Will a Chinese domestic AI chip achieve verified Nvidia H100 performance parity by Dec 31, 2028?

Chinese domestic chip development has accelerated under export control pressure, with Huawei Ascend 910C/910D and SMIC-fabricated alternatives showing measurable capability gains. By December 2028, the H100 will be a six-year-old part — single-chip parity with an aging benchmark is materially easier than frontier parity. The verification path is the critical question: Chinese vendors do not submit to MLPerf, so the path to 'verified parity' runs through SemiAnalysis-style independent system testing or equivalent independent technical media coverage of standardized benchmarks. The political incentive to claim parity exceeds the technical incentive to demonstrate it; verification standards matter.

RESOLUTION CRITERIA

TRUE if a Chinese-designed, China-fabricated AI chip is verified by independent third-party benchmark (not vendor or state-sponsored testing) to match or exceed Nvidia H100 performance on standardized AI training (MLPerf) or inference (industry-standard benchmarks) workloads, with results published in peer-reviewed journals or independent technical media (e.g., SemiAnalysis, Chips and Cheese, equivalent) before Dec 31, 2028.

▲ FOR

Huawei Ascend 910C/910D roadmap claims 2025-2027 parity, with 910D sampling May 2025

SMIC 7nm production ramping despite EUV restrictions

State-coordinated chip industry investment exceeds $150B since 2023

DeepSeek and Qwen demonstrate frontier capability on domestic silicon

H100 is a 2022 chip; by Dec 2028 it will be six years old, lowering the parity bar substantially

Inference parity on Chinese-optimized workloads is closer to current state than training parity; SemiAnalysis testing shows CloudMatrix 384 already outperforming GB200 NVL72 on some metrics

▼ AGAINST

Parity claims historically fail independent verification

Software stack (CUDA equivalent) lags hardware by 18+ months

Export controls limit access to advanced lithography

Real-world workload performance gaps consistently exceed benchmark gaps

Chinese vendors do not submit to MLPerf; verification path requires SemiAnalysis or equivalent independent third-party testing