R2 Failed on Ascend. DeepSeek Wants the Die.

DeepSeek — the Hangzhou lab whose R1 shock made Nvidia's January bloodbath look tame — is now designing its own inference chip, according to three people who spoke to Reuters on July 7. The project is roughly a year old, still early, and aimed at the one workload DeepSeek can actually run on Chinese silicon today: answering queries, not training trillion-parameter models from scratch.

What's new: Reuters reported Monday that DeepSeek has been in talks with outside chip-design firms, foundries, and memory suppliers while quietly hiring silicon engineers — without posting public job listings for those roles. The chip would serve inference only: generating responses from models already trained elsewhere. DeepSeek declined to comment.

Why it matters: This is vertical integration under duress. DeepSeek built its reputation squeezing frontier performance from other people's GPUs. Now it wants to specify the silicon itself — the same strategic pivot OpenAI made with its Jalapeno inference chip in June, but driven by export controls rather than margin envy.

The numbers:

• Nvidia shares fell about 1.6% to roughly $191.70 on July 7 after the Reuters report, per market data relayed by trade press.
• Huawei still holds roughly half of China's estimated $50 billion domestic AI chip market — but Reuters noted that grip is weakening as Alibaba and Baidu build rival processors.
• DeepSeek closed a maiden $7.4 billion funding round in June at a valuation above $50 billion, per The Information — weeks after Reuters reported Washington had approved but not published an Entity List designation for the company.
• Industry analysts estimate inference now accounts for roughly 70% of total AI compute demand — the segment where specialized, lower-power chips can undercut general-purpose GPUs.

The backstory: DeepSeek already knows what happens when it bets on domestic training silicon. The Financial Times reported in August 2025 that Beijing pushed the lab to train its R2 successor on Huawei Ascend accelerators — and that after months of work, unstable chips, slow interconnects, and immature software, DeepSeek could not complete a single successful training run. The company pivoted back to Nvidia hardware for training while relegating Huawei chips to inference duty.

That failure delayed R2 and exposed the division of labor Chinese AI now lives with: train on whatever Nvidia silicon you can smuggle or license; run inference on Ascend, H800 leftovers, or whatever domestic part yields. A home-grown inference chip would let DeepSeek co-design hardware around its sparse mixture-of-experts architectures — potentially driving token costs even lower on the workload that actually scales with users.

The geopolitics: Washington has already approved DeepSeek for the Commerce Department's Entity List, according to a June 16 Reuters exclusive — but held off publishing the designation alongside more than 100 other Chinese firms flagged as security risks, trying to avoid escalating trade tensions with Beijing. Anthropic separately identified DeepSeek among Chinese labs running distillation campaigns against Claude; in a June 10 Senate Banking Committee letter, it accused Alibaba's Qwen lab of 28.8 million fraudulent exchanges — the largest known attack of its kind. OpenAI has warned lawmakers of similar targeting by DeepSeek.

DeepSeek is not on the formal blacklist yet. But the chip project reads like insurance against the day Commerce finally publishes the list — or the day the last H800 channel closes.

What we cannot verify: Reuters cited three anonymous sources; DeepSeek has not confirmed the project, named a foundry partner, disclosed a prototype, or published benchmarks. Trade press relaying Reuters has identified SMIC as a likely fabrication candidate given U.S. restrictions on TSMC access — but no production timeline or node specification has been reported. The R2 training failure comes from FT reporting in August 2025, not from DeepSeek's own disclosures.

Convina's view: DeepSeek is not building a Nvidia killer — it is building an exit ramp. Training frontier models still requires hardware Washington can choke; inference is where Chinese labs can actually win on cost per token. The Reuters report landed the same week Samsung posted record chip profits and Seoul hit a circuit breaker anyway — proof that markets are pricing perfection while the labs underneath are scrambling for silicon sovereignty. If DeepSeek ships even a modest inference part, the competitive question stops being whether its models are cheap and starts being whether anyone outside China needs Nvidia's margin to answer a prompt.