Huawei's 2027 AI Chip Will Trail Nvidia by Three Generations—And Sell to Half of China Anyway

Huawei's own keynote, transcribed from the September 18, 2025 speech, is unambiguous on the 960: it doubles compute, doubles memory bandwidth, doubles memory capacity, and doubles the number of interconnect ports versus the 950. It also introduces HiF4, Huawei's proprietary 4-bit numeric format. What the keynote does not promise is parity with Nvidia's current Blackwell line. Independent analysis from Epoch AI puts the on-paper gap starkly: the 950PR is roughly half of an Nvidia H100 from 2022; the B300 delivers about 6.75x the FP4 throughput of the 950. The 960 closes part of that gap, but it does not close it.

The 2027 chip is the story investors will trade in 2026, but the market-share story has already happened. Bernstein, cited by AP and the Washington Post, estimates Nvidia's China AI accelerator share collapsed from about 40% in 2025 to roughly 8% in 2026, while Huawei rose from a similar 40% to about 50%. TrendForce's separate domestic-server count is even more decisive: Chinese suppliers led by Huawei and Cambricon moved from 46% of the domestic AI server market in 2025 to 56% in 2026, with foreign suppliers dropping from 34% to 21%.

Whether the Ascend 960 ships on time depends on three bottlenecks Huawei does not control: leading-edge wafer capacity at SMIC, HBM supply at CXMT, and competitive pressure from Cambricon for the second-tier Chinese buyer. All three are moving in the right direction in 2026—but none of them is at the scale Nvidia's suppliers enjoy.

The paradox of the 2027 roadmap is that the loser of the China market is the biggest winner of everything else. Nvidia reported Q2 FY2027 revenue of $96.2 billion with data-center revenue of about $89 billion, and guided Q3 FY2027 to $108 billion (plus or minus 2%), well ahead of the $104 billion consensus. Hyperscaler revenue of roughly $49 billion grew 13% sequentially. The China hole, the H200 reopening, the Rubin ramp—they are line items, not the story. The story is that the AI capex cycle has years left.

The 2027 chip will not struggle for buyers. Alibaba, Tencent, ByteDance, and Baidu together are on track for roughly $102 billion of AI capex in 2026, and they are not picking sides. ByteDance alone has earmarked about $5.6 billion for Huawei Ascend purchases in 2026 (roughly RMB 40 billion) on top of a separate $14 billion Nvidia H200 budget. Alibaba raised $10.2 billion in August 2026 specifically for AI infrastructure, halfway through a $56 billion three-year capex plan. The result is a bifurcated compute stack inside every major Chinese hyperscaler: Huawei for domestic training scale, Nvidia for advanced workloads and offshore inference.

Strip the geopolitics away and the 2027 Huawei chip is a known quantity: an H100-class AI accelerator on SMIC's 7nm, with HBM3E from CXMT, scaling to a million-NPU cluster. The honest read is that this is a 2022-vintage Nvidia chip shipping in 2027, by which time Nvidia's Rubin generation is ramping. The 50% China share Huawei has today does not require catching up on hardware; it requires only continuing to ship at a price Chinese hyperscalers will pay.

For investors, the practical question is who catches the spillover. TSMC is the structural beneficiary: the only foundry that can serve Nvidia's actual demand, and the only place outside China with the capacity to absorb the redirected spend. SK Hynix and Micron supply the HBM Nvidia needs; CXMT is the bottleneck Huawei is trying to break. SMIC trades at 9.4x trailing sales with margin expansion in flight. Cambricon is the second-source story, not the main one. And Nvidia, even after conceding China, is still the world's most profitable semiconductor company by a wide margin, with $108 billion of Q3 FY2027 revenue to prove the rest of the world is still buying.