DeepSeek is reportedly designing its own AI chip to sidestep US curbs

DeepSeek is reportedly developing its own AI chip, moving it from writing software that runs on other companies’ hardware to specifying its own silicon. If this transition materializes, it strengthens China’s vertically integrated AI stack, though the impact remains uncertain given the report is based on sources.

Analysis

The market should treat this less as an immediate earnings event and more as a bargaining-power shift. If DeepSeek can specify its own silicon, the key economic effect is not headline chip revenue but tighter control over inference cost per token, which can widen model margins and reduce dependence on imported accelerators over time. That is structurally negative for foreign GPU vendors at the margin, but only if the design is paired with a credible manufacturing path and software stack; otherwise it remains a strategic signal rather than a volume threat.

Near term, the likely winners are the Chinese semiconductor plumbing names that sit closest to design-to-fab translation: foundry, advanced packaging, and domestic accelerator ecosystems. The second-order effect is that local hyperscalers may accelerate dual-sourcing and benchmark domestic chips for lower-end inference workloads first, which pressures premium pricing on imported GPUs before it meaningfully displaces unit demand. The overhang for NVDA/AVGO/TSM is more about sentiment and future China TAM haircut than immediate lost revenue.

The contrarian risk is that the market may be underestimating how quickly a software-native lab can optimize around constrained hardware, but also overestimating the probability of successful leading-edge production. The falsifier is simple: no foundry relationship, no package tape-out, no meaningful deployment footprint in 1-3 months. If there is still no evidence of volume by the next earnings season, this becomes a narrative trade, not a fundamental one.

Key Decisions for Investors

• Do not chase the headline in NVDA/TSM today; wait 1-3 weeks for confirmation of a foundry or packaging partner before expressing a view.
• If confirmation emerges, consider a small relative-value long SMIC (0981.HK) / short NVDA pair for 1-3 months; the trade works only if the story shifts from R&D to production.
• Add Cambricon (688256.SS) or Hygon (688041.SS) to the watchlist as domestic accelerator beneficiaries, but size only on evidence of procurement or design wins.
• For portfolio hedging, use the move as an alert to trim overweight exposure to AI hardware multiples that already discount China demand; the risk is multiple compression more than immediate revenue loss.