China's AI Chip Makers Raise Prices Up to 50% as HBM Shortage Tightens Supply Chain Grip
Huawei and Cambricon have quietly lifted asking prices on key AI accelerators by as much as 50% in two months, exposing a structural bottleneck in high-bandwidth memory that threatens to inflate computing costs across China's booming AI industry well into 2027.
The price moves, first reported by Reuters on September 10 citing three people with direct knowledge of the matter, mark a significant inflection point for China's domestically produced AI compute stack. Until recently, Chinese chip vendors had competed aggressively on price to displace Nvidia hardware restricted under U.S. export controls. The latest increases suggest that demand has now outrun the domestic supply chain's ability to absorb cost pressures at the component level — a dynamic that shifts pricing power decisively toward suppliers.
Market participants have taken note. Cambricon's shares have drawn renewed scrutiny from institutional desks tracking the AI infrastructure theme, even as the company moved to distance itself from the reported figures, telling 21st Century Business Herald on September 11 that it had "not issued any price-increase announcement" and urged investors to rely solely on official disclosure channels.
HBM Scarcity Forces Vendors to Pass Through Rising Input Costs
The proximate trigger is high-bandwidth memory. HBM sits at the heart of every modern AI accelerator, and its supply remains structurally constrained. Micron Technology Chief Executive Sanjay Mehrotra told analysts on June 24 that expanding HBM output is inherently slow — gated by greenfield fab construction timelines, skilled labor availability, and energy infrastructure — and that the market will remain tight beyond 2027. That forecast has now materialized as a direct cost input for Chinese chip vendors who must source or develop HBM independently of Western suppliers.
The pass-through is visible in Huawei's product line. The Ascend 950PR accelerator card has risen from approximately RMB 60,000 (US$8,333) at the start of 2026 to more than RMB 80,000 (US$11,111), a 33% increase. The Ascend 910C board has climbed from roughly RMB 90,000 (US$12,500) to above RMB 110,000 (US$15,278) over the same period. Most striking is the Ascend 950DT, Huawei's forthcoming flagship training and inference card announced by rotating chairman Xu Zhijun at Huawei Connect in September 2025. Its indicative asking price has already exceeded RMB 250,000 (US$34,722) — a 20% to 50% premium over quotes issued just two months ago, depending on contract structure. The 950DT, slated for a Q4 2026 launch, features Huawei's proprietary HiZQ 2.0 memory technology with 144 GB capacity and 4 TB/s bandwidth.
Cambricon's next-generation "690" chip — not yet formally released but already being quoted to prospective buyers — has seen indicative pricing rise 20% to 30% over the same two-month window, according to Reuters. Guohai Securities, in an August 11 research note, described the Siyuan 690 as having entered mass production in early 2026, featuring dual-die packaging, FP16 compute exceeding 700 TFLOPS, and 196 GB of HBM3 memory — a substantial step up from the prior-generation Siyuan 590.
Demand Acceleration Compounds the Pressure, Doubling ByteDance Orders
Supply constraints alone do not explain the full magnitude of the price shift. Demand is simultaneously surging. Reuters reported in June that Iluvatar CoreX was on track to deliver at least 50,000 AI chips to ByteDance this year, primarily for inference workloads supporting the expansion of its Doubao AI assistant. By September 10, that figure had doubled to 100,000 units, with Iluvatar redirecting chips originally earmarked for internal use to fulfill the ByteDance order — a telling sign of how rapidly hyperscaler demand is overwhelming available supply.
Reuters identified Huawei as ByteDance's largest AI chip supplier, followed by Cambricon and Iluvatar. The concentration of China's largest AI application developer around three domestic vendors creates a captive demand base that structurally supports further price increases. Moffat-style competitive pricing discipline — which characterized the market when Chinese vendors were primarily fighting for design wins against each other — is giving way to a seller's market dynamic.
Iluvatar and Metax have also made comparable price adjustments, per Reuters, though neither company has confirmed the reports publicly.
Structural Implications Reshape China's AI Cost Curve
The repricing cycle carries consequences that extend beyond individual vendor margins. For Chinese cloud providers and AI model developers, rising accelerator costs compress the economics of scaling training runs and inference clusters. The cost-per-FLOP advantage that domestic chips once offered relative to smuggled or gray-market Nvidia hardware is narrowing, complicating procurement decisions for budget-constrained AI startups.
For investors, the price increases represent a double-edged signal. Higher average selling prices improve near-term revenue and gross margin optics for Cambricon — a listed company on the Shanghai STAR Market — and for Huawei's unlisted intelligent computing division. But they also raise the risk of demand destruction if enterprise customers defer purchases or seek alternative configurations. Cambricon's explicit caution that it "cannot make judgments on future product price trends" reflects that uncertainty.
The deeper structural question is whether China's domestic HBM production capacity — still nascent relative to Samsung, SK Hynix, and Micron — can scale fast enough to alleviate the bottleneck before it becomes a ceiling on the country's AI ambitions. With Micron projecting supply tightness past 2027, the answer, for now, appears to be no.
Related Coverage:
Huawei’s Kirin 9050 Pro Tests a New Scaling Path Under Chip Sanctions