Zhipu AI jumps on report it is exploring custom AI chip for GLM models

Zhipu AI shares in Hong Kong jumped 9.9% after The Information reported the company is exploring a custom ASIC for its GLM large language models. The project is early (likely >2 years) but reflects accelerating demand—daily token usage rose 27x in the first week after the GLM-5.2 launch—amid U.S. export restrictions limiting access to Nvidia’s most advanced chips. The initiative aims to reduce reliance on third-party suppliers and lower AI inference costs/energy use, with the next key catalyst being selection of a chip design partner.

Analysis

The signal is less about one issuer and more about the direction of travel: Chinese AI buyers are moving from “rent GPU capacity” to “own the inference stack.” That is a structural negative for foreign accelerator pricing power over time, but the near-term earnings impact on NVDA is limited because ASIC programs are slow, software-heavy, and usually start with narrow inference workloads rather than broad model training. The first-order market move is likely sentiment-driven; the second-order effect is a gradual transfer of spend toward domestic chip design, packaging, and foundry capacity.

For BABA, the read-through is mildly constructive if it helps normalize domestic AI infrastructure spending and lowers inference costs for cloud and application layers. For GOOGL, the broader validation of custom silicon economics reinforces the strategic premium for firms that control both model distribution and compute architecture, but it is not a direct China revenue story. The real beneficiaries are likely the suppliers behind the curtain, while the app-layer company itself only wins if cheaper compute converts into faster monetization.

The contrarian mistake is to treat this as an immediate competitive threat to NVDA. An ASIC only matters after partner selection, tape-out, yield, software porting, and scale deployment; that is a 12-24 month path, not a tape-worthy overnight shift. The better short thesis is not “NVDA loses China tomorrow,” but “the market is overpaying for the scarcity narrative if inference workloads keep migrating to bespoke silicon.”

Key Decisions for Investors

• Do not short NVDA outright on this headline; if anything, use strength to add only as part of a hedged basket, because the fundamental hit is delayed and contingent on design execution.
• Initiate a 3-6 month pair: long GOOGL / short NVDA, expressing the view that vertically integrated AI compute stacks deserve a premium while GPU pricing power faces gradual erosion.
• Small tactical long BABA on pullbacks as a domestic AI infrastructure beneficiary, but keep size modest; thesis breaks if China tech capex or cloud demand disappoints.
• Set an alert for any named chip partner or tape-out milestone in 1-3 months; that is the first verifiable catalyst that would convert this from narrative into real capex displacement.
• If NVDA reports no deterioration in China-linked demand or data center guidance over the next two earnings cycles, cover any anti-NVDA exposure—the market is likely over-discounting the near-term impact.