China's AI chipmakers just hiked prices 20–50% — and the HBM rent flows uphill to SK Hynix, Samsung, and Micron

Reuters reported on Sept 10, 2026 that [Huawei Technologies] lifted the indicated price of its Ascend 950DT accelerator card above 250,000 yuan (~$37,255) — the first confirmed public price quote for the part and a jump of 20% to 50% from what customers were being quoted in July. Cambricon Technologies repriced its still-unlaunched Siyuan 690 20% to 30% above July indications, and smaller rivals MetaX and Iluvatar CoreX made similar moves with no specific figures disclosed. Older parts are not spared: the Ascend 950PR has risen to above 80,000 yuan from roughly 60,000 yuan at the start of the year, and the Ascend 910C to above 110,000 yuan from about 90,000 yuan.

This is not a normal channel-stocking cycle. Reuters directly attributed the increase to soaring high-bandwidth memory costs and the use of grey-market HBM channels — supplies that, per the report, typically cost 'several times' what buyers outside China pay. The export-control clock matters: Washington tightened restrictions on advanced HBM exports to China in December 2024, and the December-2024 framework has been the binding constraint that pushed Chinese chipmakers onto second-tier suppliers.

The HBM bottleneck is not new, but in 2026 it has hardened into a structural shortage rather than a cyclical squeeze. SK Hynix CEO Kwak Noh-jung told Reuters on July 10, 2026 — the day the company listed on Nasdaq — that 2027 'will be the worst year in the industry's history from the supply perspective,' with demand projected to outstrip supply beyond 2030. Bank of America sized the 2026 HBM market at $54.6 billion, up 58% year over year, while TrendForce tracks HBM4 contract pricing of $31–$36 per GB for 2H 2026 — roughly $17–18 per GB for the HBM3E generation it is replacing.

SK Hynix ran a record Q2 2026 — revenue of 79.3 trillion won (~$54.6 billion), up 51% quarter-on-quarter and 257% year-over-year, with operating profit of 60.5 trillion won (557% YoY). S&P Global put HBM revenue alone at 9.4 trillion won for the quarter, broadly in line with consensus. The growth is unmistakable, and so is the constraint: SK Hynix has guided to doubling wafer purchases and is prioritizing DDR5 expansion in some segments to manage capacity — a sign that even the leader is choosing which demand it can serve.

For Chinese chipmakers, the structural reality is harsher. CXMT — China's only DRAM/HBM producer — only began small-scale HBM3E production in late August 2026, and only for two known customers in test: T-Head (Alibaba's chip unit) and Cambricon Technologies. Mass production is targeted for end-2026, and meaningful HBM3E volume is more realistically a 2027 story. Until then, the supply stack is a Chinese version of 'several times' Western pricing plus grey-market availability — exactly the cost Reuters says Chinese chipmakers are now passing through.

Cambricon Technologies is the listed Chinese name with the cleanest exposure to this transmission. Q1 2026 revenue hit 2.88 billion yuan, up 159% year-over-year from 1.11 billion yuan in Q1 2025; net income reached 1.01 billion yuan, up 185% YoY. The company is targeting roughly 500,000 AI accelerators in 2026, including as many as 300,000 Siyuan 590/690 units — though current yields on the 590/690 sit near 20% per Bloomberg sources, capping the upside. A 20–30% price hike on the 690 lifts unit ASPs without changing the unit ceiling, so the first-half 2026 margin expansion (gross margin 54.3% in Q1 2026 vs. 56.0% in Q1 2025) should re-accelerate in 2H 2026 as the new pricing flows through.

SMIC is the foundry that actually fabricates these chips — the take-rate per accelerator does not change with HBM cost, but wafer volume does. SMIC's Q2 2026 revenue rose to $3.01 billion (+18% YoY), with utilization at 93.7% and pricing up more than 10% in key segments per Reuters' August 13 report. SMIC also raised prices for advanced chipmaking by up to 15% in August 2026 per Reuters — meaning SMIC has its own pricing-power story that compounds the accelerator vendors' pass-through.

Every dollar Chinese chipmakers add to accelerator pricing above their grey-market HBM cost accrues to the HBM producer at the top of the stack. The HBM market remains a three-player oligopoly — SK Hynix, Samsung Electronics, and Micron Technology — and all three just printed record quarters. SK Hynix took 62% of HBM shipments in Q2 2025 and is expected by UBS to capture roughly 70% of NVIDIA's HBM4 share for the Rubin platform in 2026. Samsung finally passed NVIDIA's 12-layer HBM3E qualification in September 2025 — its third try — and is now selling out its 2026 HBM allocation.

Micron Technology is the cleanest US-listed expression of the thesis. Q3 FY2026 revenue hit a record $41.5 billion, up 73% quarter-on-quarter, with operating profit of $33.3 billion — a 55.9% net margin. The Street is modeling an analyst target of $1,513 against a recent print near $1,028, with 44 of 48 analyst ratings at Strong Buy or Buy. The narrative is straightforward: every Chinese accelerator shipped at a 20–50% markup allocates a disproportionate share of that markup to the HBM producer at the top of the bill of materials.

The customers absorbing these price hikes are the same names that scrambled for Huawei chips after DeepSeek V4 launched in April 2026. Per Reuters, ByteDance, Tencent, and Alibaba all reached out about new orders once Huawei's Ascend 950 supernodes were scheduled to ship at scale — Huawei plans to ship roughly 750,000 Ascend 950PR units in 2026 and expects AI chip revenue to rise about 60% to ~$12 billion for the year. ByteDance alone is reportedly planning around $5.7 billion in Huawei AI chip purchases.

The China-side math compounds the Western HBM case in a way the headlines miss. TrendForce's 2025 China AI accelerator shipment data already showed Huawei at 20.3% share and Cambricon Technologies at 2.9%, with Nvidia at 55% — a base that is shifting toward 90% domestic share in 2026. The result: more HBM consumed per China-shipped accelerator (because Chinese designs trail NVIDIA's HBM efficiency), more grey-market premium paid (because export controls restrict legal channels), and more rent flowing back to SK Hynix, Samsung Electronics, and Micron Technology. NVIDIA keeps the dollars in the data-center GPU, but the memory dollars migrate.

The Reuters report is a real-time read on 2H 2026 contract pricing for Chinese AI accelerators. Three catalysts matter in the next two quarters: (1) SK Hynix's next HBM contract repricing window — currently flagged by Counterpoint Research to push HBM4 contract pricing toward $35–$36/GB in 2H 2026; (2) CXMT's HBM3E yield data from the T-Head and Cambricon Technologies qualification lines, with first revenue shipments possible by year-end; (3) SMIC's pricing update after the 10–15% August 2026 wafer hikes land in customer bills.

Over the 1–3 year horizon, three structural scenarios compete. The base case: Chinese grey-market HBM costs remain elevated through 2027 as CXMT scales HBM3E, keeping the accelerator pricing-power trade intact and Western HBM margins at cycle peaks. The upside for Chinese chipmakers: a meaningful CXMT HBM3E ramp by mid-2027, compressing grey-market premia and forcing Cambricon/Huawei to compete on performance rather than availability — bad for their margins but good for unit volume. The downside for everyone: a 2027 inventory correction at the hyperscaler level that hits SK Hynix, Samsung Electronics, and Micron Technology together — CEO Kwak's 'demand outstripping supply beyond 2030' guidance is the bull case, but it depends on hyperscaler capex holding.