HBM Shortages Rewrite the Economics of AI Chips
High-bandwidth memory has become a strategic input for AI accelerators. Price increases at Chinese chipmakers show how memory scarcity, export controls and allocation decisions can reshape the economics of domestic AI hardware.
High-bandwidth memory has become a strategic input for AI accelerators. Price increases at Chinese chipmakers show how memory scarcity, export controls and allocation decisions can reshape the economics of domestic AI hardware.
The memory layer sets the pace
AI accelerators are often discussed as if the processor is the whole product. It is not. High-bandwidth memory, or HBM, sits beside the compute engine and supplies data at very high speed. Its vertically stacked design allows an accelerator to work through large models without waiting as often on slower memory paths. When HBM is scarce, the constraint reaches directly into the price and availability of finished AI cards.
Reuters reported that Chinese AI chipmakers including Huawei and Cambricon raised prices as HBM costs squeezed efforts to build alternatives to Nvidia. The reporting matters because it exposes a hidden dependency. A domestic processor can be designed locally and still depend on a globally constrained memory supply chain.
Price changes are a market signal
Reuters reported an indicated price above 250,000 yuan for Huawei’s Ascend 950DT card, a 20% to 50% increase from quotes given customers about two months earlier, depending on contract terms. Cambricon’s planned 690 chip was reportedly repriced 20% to 30% higher. These are not simply price-list updates. They show suppliers attempting to pass through a component shock while demand remains strong.
Older products also moved higher. Reuters reported that Huawei’s Ascend 950PR rose from roughly 60,000 yuan per card at the start of 2026 to more than 80,000 yuan, while the Ascend 910C board rose above 110,000 yuan from about 90,000 yuan. The pattern indicates that scarcity affects installed-base expansion, not only premium launches.
Export controls change the sourcing premium
Export restrictions on advanced HBM products to China have pushed buyers toward grey-market channels, according to the Reuters report. That route can add cost, uncertainty and compliance risk. The memory itself may be identical, but the transaction is not. Buyers pay for scarcity, intermediation and the possibility that a supply line disappears.
This is why national self-sufficiency is difficult to define. It is not enough to produce a processor inside one country. A resilient system needs memory, packaging, substrates, testing equipment, software and reliable logistics. Any missing layer can become the price-setting bottleneck.
The accelerator bill of materials is changing
When HBM takes a larger share of total cost, chip designers face a new tradeoff. More memory capacity can improve performance for large models, but it can also raise the bill and expose the product to allocation risk. Designers may optimize memory use, reduce bandwidth requirements or create products that can accept different memory configurations.
Buyers will make similar calculations. The cheapest accelerator on a quote sheet may not be the cheapest deployed system if it needs more cards, special networking or scarce memory modules. Total cost includes power, software migration, utilization and the probability of receiving the next shipment.
Allocation beats open competition in a shortage
Reuters reported that Iluvatar CoreX doubled planned shipments to ByteDance to 100,000 units this year and diverted GPUs intended for its own use. That is a revealing response to constrained supply. In a normal market, a supplier might spread inventory across customers. In a strategic shortage, large buyers with urgent workloads can receive priority.
Allocation decisions create winners and losers before a product is fully mature. A company with a major platform customer can justify expanding production, while smaller buyers may face longer waits or higher prices. For investors, customer concentration is therefore both a growth asset and a bargaining risk.
Domestic substitution will cost more at first
Beijing wants Chinese companies to replace Nvidia processors, but expensive HBM raises the cost of the domestic alternatives intended to fill that gap. That does not make substitution impossible. It means policy support, anchor customers and software compatibility may be needed to bridge the early economics.
Domestic products can gain value even when they are not the lowest-cost option. They may offer availability, regulatory certainty or data-residency benefits. The commercial question is whether those benefits compensate for performance gaps and higher system costs. The answer will differ by workload and customer.
Software is part of the pricing equation
Hardware prices cannot be evaluated separately from the software stack. A cheaper accelerator that requires extensive porting can erase its initial advantage. Conversely, a chip with strong compiler tools and a familiar programming model can win business despite a higher card price.
The HBM shortage strengthens that logic. If memory limits make every card more valuable, operators will push for higher utilization. Better scheduling, quantization and model optimization can reduce the number of accelerators required. Software that saves memory becomes a form of supply-chain management.
The next bottleneck may be packaging
HBM is connected to advanced packaging, which brings its own equipment and yield constraints. As processors and memory are assembled into larger packages, manufacturing complexity increases. A wafer that looks plentiful at the chip stage does not guarantee enough tested, packaged accelerators.
This widens the diligence checklist for semiconductor investors. Capacity announcements should be examined across memory supply, packaging throughput, test capacity and board integration. The weakest link determines how many usable products reach customers.
What the repricing tells buyers
The reported price increases are a warning against treating AI chips as interchangeable commodities. A buyer should secure supply terms, clarify memory configuration, understand delivery priorities and model the cost of switching software stacks. Procurement contracts may become as strategic as engineering road maps.
The new economics of AI chips are therefore about access as much as efficiency. HBM scarcity turns a component into a strategic lever, export controls turn logistics into policy, and allocation turns customer relationships into capacity. The companies that manage all three will have more room to compete. Buyers should also track contract duration, replacement cycles and the resale value of older cards. A system bought at a high spot price may remain useful if software support lasts, while a cheaper product can become a stranded asset if tools or memory supply disappear.