Factbox-DeepSeek-V4, the Chinese AI model adapted for Huawei chips

DeepSeek previewed V4, a new open-source AI model designed to run on Huawei chips, with two variants: Pro and Flash. The model supports a 1-million-token context window and is positioned as stronger in agentic coding, STEM, and competitive programming, though Pro still trails top closed-source systems in some areas. The release underscores China’s push for AI hardware self-sufficiency amid U.S. export controls, but compute constraints keep Pro availability limited for now.

Analysis

The real tradeable shift is not the model itself but the hardware validation loop it creates for China’s domestic AI stack. If a frontier open model can be tuned to run acceptably on Ascend, the marginal buyer of inference and agentic workloads in China becomes less Nvidia-dependent, which is a structural negative for NVDA’s China mix and a medium-term positive for Huawei-adjacent ecosystem names. The second-order effect is that software developers optimizing for domestic chips may begin to set de facto standards around them, creating a local moat that compounds through tooling, compilers, and deployment libraries rather than raw silicon performance.

The biggest near-term beneficiary is not the chip vendor alone, but the broad set of Chinese cloud, enterprise software, and systems integrators that can market “sovereign AI” deployments without US supply-chain exposure. That should help domestic AI adoption economics over the next 6-18 months, especially in regulated verticals where procurement risk matters as much as benchmark performance. The flip side is that the current cap on higher-end service availability implies monetization is still constrained by compute scarcity, so the adoption story can outrun revenue realization for several quarters.

For GOOGL, this is a reminder that open-source models are still closing the gap in agentic workflows, which compresses differentiation at the application layer and raises the importance of proprietary data and distribution. The market often overprices “model quality” and underprices “runtime control”; if enterprises can get 80-90% of the capability on cheaper infrastructure, vendor switching costs fall faster than expected. That’s mildly bearish for the premium cloud/AI stack in the near term, but not enough to change the secular thesis unless domestic Chinese deployment scales faster than expected.

Contrarian read: the move is likely more important as a policy signal than as an immediate earnings signal. The market may be underestimating how quickly domestically optimized AI could become the default procurement standard in China, but overestimating the speed at which that translates into true Nvidia replacement at scale. The constraint is not just chip performance; it is the ecosystem depth of CUDA, tooling, and developer familiarity, which likely keeps NVDA relevant globally even if its China share erodes.

Key Decisions for Investors

• Short NVDA on rallies over the next 1-3 months via a China-policy basket hedge; risk/reward favors a modest tactical short because incremental China sentiment can compress multiple even if core demand remains intact.
• Long a China domestic AI ecosystem basket for 6-12 months (Huawei supply chain / local cloud / enterprise software proxies) as a way to express the self-sufficiency theme without taking single-name execution risk.
• Pair trade: long GOOGL / short NVDA for 3-6 months if AI market leadership continues shifting toward agentic software and away from GPU-driven scarcity economics; use tight stop if China hardware adoption disappoints.
• Avoid chasing headline strength in open-model beneficiaries until service availability and monetization improve; wait for evidence of sustained domestic procurement, not just benchmark wins, before adding size.
• For higher-conviction hedging, buy medium-dated NVDA put spreads financed against a basket of China tech longs; this captures policy-driven downside while limiting premium bleed if the market rotates back to broad AI beta.