DeepSeek Narrows AI Performance Gap With US Rivals to Three Percent
The performance gap between leading US and Chinese artificial intelligence models has contracted to a record low of 3 per cent, according to a Bloomberg Intelligence analysis. The rapid convergence follows the September release of DeepSeek's V4.1 Flash model, raising fresh questions regarding the long-term effectiveness of US hardware export restrictions.
Multi-head Latent Attention (MLA) compresses Key-Value cache memory via low-rank projections, allowing efficient inference and training on modest hardware. Mentions of open-weight models in US corporate earnings calls grew six-fold year-over-year in August–September 2026, accelerating enterprise model-mixing strategies.
Bloomberg Intelligence senior analyst Robert Lea reported that the benchmark performance deficit stood at 15 per cent earlier in the year before narrowing to 9 per cent in May. The latest assessment highlights how quickly Chinese laboratories are optimizing models and training methodologies to extract maximum capability from available domestic hardware.
Benchmark Performance and LiveBench Rankings
DeepSeek V4.1 Flash secured sixth position globally on LiveBench in September, achieving a score of 81.1 compared with 83.4 for leading systems from Anthropic. The result marks the highest standing achieved by a Chinese model since the release of DeepSeek's R1 reasoning model in 2025, demonstrating performance comparable to Western frontier systems.
DeepSeek's V4.1 Flash ranked sixth globally on LiveBench, becoming the highest-ranked Chinese model since the startup launched its reasoning model R1 in 2025.
Despite these capability gains, Chinese representation among top-tier models remains narrow. Bloomberg Intelligence noted that only three of the fifteen models ranked on LiveBench originate from Chinese developers, indicating that while peak performance has converged, breadth of deployment remains concentrated.
Impact of Hardware Restrictions and Efficiency
China's rapid advancement is driven by engineering efficiencies and advanced optimization for domestic processors. Analysts suggest these developments challenge the intended impact of US export controls implemented by the Bureau of Industry and Security, which aimed to restrict advanced semiconductor shipments to slow foreign competition.
While rules have seen minor adjustments allowing case-by-case reviews for certain hardware exports, Chinese labs have demonstrated an ability to achieve competitive benchmark results with significantly lower capital expenditure than major US hyperscalers, who face mounting scrutiny over multi-billion-dollar infrastructure spending.
Profitability Pressures and Domestic Market Dynamics
Despite technical parity, Chinese artificial intelligence companies encounter severe commercial headwinds. Analysts project that domestic AI operations could remain unprofitable until 2030 due to intense price competition and token-supply pricing wars.
The domestic market currently hosts more than 1,100 large language models, creating a crowded ecosystem where major consumer applications such as ByteDance's Doubao pursue monetization while competitors like DeepSeek and Tencent maintain free service models. Industry observers note that achieving long-term financial sustainability will require market consolidation and rationalized pricing structures.