20-50% Price Hikes Across the Board: HBM Shortage Squeezes China's AI Chipmakers

China’s campaign to replace Nvidia with homegrown AI silicon has hit an unexpected tax: its own chip prices are rising faster than almost anyone else’s. Reuters reported on September 10 that Huawei, Cambricon, MetaX and Iluvatar CoreX have all raised prices on current- and next-generation AI accelerators by 20-50% over the past two months, blaming a worsening shortage of high-bandwidth memory (HBM) — the stacked, ultra-fast DRAM that every serious AI processor depends on.

The numbers

The starkest figure comes from Huawei. Its Ascend 950DT accelerator card is now being quoted at more than 250,000 yuan (roughly $37,255) — 20% to 50% above the prices the company was giving customers just two months ago. The rack-scale Ascend 950PR has climbed around 30% to over 80,000 yuan. Cambricon has repriced its next-generation 690 chip 20-30% higher. Even older-generation cards from multiple vendors have seen quotes rise, effectively raising the price of every incremental unit of AI compute being added across China.

The hikes are not a demand signal in the usual sense — they are cost pass-through. Since the United States tightened export controls on advanced HBM shipments to China in December 2024, Chinese accelerator makers have had to source high-end memory through grey-market channels, where prices run to several multiples of what SK Hynix, Samsung and Micron charge their contracted customers elsewhere in the world. Memory is one of the largest line items in an accelerator’s bill of materials, so when HBM costs triple, finished-board quotes have nowhere to go but up.

Why HBM is the choke point

HBM is the ingredient that makes large AI models practical. By stacking DRAM dies vertically and placing them millimeters from the compute die, HBM delivers the memory bandwidth that transformer inference and training demand. The advanced-HBM market is dominated by South Korea’s SK Hynix and Samsung along with America’s Micron — none of which can freely sell their top-tier products into China under current US rules.

That leaves Chinese vendors in an awkward position: their domestic logic chips are improving fast — Huawei has been sharply ramping Ascend production, targeting roughly 600,000 units of its most advanced parts in 2026 — but each accelerator still needs to be fed with HBM that China cannot yet manufacture at competitive quality or volume. Domestic memory efforts exist, but they trail the leading edge by a meaningful margin. The result is a structural arbitrage: Chinese AI chipmakers can design cards competitive with last-generation Nvidia hardware, but only by paying grey-market prices for the memory that makes them work.

Who wins, who pays

The squeeze is already reshaping allocation decisions. One source cited in the reporting said Iluvatar CoreX doubled its shipments to ByteDance to 100,000 GPUs this year — partly by diverting units that had originally been earmarked for internal use. When supply is scarce and expensive, big buyers with deep pockets and strategic relationships absorb what’s available, while smaller AI labs and enterprises face the sharpest effective price increases.

For the chipmakers themselves, the price hikes land at a delicate moment. Several have just turned the corner on profitability: MetaX posted a net profit of 612 million yuan in the first half of 2026, reversing a 186-million-yuan loss a year earlier, and analysts tracking the sector expect Huawei and Cambricon to be the biggest winners as homegrown accelerators climb toward supplying an estimated 90% of China’s domestic AI chip market. Rising prices protect those margins, but they also test just how price-sensitive China’s AI buildout really is.

The pass-through economics

There is a simple reason these costs are showing up as higher prices rather than thinner margins. When HBM is the scarce ingredient, the companies that manage to secure it gain pricing power — buyers cannot substitute cheaper memory without redesigning the entire card. Higher HBM costs therefore flow through to finished-board prices almost mechanically.

For hyperscale buyers like ByteDance, higher quotes mean a higher cost per unit of training capacity. Over time, that pushes spending toward whatever hardware is actually available for a given job, nudges the market toward inference-heavy deployments, and could slow or reshape China’s data-center expansion — even if the headline count of GPUs shipped keeps rising.

The bigger picture

The price hikes expose the true cost of the export-control era for China’s AI ambitions. Import substitution works when every layer of the stack can be localized; AI accelerators cannot, because HBM remains a Korea-and-America oligopoly. Every restriction that raises China’s memory cost raises the cost of the entire domestic AI stack built on top of it.

It also cuts the other way. Grey-market HBM demand at multiples of world prices is itself a signal — it shows how much Chinese buyers will pay rather than slow down, and how far the country’s AI buildout is from being throttled by cost alone. The 250,000-yuan Ascend 950DT quote is simultaneously a symptom of scarcity and proof that the demand behind China’s compute buildout remains strong enough to absorb a 50% price shock without collapsing.

For the global industry, the lesson is that memory — not logic — is currently the binding constraint on AI compute everywhere, and nowhere more visibly than in China. Until domestic HBM closes the gap or export policy shifts, China’s Nvidia-replacement project will keep paying a memory premium, one price hike at a time.