China's AI chip prices jump 50% as the memory shortage bites
A memory shortage is now setting the price of Chinese AI compute, and California just created the job of auditing AI. Both stories are about the same thing: inputs and oversight that nobody can source cheaply.
Huawei has raised the indicated price of its Ascend 950DT accelerator to more than 250,000 yuan — about $37,000 — a jump of 20% to 50% from quotes it gave customers just two months ago. Reuters reported the increase on Thursday, citing three people familiar with the pricing; Cambricon has repriced its next-generation part, tentatively called the 690, 20% to 30% higher than its indications two months ago, and smaller rivals MetaX and Iluvatar CoreX moved by similar amounts. Older hardware has not been spared: the Ascend 950PR has gone from roughly 60,000 yuan at the start of the year to more than 80,000, and the 910C board from about 90,000 to more than 110,000. None of the four companies responded to requests for comment; Cambricon separately told Chinese media it has issued no such announcement, which is not the same as a denial.
The mechanism is memory, not margins. High-bandwidth memory is stacked DRAM sitting next to an accelerator, and it is a large share of an accelerator's bill of materials — so when HBM prices rise, finished cards follow almost immediately. Washington restricted advanced HBM shipments to China in December 2024, and Chinese buyers have leaned on grey-market channels since, where the same parts cost several times what SK Hynix, Samsung and Micron charge elsewhere. Beijing is simultaneously pressing domestic firms to replace Nvidia, which makes the timing worse: the replacement is getting more expensive at exactly the moment the policy demands more of it.
California now has a state registry for AI auditors. Governor Gavin Newsom signed two bills on Wednesday — SB 813 from Senator Jerry McNerney, which sets up "independent verification organizations" empowered to assess AI systems for compliance with state law, and AB 1405 from Assemblymember Rebecca Bauer-Kahan, which creates a registry for the auditors themselves and standards for their independence and integrity. Together they are the first US framework that requires someone other than the developer to check the work. Bauer-Kahan's framing was blunt: the state "cannot expect industry to simply grade its own homework."
The timing is not subtle — this is the same week OpenAI went to Congress asking for a national rulebook, and McNerney cited this week's news that "the most powerful AI systems teamed with AI agents pose real threats to humanity." The practical read: California is building the evaluation market before Washington exists, and SB 813 codifies a recommendation from Newsom's own blue-ribbon AI panel. The interesting question is whether a state registry becomes the de facto national standard the way SB 53's transparency rules already have, or whether it collides with a federal framework that preempts it.
Samsung showed a prototype that stacks HBM directly on top of the accelerator. Unveiled at the FMS 2026 conference in California, "zHBM" puts the memory on the z-axis above the xPU instead of beside it, which shortens the distance data travels. Samsung claims up to eight times the data-processing performance of HBM5 with three times the performance per watt and less than half the thermal resistance, and it lets customers drop their own IP into an interlayer between memory and processor. Treat the multiples as a vendor's slide — but the direction of travel is the point: when memory distance becomes the bottleneck, the industry's answer is to stop shipping memory as a separate part.
What to watch: whether Huawei's Q4 950DT launch ships at the new price, and which federal AI framework — if any — tries to preempt California's auditor registry.
If a domestic AI chip costs 50% more than it did in July, does China's buildout slow down, or does Beijing just pay the memory premium? Tell us in the comments.
