Alibaba unveils new AI chip in push for domestic alternatives

Alibaba unveiled its Zhenwu M890 AI chip, which it says delivers 3x the performance of the prior Zhenwu 810E and is designed for AI agent workloads. The company also laid out a chip roadmap with the V900 due in Q3 2027 and J900 in Q3 2028, signaling continued in-house silicon investment amid U.S. export curbs on advanced Nvidia processors. Alibaba said it has shipped over 560,000 Zhenwu units and that its new Panjiu AL128 server is available immediately to Chinese enterprise customers.

Analysis

Alibaba is signaling that the next leg of China’s AI capex cycle is moving from model training rhetoric into deployment infrastructure. The most important second-order effect is that agent workloads are memory- and bandwidth-intensive, which favors vertically integrated cloud stacks and accelerator clusters over standalone model providers; that shifts bargaining power toward domestic hyperscalers that can bundle chips, servers, cloud, and model access into one procurement decision.

For NVDA, the immediate revenue impact is likely limited, but the strategic risk is that China’s demand mix becomes progressively less addressable just as local alternatives improve enough to absorb enterprise workloads. The bear case is not a sudden loss of Chinese revenue; it is a slow erosion of pricing power and design wins in edge-to-midrange inference, where domestic chips can win on availability, compliance, and system-level integration even if they remain behind at the frontier.

The more interesting catalyst set is over the next 12-24 months: if agent adoption accelerates, compute demand should re-accelerate across domestic Chinese cloud budgets, but the spend may increasingly bypass U.S. silicon entirely. That creates a near-term paradox: the broader AI capex wave remains bullish for semicap equipment, memory, and networking, while the China-specific supply chain becomes a headwind for NVDA and potentially a tailwind for local server integrators and cloud platforms.

Consensus is probably underestimating how quickly enterprise buyers in China will accept “good enough” chips when the alternative is export uncertainty and supply fragmentation. The market may also be overestimating how much of Nvidia’s China exposure is protected by software lock-in; for agentic inference, software moat matters less than cluster-level performance per watt and procurement certainty.

Key Decisions for Investors

• Reduce NVDA exposure tactically on strength over the next 1-3 weeks; use rallies to trim rather than chase, since the risk is gradual share-loss in China rather than a single headline event.
• Initiate a pair trade: long selected China AI infrastructure beneficiaries vs short NVDA into the next 1-2 quarters; focus on the thesis that domestic demand will be redirected to local stacks as agent workloads scale.
• Stay constructive on semicap and networking beneficiaries of rising AI capex over 6-12 months; the domestic China substitution story is bearish for NVDA but not for the broader global compute buildout.
• For more aggressive expression, buy NVDA 3-6 month puts financed with upside calls if implied vol is elevated; the trade targets a valuation compression from incremental China de-risking while limiting premium outlay.
• Watch for evidence of enterprise deployment uptake in China over the next two reporting cycles; if domestic chip shipments accelerate faster than expected, extend the NVDA underweight and avoid mean reversion complacency.