DeepSeek's R2 model delayed as Beijing pressures the lab to ditch Nvidia for Huawei chips
Aug 15, 2025
Key Points
- DeepSeek's R2 model release is delayed because Beijing pressured the lab to abandon Nvidia chips for Huawei's Ascend hardware, which encountered serious training obstacles including interconnect failures and memory problems at scale.
- Huawei's infrastructure, despite theoretical capability parity, lacks Nvidia's software maturity and reliability—gaps that no amount of cheap power can fix, exposing China's struggle to build sovereign AI chip capacity.
- If DeepSeek ships an o3-level model at 100x lower inference cost, it would crater margins for application-layer companies relying on OpenAI's API, making the timing and performance of R2 a market-defining event.
Summary
DeepSeek's R2 model, originally scheduled for May release, remains delayed. The Financial Times reports that Chinese government pressure forced the lab to abandon Nvidia chips in favor of Huawei hardware.
DeepSeek, the high-frequency trading firm turned open-source AI lab under High Flyer, has been forced to switch from Nvidia to Huawei's Ascend chips. Using Nvidia hardware is not technically illegal, but Beijing has made it politically untenable. The shift reflects China's push to build sovereign chip capacity, though Huawei's infrastructure is not yet mature enough for the task.
Huawei's Ascend CloudMatrix 384 was initially seen as competitive with Nvidia on raw capability—less efficient per dollar and more energy-intensive, but functionally equivalent with cheap power. That assumption appears wrong. The technical gap runs deeper than flops-per-watt. Cuda's reliability, driver maturity, and chip architecture deliver qualitative advantages that cheap energy cannot overcome. When DeepSeek tried to train at scale on Huawei hardware, the training run hit serious obstacles including interconnect failures, power management problems, and memory issues. These bottlenecks explain why AI researchers command high salaries: resolving a single training failure can be worth tens or hundreds of millions.
The delay matters because inference cost—price per million tokens—will determine whether DeepSeek's next model disrupts the market or merely competes within it. If DeepSeek ships an o3-level model at 100x lower cost, it would crater gross margins for every application-layer company relying on OpenAI's API, even without displacing OpenAI itself. Companies like Perplexity have already shown the path: fine-tune DeepSeek's base model to improve usability, then leverage its cheap inference to flip unit economics positive.
The Financial Times account faces skepticism. T. York Taxis argues the story relies on half-baked rumors and patronage politics rather than hard confirmation. Beijing's push to champion Huawei as the national AI champion is real—the party asked Huawei and DeepSeek to partner in February to beat America at AI—but DeepSeek may have quietly told officials that Huawei's ecosystem is too immature for training and defaulted to Nvidia's H100s instead. On this reading, the R2 timeline never made public sense, and the delay may simply reflect DeepSeek's unwillingness to ship an inferior model.
The unresolved question is whether DeepSeek breaks ranks with Beijing and orders more H20s (Nvidia's China-approved but four-year-old, heavily throttled variant) to ship R2 on schedule, or swallows the delay and waits for Huawei to mature.