not a wall

The Biden administration’s October 2023 rules cut off H100 and A100 shipments. The updated October 2024 rules extended the net to H20 and theoretically to Blackwell. On paper, Moonshot shouldn’t get a single wafer. In practice, three channels stay open: pre-order agreements signed before the cutoff, third-country transshipment through Singapore and the UAE, and Nvidia’s own “compliant” SKUs with nerfed NVLink speeds that still outclass domestic alternatives. Moonshot’s procurement team knows every loophole by heart.

Domestic silicon isn’t ready for prime time

Huawei’s Ascend 910B and Cambricon’s MLU590 are credible on inference. For dense, month-long pretraining runs with 10 trillion tokens, they still hit memory-bandwidth walls and driver maturity issues. Moonshot’s engineers have run the internal bake-offs. They know exactly how many Blackwell nodes replace a 2,000-card Ascend cluster. The answer is fewer than you’d think, and the power bill drops 35 percent. That math drives the procurement urgency.

What this means for the rest of 2025

• Expect more “compliant” SKU announcements from Nvidia — same die, fused-off NVLink, sold at a premium.
• Watch Singapore customs data; Blackwell transshipment volumes will become a leading indicator for Chinese model releases.
• Domestic foundries (SMIC, CXMT) will get pressure to accelerate 5 nm and 3 nm tape-outs, but yield curves don’t negotiate.

The uncomfortable truth for policymakers: compute hunger in Beijing moves faster than rulemaking in Washington. Moonshot’s next model will train on Blackwell, one way or another. The only variable is whether Nvidia books the revenue directly or watches it route through intermediaries. Either way, the model ships on schedule.

Source: original report