Exclusive-China’s DeepSeek developing its own AI chip, sources say

DeepSeek is developing an in-house AI inference chip, aiming to reduce reliance on Nvidia and Huawei amid U.S. export controls. The chip is targeted for inference (response generation) rather than training, and the effort is still early, involving discussions with chip-design, foundry, and memory partners. If successful, it could strengthen China’s domestic AI compute capabilities and add pressure to Huawei in the ~$50B China AI chip market, though design and manufacturing constraints (notably foundry and high-bandwidth memory access) remain major hurdles.

Analysis

The market should treat this as a China-stack localization story, not a fresh earnings threat to US AI leaders. Nvidia’s China exposure is already structurally impaired by export controls, so the incremental risk is mostly mix and sentiment, not a new unit shock. The bigger second-order effect is inside China: if DeepSeek can lower inference cost with domestic silicon, it raises the odds that Chinese AI deployment shifts from imported accelerators to local ASICs, foundries, memory, and packaging over 6-18 months.

Near term, the key question is feasibility, not intent. Inference chips are where cost savings matter most, but they are also where HBM access, software stack maturity, and yield discipline become gating items. That means the next 1-3 months catalyst path is mostly hiring, partner disclosures, and any evidence of taped-out silicon; until then, the move is more narrative than revenue. If the chip slips, the likely beneficiary is still the existing domestic supply chain rather than a restart in Nvidia demand.

Contrarian take: consensus may be underestimating how hard it is for Chinese AI firms to replicate a usable inference stack, and overestimating the immediate damage to Nvidia. The more durable winner, if this works, is the Chinese platform layer that can internalize compute costs; the loser is Huawei’s current share of domestic AI silicon. For BABA and BIDU, the trade is less about chip economics and more about whether cheaper inference translates into better cloud monetization without another round of capex drag.

Key Decisions for Investors

• Do not add to a short NVDA solely on this headline; treat any 2-3% gap lower as a fade unless the company later shows China mix deterioration or guidance cuts.
• Small relative-value: long AVGO / short NVDA over 1-3 months to express the broader custom-inference-ASIC theme while limiting single-name China policy risk; stop if AVGO design-win commentary softens.
• Watchlist, not buy: BABA and BIDU only become attractive if they show AI usage monetization outpacing incremental chip capex; otherwise the chip race is margin-dilutive.
• Alert item: if Chinese foundry/HBM sourcing is confirmed within 30-60 days, reconsider a basket long of China semis versus US GPU names; if not, assume the story remains aspirational.