China’s frontier AI training constrained by compute gaps
Epoch AI data cited by 36Kr says Chinese frontier models have averaged about seven months behind US models since 2023, while Stanford University’s 2026 AI Index reports that the performance gap has narrowed to about 2.7%. Epoch AI estimates put DeepSeek-R1 training compute at about 4×10^24 floating-point operations, versus roughly 6.6×10^25 for GPT-5 and 5×10^26 for Grok 4. The analysis examines China’s constraints in high-end GPUs, HBM, networking, software, power and cluster utilization, and recommends system engineering and model efficiency in the short term, capacity expansion and large-scale validation in the medium term, and domestic equipment and talent development in the long term.
The report provides a sourced framework for understanding why efficiency-led model development may narrow China’s AI performance gap without closing its training-compute deficit.
过去三年,中国大模型进步迅速,却始终未能真正站上全球前沿。Epoch AI的能力指数显示,自2023年以来,每一个处于前沿的模型均出自美国,中国模型平均落后约7个月,最短4个月,最长14个月。
与此同时,差距正在持续收窄。
斯坦福大学《2026年AI指数报告》显示,中美最佳模型的性能差距已由2023年5月的三成以上,缩小至2026年3月的约2.7%。DeepSeek-R1在2025年初一度追平美国最强模型,但很快又被拉开距离。
差距不断缩小,为何始终无法反超?根本原因在于训练算力。
据非营利AI研究机构Epoch AI估算,DeepSeek-R1的训练计算量约为4×1024次浮点运算,Qwen3-Max约为1.5×1025次;同期美国的GPT-5约为6.6×1025次,Grok 4约为5×1026次,高出一个数量级。
在算力受限的条件下,中国大模型走出了一条“追赶加提效”的道路。DeepSeek-V3仅用2048张H800完成预训练,依靠混合专家架构和8位精度训练,将正式训练的算力成本压缩至约560万美元(不含前期研发与试验)。同一时期,Meta训练Llama 3.1 405B动用了约1.6万张H100。
#Epoch AI#Stanford University#DeepSeek-R1#Qwen3-Max#GPT-5#Grok 4#DeepSeek-V3#Meta#Llama 3.1 405B#Huawei Ascend 950DT#NVIDIA B200#NVIDIA GB200 NVL72