"The biggest gap between us and the US lies in resources," Liang Wenfeng reportedly told investors at a recent meeting, tying the difference in talent, model quality, and applications directly back to raw computing power. The minutes from that meeting, which neither Liang nor DeepSeek has officially confirmed, paint a picture of a company that believes the hardware ceiling is about to lift.
Liang's core claim is straightforward: chip supply is currently the last meaningful bottleneck holding Chinese AI back, and he expects it to stop being a problem within twelve months. DeepSeek is currently testing around 16,000 Huawei Ascend 950 chips, and Liang told investors those chips match Nvidia's GB200 and GB300 supernodes on both performance and price. That is a striking assertion given how aggressively the US has used export controls to keep advanced Nvidia silicon out of Chinese hands. The meeting took place in the context of DeepSeek raising outside funding for the first time, which gives Liang's optimism at least some commercial weight behind it.
On the raw scale of the gap with the West, Liang was candid. The largest frontier models in the US run at roughly 800 billion active parameters; China's best sit at tens of billions, a tenfold spread. DeepSeek's own estimate puts it 12 to 18 months behind US leaders, though the company insists it achieves comparable output on about one-twentieth of the compute. The stated target is to close that lag to three to six months. Huawei is not the only domestic card on the table: Alibaba and Moore Threads are also pushing alternatives to Nvidia hardware, and Chinese researchers are actively working to reduce dependence on Nvidia's CUDA programming ecosystem, which has long been the invisible lock-in keeping developers tethered to US chips.
DeepSeek is also reportedly designing its own inference chip, which would put it in direct competition with both Huawei and Nvidia simultaneously. If that project moves forward, it signals a broader ambition beyond just running models efficiently on someone else's silicon. The chip design push, combined with the Ascend 950 tests and the CUDA workarounds, suggests China's AI hardware strategy is converging on several fronts at once rather than betting on a single fix.
