Chinese AI Models Narrow Gap with U.S. Leaders as DeepSeek’s V4.1 Flash Gains Ground

Senior analyst Robert Lea of Bloomberg Intelligence reported that, as of the Bloomberg Intelligence report published on October 4, top Chinese large-language models trail their U.S. counterparts by only 3 percent on benchmark scores. The narrowing follows the September release of DeepSeek’s V4.1 Flash model, which achieved a LiveBench score of 81.1, compared with Anthropic’s best score of 83.4. LiveBench ranked V4.1 Flash sixth globally, the highest position for a Chinese model since DeepSeek’s R1 debut in 2025.

U.S.-China AI Gap Narrows to 3% as DeepSeek Matches Anthropic's Performance — BigGo Finance

The current gap contrasts with a 15 percent difference earlier in the year and a 9 percent gap in May, according to Lea. The shift echoes a January 27 2025 event when Nvidia’s market value fell by $589 billion after investors questioned the need for costly U.S. chips given Chinese labs could produce comparable models at lower cost. Although Nvidia’s shares recovered, the episode highlighted vulnerability in the U.S. hardware-centric AI strategy.

DeepSeek Narrows AI Gap With US to Just 3 Percent, Bloomberg Says

• LiveBench gap: 3 percent (Oct 4) vs. 9 percent (May) vs. 15 percent (early 2026).

Bloomberg

• DeepSeek V4.1 Flash score: 81.1 points; Anthropic top score: 83.4 points.

U.S.-China AI Gap Narrows to 3% as DeepSeek Matches Anthropic's Performance — BigGo Finance

• Global ranking: DeepSeek V4.1 Flash placed 6th; only three of the top 15 LiveBench models are Chinese.

U.S.-China AI Gap Narrows to 3% as DeepSeek Matches Anthropic's Performance — BigGo Finance

• Model ecosystem: Over 1,100 large-language models are active in China, creating intense price competition and low-margin token supply, according to Lea.

Bloomberg

The U.S. Bureau of Industry and Security issued a rule in January 2026 that shifted export reviews for Nvidia’s H200 and AMD’s MI325X chips from a presumption of denial to case-by-case approval, provided applicants demonstrate that shipments will not starve U.S. customers and that Chinese buyers meet compliance conditions. This adjustment reflects a loosening of the export-control strategy that originally aimed to limit Chinese access to advanced GPUs.

DeepSeek Narrows AI Gap With US to Just 3 Percent, Bloomberg Says

In a September 30 announcement, DeepSeek and Huawei released open-source programming libraries for Huawei’s Ascend 950 AI chips, including DeepGEMM-Ascend for matrix operations and DeepEP-Ascend for inter-chip communication. The tools also add support for the TileLang programming language, which DeepSeek describes as offering a “simpler programming model” than Nvidia’s CUDA. Huawei indicated that its Ascend 950 supernodes fully support DeepSeek’s V4 models and that the chips contributed to training the lighter V4-Flash model.

DeepSeek and Huawei release open-source Ascend AI programming tools to reduce reliance on Nvidia CUDA ecosystem — tools include compute and communication libraries, as well as Ascend support for TileL U.S.-China AI Gap Narrows to 3% as DeepSeek Matches Anthropic's Performance — BigGo Finance

Lea projected that China’s AI sector may remain unprofitable until 2030 because of a brutal price war and low-margin token supply. He suggested that sustainable profitability will require a cooling of competitive pressures, an industry shake-out, and more rational pricing. The narrowing benchmark gap, combined with open-source tooling that reduces reliance on Nvidia’s ecosystem, raises questions about the long-term effectiveness of U.S. export controls and the sustainability of U.S. technological dominance in AI.

Bloomberg