Alibaba has developed a new artificial-intelligence chip aimed at handling a wider range of AI inference tasks and, crucially, remaining compatible with Nvidia’s software stack, letting engineers reuse programs written for Nvidia hardware. The processor is fabricated in China, a shift from Alibaba’s earlier AI chip that was built at TSMC in Taiwan. The part is in testing now.
Why this matters
- Software compatibility: WSJ reports Alibaba’s new chip works with Nvidia tools, lowering switching costs for developers who have built on CUDA-based workflows.
- Made in China: Local fabrication reduces exposure to export curbs that have complicated Nvidia’s China sales.
- Timing: It lands as Chinese regulators urge or caution domestic firms about buying Nvidia’s H20 chips, citing security concerns—complicating Nvidia’s path even after Washington allowed resumed H20 sales under certain license terms.
Backdrop: China is racing to fill the Nvidia gap
- Policy pressure: Bloomberg and Reuters report Beijing has urged companies to avoid Nvidia’s H20 (especially for government uses) and cautioned on purchases over information-risk concerns; The Information and others said Nvidia even asked suppliers to pause H20 work last week.
- Local substitutes:
- MetaX (Shanghai) rolled out an H20 replacement in July, claiming larger memory for memory-heavy AI tasks (with higher power draw).
- Cambricon shares have surged; the firm flagged ¥5–7bn 2025 revenue guidance and warned on trading risks after a sharp rally.
- State funding: Beijing launched a ~¥60bn (US$8.2bn) AI fund this spring to back early-stage AI projects and chip independence.
What’s actually new about Alibaba’s chip
Broader inference focus, Nvidia-tool friendly, China-built. According to WSJ, Alibaba’s part is more versatile than its predecessors and compatible with Nvidia software, letting teams repurpose existing code. Manufacturing is domestic, unlike the prior generation made at TSMC. It is not pitched as a training monster to rival Nvidia’s H100/Blackwell, but as a practical inference workhorse for Chinese data centers today.
Reality check: China’s strengths and gaps
- Inference is catching up: Multiple Chinese vendors (Huawei, Alibaba, MetaX, Cambricon) are pushing good-enough inference for many workloads.
- Training still lags: Due to export limits on cutting-edge tools and manufacturing, China remains behind on the highest-end training chips and reliable mass production at advanced nodes.
What to watch next
- Specs & benchmarks: Alibaba has not publicly disclosed full specs or MLPerf-style results; watch for model throughput, memory bandwidth, and power numbers to compare vs. H20 and domestic peers. (WSJ/Reuters say chip is in testing.)
- Toolchain details: How complete is the CUDA-compatibility story (operators, kernels, driver stack)? Developer friction will decide real adoption.
- Procurement signals: If Chinese cloud players shift AI inference orders from Nvidia to domestic parts, you’ll see it in cloud segment updates and order disclosures.
- Policy trajectory: Any formal guidance from Beijing on H20 usage—and Washington’s licensing terms—will shape both Nvidia’s and local vendors’ near-term share.
Alibaba’s move gives China another homegrown inference option that feels familiar to Nvidia-trained engineers—reducing the switching cost at exactly the moment regulators are steering buyers toward domestic chips. But closing the gap on frontier training silicon remains the long pole.
Disclosure: This article does not represent investment advice. The content and materials featured on this page are for educational purposes only.
Related: Nvidia Q2 2026 Earnings Preview and Prediction: What to expect
Why Warren Buffett and Hedge Funds Are Betting on UnitedHealth Stock (UNH)
