Huawei, Cambricon Hike AI Chip Prices as HBM Squeeze Bites

Huawei has pushed the indicated price of its flagship Ascend 950DT AI accelerator card above 250,000 yuan, roughly $37,255, a jump of 20% to 50% from what it was quoting customers just two months earlier. Cambricon, the other major name in Chinese AI silicon, has repriced its upcoming "690" chip 20% to 30% higher over the same window. Reuters broke both numbers on September 10, and they say something narrower and sharper than the general memory shortage story that's been running all year.

• Huawei's Ascend 950DT, its most advanced chip and due to ship in volume in Q4 2026, now lists above ¥250,000 ($37,255), up 20% to 50% from July 2026 quotes.
• Cambricon's next-generation "690" chip carries a 20% to 30% markup versus the pricing it indicated just two months ago.
• The driver is HBM: since Washington tightened export controls on advanced high-bandwidth memory to China in December 2024, Chinese chipmakers have leaned on grey-market channels that charge several times what buyers elsewhere pay for the same memory.
• The higher chip prices raise the cost of China's domestic AI buildout right as Alibaba, Tencent, Baidu, ByteDance and DeepSeek race to scale compute without access to Nvidia's top-tier accelerators.

Why are Huawei and Cambricon raising prices now?

Both companies are pricing in a cost they can't hedge away. HBM, the memory stacked directly on top of an AI accelerator die, has become the tightest bottleneck in the entire AI hardware chain this year, and GenZTech covered that broader DRAM and HBM supercycle back in June and July. What's different here is that Huawei and Cambricon aren't just paying more for memory the way everyone is. They're paying a China-specific premium on top of the global shortage, because the highest-performance HBM3E and HBM4 parts from SK Hynix, Samsung and Micron are largely off-limits to them under US export rules. A card that costs $37,255 today cost somewhere between $25,000 and $31,000 in July, depending on the contract, and the gap is almost entirely memory cost passed straight through to the buyer.

RelatedSamsung Hiked Foundry Prices After Losing a Third of Share

What is the grey market for HBM, and how does it work?

Since December 2024, when Washington expanded controls on advanced memory products bound for China, Huawei, Cambricon and the rest of China's AI chip industry have been formally cut off from buying cutting-edge HBM directly from SK Hynix, Samsung or Micron. Demand for AI compute inside China didn't slow down to match, so a grey market filled the gap. Memory gets rerouted through intermediaries, shell buyers and third countries before it reaches a Chinese packaging line, and each hop in that chain adds margin. Reuters' sourcing puts the resulting cost at several times what a buyer outside China pays for identical parts. That's not a rounding error in a bill of materials. HBM is one of the most expensive components in any AI accelerator, so a multiple on its price shows up almost directly in the price of the finished chip, which is exactly the 20% to 50% swing Huawei just posted.

How does the Ascend 950DT compare to Cambricon's 690 and Nvidia's H20?

The three chips sit in different lanes of the same race. The Ascend 950DT is Huawei's flagship, not yet shipping in volume but already being quoted to customers ahead of a Q4 2026 launch. Cambricon's 690 is further out and still being priced off indicative terms rather than firm orders. Nvidia's H20, the cut-down accelerator built specifically to comply with US export rules and stay legal for sale into China, remains the closest thing to a benchmark for what Chinese buyers can get from the outside world, though its performance trails Nvidia's global lineup by design. None of the three is a clean substitute for the others, but together they define the actual menu available to a Chinese cloud provider shopping for AI compute in late 2026.

What does this mean for China's AI race?

China's cloud providers and AI labs were already building around a hardware disadvantage. Nvidia's H100, H200 and B200-class chips are barred from sale into China, leaving domestic buyers dependent on Huawei, Cambricon and the export-compliant H20 to fill racks. Now the chips built to close that gap are getting meaningfully more expensive to buy, on top of already trailing Nvidia's top tier on raw performance. For a lab like DeepSeek or a hyperscaler like Alibaba, that means either paying more per unit of compute than planned, delaying cluster buildouts, or accepting a slower pace of scaling than US-based rivals with unrestricted access to Nvidia's newest silicon. None of that stops China's AI buildout. It does make it costlier and slower than it would be without the export controls, which is the effect those controls were designed to have.

RelatedAMD buys Taalas, betting on AI models etched into silicon

What it means for the market

The ripple effects reach past China's borders. SK Hynix, Samsung and Micron aren't direct beneficiaries of grey-market sales, since those channels route around them rather than through official contracts, but the episode confirms just how much pricing power HBM makers hold right now: demand is strong enough that even buyers locked out of the legal channel will pay multiples to get supply anyway. That reinforces the case for continued HBM price strength industry-wide, which matters for every AI company budgeting compute costs, not just Chinese ones. For Nvidia, a costlier Ascend 950DT and 690 don't open the Chinese market back up, since export rules rather than pricing are what keep Nvidia's best chips out, but they do make the H20, Nvidia's remaining China-legal product, look more competitive on a dollar basis than it did in July. And for Chinese AI labs, the math is blunt: compute that was already scarce relative to US rivals just got more expensive to acquire, squeezing margins on every model they train or serve.

The bigger picture

Put together, Huawei's and Cambricon's price hikes are less a story about memory shortages in general and more a story about what sanctions do to a market once the legal supply chain closes. Grey markets don't disappear demand, they just reprice it, and the buyers on the wrong side of an export control end up funding that repricing directly. That dynamic isn't going away as long as December 2024's HBM restrictions stay in place, and it's worth watching whether Huawei and Cambricon can hold Q4 2026 launch timelines at these new prices, or whether cost pressure ends up being the thing that slips the schedule instead.

Original analysis by GenZTech. Source: Reuters via Investing.com