HBM Shortages Are Repricing China’s AI Accelerators

High-bandwidth memory (HBM) has become a practical constraint on China’s AI accelerator rollout. As supply tightens and memory sourced through unofficial channels becomes more costly, vendors are raising indicative prices, prioritizing major customers, and leaving buyers with less certainty about when they can assemble a full, consistent cluster. 1

The clearest price moves

Huawei has lifted the indicated price of its Ascend 950DT accelerator card to more than RMB 250,000 (about $37,255). Reuters reported that this was 20%–50% higher than customer quotes from two months earlier, with the increase depending on contract terms. Bloomberg separately reported an increase of roughly 60% over three months—a different comparison period that can coexist with Reuters’ range. 1

Cambricon has reportedly repriced its planned next-generation processor, provisionally called the 690, at 20%–30% above indications made two months earlier. No exact headline price was reported. 5

Smaller competitors MetaX and Iluvatar CoreX have also raised prices by similar amounts, according to Reuters reporting. The available evidence does not establish a model-by-model list price or a precise percentage for either company. 16

Earlier Huawei products are reportedly affected as well. The Ascend 950PR rose from around RMB 60,000 at the start of 2026 to more than RMB 80,000, while the older 910C climbed from roughly RMB 90,000 to more than RMB 110,000, according to TrendForce’s report on the Reuters findings. These figures should be treated as market-reported indications, not universally applicable transaction prices. 9

Why an indicative card price is not the final price

The 950DT’s RMB 250,000 figure is a useful signal of cost pressure, but it is not a standardized global list price. Huawei and Nvidia both sell AI processors in batches of eight or more, and reported pricing varies with order size and the amount of accompanying equipment a customer buys, including networking.

That matters for procurement. A customer comparing quotes should separate:

• accelerator count and memory configuration;
• server, network, storage, and integration equipment;
• committed shipment batches and delivery dates;
• price-adjustment clauses tied to component availability;
• hardware acceptance criteria, spares, and support commitments.

A seemingly comparable per-card quote can therefore imply a materially different delivered cluster cost and delivery risk.

Huawei’s roadmap: launch timing remains, supply certainty does not

Huawei has publicly said its Ascend 950DT will be available in the fourth quarter of 2026. The reported HBM-driven repricing has not changed that stated window. However, a launch target is not the same as a commitment that every prospective buyer will receive volume supply in that quarter. 1

Huawei’s roadmap also makes memory central to the product distinction. Reuters reported that the 950PR will integrate Huawei’s proprietary HiBL 1.0 HBM, with 128GB of memory and 1.6TB/s of bandwidth in Huawei’s stated specifications. The higher-performance 950DT is planned to use HiZQ 2.0, with 144GB and 4TB/s of bandwidth in those specifications. 15

Huawei has not disclosed where or how those memory technologies are manufactured, according to reporting on the roadmap. That leaves capacity, qualification, and ramp timing as important unknowns even where product specifications have been announced. 9

The 950PR has already moved beyond a purely roadmap-stage product: Reuters reported in March that customer testing had gone well and that ByteDance and Alibaba planned orders after sampling. The same report described an approximately RMB 50,000 version using DDR memory and a premium HBM version at around RMB 70,000 at that time. 16

A B200 price comparison is not a performance equivalence claim

Bloomberg reported that the 950DT’s suggested price was set in line with Nvidia’s B200. This is best understood as a pricing-positioning reference. It does not establish that the two products offer equivalent compute performance, software maturity, availability, power efficiency, networking, or total cost to deploy at cluster scale.

For buyers, the relevant comparison is an executed system: usable model performance, software compatibility, interconnect behavior, power and cooling needs, and the vendor’s ability to deliver enough matched cards in the required time frame.

Allocation is becoming as important as price

The HBM shortage is reshaping who gets hardware. One report said Iluvatar CoreX doubled its planned shipments to ByteDance to 100,000 GPUs in 2026, including by diverting units originally intended for its own use.

That kind of reallocation makes a large customer more likely to receive hardware, but it also reduces uncommitted supply available to other buyers. The same report described Huawei as ByteDance’s largest domestic AI-chip supplier, followed by Cambricon and Iluvatar CoreX. This is a supplier ranking for ByteDance, not a national market-share ranking.

More broadly, sourcing constraints are amplified by the high cost of HBM. Reporting has said that Chinese firms have relied on stockpiled memory or supplies obtained through black-market channels after U.S. restrictions on advanced HBM exports, with those channels carrying substantial premiums.

What 2026 cluster buyers should do

The practical takeaway is to plan for a supply-constrained system market rather than a commodity-card market.

Use binding, component-specific commercial terms. A purchase agreement should identify the accelerator model, memory configuration, quantities, shipment tranches, acceptance tests, support obligations, and remedies if memory-equipped cards arrive late.

Phase deployments. For products such as the 950DT and Cambricon 690, reserve datacenter capacity and budget, but tie final expansion decisions to qualified hardware and demonstrated multi-node software performance.

Design for allocation risk. A cluster only works as planned if enough compatible accelerators, memory-equipped boards, servers, and networking arrive in the same deployment window. A low nominal card price is of limited value if supply is fragmented across months or configurations.

Keep alternatives where feasible. Multi-vendor architecture cannot eliminate software and operations complexity, but it can reduce dependence on a single supplier’s HBM allocation and delivery schedule.

For Chinese AI accelerator makers, the immediate effect of the HBM shortage is clear: pricing is rising and allocation is becoming more selective. The harder question—whether announced products will be available in sufficient volume for a specific 2026 cluster—can only be answered through enforceable procurement commitments and tested, delivered systems. 1