Chinese AI chipmakers Huawei and Cambricon are raising accelerator prices as the global high-bandwidth memory shortage ripples into Beijing's push for a domestic Nvidia alternative. Reuters reported on September 10 that Huawei's upcoming Ascend 950DT has been quoted at more than 250,000 yuan — roughly $37,000 — with some quotes up 20% to 50% over prior guidance, while Cambricon has raised prices on its next-generation 690 by 20% to 30%. MetaX and Iluvatar CoreX have followed.
Memory, not fabs, is the choke point
High-bandwidth memory has become the AI supply chain's tightest constraint because every serious accelerator needs it, and the leading edge is dominated by SK Hynix, Samsung and Micron. China faces an added hurdle because U.S. export controls restrict access to the most advanced HBM stacks, pushing domestic buyers into more expensive back-channel sourcing. The result: even chips fabricated locally now carry a memory bill that erases much of the "domestic alternative" price advantage.
What the price hike does to Beijing's plan
Beijing has spent two years financing a domestic accelerator ecosystem meant to reduce Nvidia dependency. The HBM crunch complicates that plan: Huawei is still shipping impressive silicon in phones, but data-centre accelerators are the workload that matters for training frontier models, and their bill of materials is now moving the wrong way. It also shows that the AI-infrastructure bottleneck has shifted beyond GPUs themselves — memory, packaging, networking and electricity now decide who trains what, and when.
Global read-through
The read-across is that even non-Chinese buyers should expect firmer accelerator pricing into 2027. If China is paying more for Ascend and Cambricon parts to sidestep export controls, hyperscalers in the U.S. and Europe are competing for the same limited HBM3E and HBM4 supply — which was the subtext of Nvidia's 2 GW Australian buildout and every other multi-year capacity deal signed this quarter.
Reporting based on coverage from Reuters and the Financial Times.