Chinese domestic AI chips 3 years behind US-Taiwan
Producing domestic chips about three years behind the U.S.-Taiwanese frontier.
What AI 2027 Predicted
The scenario states that China is “producing domestic chips about three years behind the U.S.-Taiwanese frontier.” This is presented as the state of play by mid-2026, reflecting the combined impact of US export controls and China’s efforts to build indigenous chip manufacturing through Huawei and SMIC.
How We Track This
We monitor:
- Huawei Ascend chip specifications and benchmarks vs. NVIDIA current-generation
- SMIC process node achievements vs. TSMC frontier nodes
- Industry benchmark comparisons (training throughput, memory bandwidth, power efficiency)
- Analyst reports from Institute for Progress, CFR, SemiAnalysis
- Huawei’s public Ascend roadmap and delivery timelines
Current Evidence
Huawei Ascend 910C vs. NVIDIA H200: Reuters reported (December 2025) that Huawei’s Ascend 910C — the most advanced Chinese AI chip available — “lags significantly behind the H200 in computing power and memory bandwidth,” per a report from the Institute for Progress. The H200 itself was already being superseded by NVIDIA’s Blackwell architecture (B100/B200/B300) at the time.
CFR Report on Gap: A Council on Foreign Relations report (December 2025) detailed that NVIDIA continues to lead Huawei “by a wide margin.” The report estimates that by 2027, the best US AI chips could be “more than 17 times more powerful” than Huawei’s top offerings, suggesting the gap may be widening in absolute terms even as process node gaps narrow.
Huawei’s Own Assessment: In September 2025, Huawei founder Ren Zhengfei acknowledged to Chinese state media that the company’s chips were “still a generation behind processors from the U.S.” He suggested packaging techniques could help close the performance gap. “A generation” in chip terms typically means 1.5–2 years.
Huawei Ascend Roadmap: Huawei’s planned roadmap includes the Ascend 950PR for 2026, Ascend 960 for 2027, and Ascend 970 for 2028. However, planned roadmaps frequently slip, and each generation faces the constraint of SMIC’s manufacturing capabilities (currently limited to ~7nm-equivalent processes while TSMC is at 3nm and moving to 2nm).
Production Volume Constraints: US officials estimate Huawei will produce only a few hundred thousand advanced AI chips in 2025, a fraction of NVIDIA’s output. This limits China’s ability to close the compute gap even as individual chip performance improves.
Tension with model gap: China’s domestic chips are approximately 3 years behind the US-Taiwanese frontier (this prediction), yet China’s best models are only 3-8 months behind US frontier models (see Leading Chinese AI lab ~6 months behind US frontier). This implies massive algorithmic compensation — Chinese labs are extracting far more capability per FLOP through training efficiency innovations. DeepSeek R1 (January 2025) demonstrated this dramatically: matching US frontier models at a reported total training cost of $5.6M, causing Nvidia to lose $589B in market value in a single day. The chip gap may matter less than expected if algorithmic efficiency continues to compensate.
Counterevidence & Limitations
- “Three years behind” is an approximation that conflates process node gap, architectural gap, and software ecosystem gap. The true picture is more nuanced.
- Huawei’s Ren Zhengfei said “one generation behind” (roughly 1.5–2 years) — this suggests the gap might be smaller than AI 2027 predicted, at least by Huawei’s own assessment.
- China is pursuing alternative approaches (advanced packaging, chiplet architectures) that could narrow the effective performance gap even without matching process nodes.
- The comparison depends heavily on what metric is used. In raw FP16 throughput, the gap may be smaller than in memory bandwidth or energy efficiency.
- SMIC’s progress on advanced nodes has been faster than some analysts expected, suggesting the gap could narrow.
What Would Change Our Assessment
- Upgrade to “confirmed”: Independent benchmarks showing Huawei’s best chip roughly matching NVIDIA chips from 2023 (i.e., A100/H100-era) by mid-2026
- Downgrade to “behind”: Evidence that China is closing the gap faster than predicted (e.g., Huawei matching H200-class performance by 2026)
- Complicate assessment: If the gap is better measured as “2 years in process node but 4+ years in software ecosystem and production volume”
Update History
| Date | Update |
|---|---|
| 2025-04 | US imposes indefinite H20 export license requirement (April 9). H20 was the last Nvidia chip available to Chinese buyers without license. METR June data suggests gap is real — Chinese models lag ~6 months on agentic tasks, partially attributable to compute constraints. |
| 2026-03 | Huawei’s Ascend 910C significantly lags NVIDIA H200 in performance. Ren Zhengfei publicly acknowledges being ‘one generation behind.’ A ~3-year gap is a reasonable characterization of the current state. |