Picture a restaurant with a world-class kitchen, twelve burners, a walk-in freezer, and a chef who never sleeps. Now take away the pass — the narrow shelf where finished plates go out to the dining room. The kitchen doesn’t slow down a little. It stops. Everything backs up behind that one strip of stainless steel.
That shelf is high-bandwidth memory, and right now it’s the most expensive real estate in AI hardware. Reuters reported on September 10 that Chinese AI chip prices have jumped 20% to 50% in the space of two months, driven by HBM supply tightening. The chips themselves didn’t get better. The pass got narrower.
What actually moved
The numbers are specific enough to be uncomfortable. Huawei’s Ascend 950PR sold for roughly 60,000 yuan per card at the start of 2026. It now fetches more than 80,000 yuan. Older Ascend parts are caught in the same updraft, which tells you this isn’t a premium being charged for a shiny new part — it’s a supply constraint pricing everything on the shelf.
Cambricon has repriced its next-generation chip, tentatively called the 690, at 20% to 30% above what it was quoting two months earlier. Smaller domestic players MetaX and Iluvatar CoreX have moved their pricing too. When the entire domestic field adjusts in the same direction inside one quarter, that’s not competitive positioning. That’s a shared input getting scarce.
The upstream picture explains why. DRAM prices surged 90% in Q1 2026 alone compared to Q4 2025, as major memory suppliers including Samsung and SK Hynix pivoted capacity toward AI-driven demand. SK Hynix has warned the shortage may persist past 2030. That is not a supply blip you wait out with a purchase order and some patience.
The timing is the awkward part
China told its largest tech firms to move off Nvidia and buy home-grown AI chips. The domestic supply chain got handed a demand surge and a mandate at the same moment the memory that feeds those chips became the scarcest component in the stack. Policy created the customers. Physics is setting the price.
Huawei has publicly positioned the 950DT as its most advanced part in this line, which means the flagship and the older inventory are both competing for the same constrained memory pool. There’s no obvious way to route around that with clever packaging or a software update.
Why this lands on agent builders
I spend most of my time looking at agent tooling — frameworks, orchestration layers, the stuff that turns a model into something that does work. It’s easy to treat hardware as somebody else’s problem three layers down. This one isn’t.
Agents are memory-bandwidth hogs in a way that single-shot chat completions are not. A conversational request loads a context, generates a few hundred tokens, and exits. An agent loop does something different:
- Long, growing context windows as tool results and intermediate reasoning accumulate
- Repeated inference passes on the same task, each one re-reading a KV cache that keeps expanding
- Multiple parallel sub-agents, each holding its own state in memory
- Retrieval steps that pull in documents and inflate the working set further
Every one of those patterns pushes on memory bandwidth and capacity rather than raw compute. Which means the component that just got 20% to 50% more expensive is precisely the component agent workloads consume most aggressively. If you’re running agents on Chinese domestic silicon, your cost-per-task curve just steepened at the exact point where you were hoping to scale.
The practical read
A few things follow from this, none of them dramatic, all of them worth acting on.
Context discipline stops being a nice-to-have. Pruning tool outputs, summarizing intermediate steps, capping loop depth — these were performance optimizations. On constrained memory they become cost controls. An agent that carries 100k tokens of context through twelve steps is paying rent on that memory the whole way.
Smaller models get more attractive for routing and classification work. Not because they’re better, but because the memory footprint of a 7B model doing tool selection is a fraction of a frontier model doing the same job with the same accuracy.
And if you’re making procurement decisions on a two-year horizon, the SK Hynix warning about supply beyond 2030 is the fact to sit with. Memory constraint isn’t a temporary condition to route around. It’s a design parameter.
The interesting shift here is conceptual. We’ve spent three years talking about AI capability in terms of FLOPs and parameter counts — how much compute can you point at a problem. The pricing data coming out of China suggests the binding constraint has moved. Not how fast you can think, but how much you can hold in mind while you do it. Agent architectures that respect that limit are going to be cheaper to run than ones that don’t, on any silicon, in any country.
🕒 Published: