DeepSeek's latest flagship model, V4, has renewed debate over the trajectory of the US-China AI race, with analysts and industry voices suggesting that China's progress in model efficiency is not translating into a meaningful reduction in capability gaps with leading US systems.
According to Bloomberg, while DeepSeek's earlier R1 model raised concerns in Washington and Silicon Valley about China narrowing the AI gap, its newly released V4 series "does not meaningfully narrow the US lead in AI capabilities." Chris McGuire of the Council on Foreign Relations said the model "is not competitive with frontier US models, and does not appear to close the gap with the US in AI".
Bloomberg also noted that even DeepSeek's internal benchmarking acknowledges a lag of roughly "3 to 6 months" behind state-of-the-art US models such as GPT-5.4 and Gemini 3.1 Pro, reinforcing the view that the performance gap remains structurally intact.
Chip access constraints increasingly define China's AI development path
A central factor driving the divergence, according to multiple sources, is access to advanced semiconductor infrastructure.
Zhang Zhi, a former ByteDance engineer and research scientist and assistant professor at Peking University, told Into Asia in an interview that Chinese AI developers remain constrained by US export controls on high-end GPUs such as Nvidia's H100, forcing many teams to rely on less capable H20 chips, which significantly slows iteration cycles. One academic cited in the report noted that while companies like Google can complete full large language model training cycles in about three months, Chinese teams often require up to six months, effectively cutting iteration speed in half due to hardware limitations.
The interview further highlights that domestic Chinese chips still lag in both performance and software ecosystem maturity, making them unsuitable for frontier-scale training in many cases. This creates a structural dependency that continues to reinforce the US lead in rapid model development cycles.
Domestic chip integration signals China's push for AI self-reliance
Despite these constraints, DeepSeek V4 also reflects China's accelerating push toward technological self-sufficiency.
CCTV-affiliated commentary cited by Yuyuantantian suggests that DeepSeek's delayed V4 release was partly due to a strategic shift toward optimizing performance on Huawei Ascend chips rather than prioritizing rapid iteration. The model has reportedly been tuned specifically for domestic hardware, moving beyond basic compatibility toward deeper hardware-software co-design.
Industry analysts cited in the same report argue that increasing adoption of domestic semiconductors is enabling early deployments of models trained entirely on local compute infrastructure. This trend is seen as part of Beijing's broader effort to build a "resilient, self-contained AI ecosystem" amid tightening external constraints.
Open source efficiency improves economics, but not frontier parity
Bloomberg also notes that while DeepSeek V4 is significantly cheaper to operate and may challenge US firms on pricing, this cost advantage does not translate into parity in frontier capabilities.
The model is described as "almost on the frontier, a fraction of the price," according to developer Simon Willison, but it remains behind leading US systems in complex reasoning tasks. Bloomberg further notes that US firms accuse Chinese models of relying on "distillation" from American systems, suggesting continued technological dependence rather than full independence.
Structural gap persists despite efficiency gains
Taken together, the sources suggest a consistent conclusion: China is improving rapidly in efficiency, cost structure, and hardware adaptation, but not closing the core capability gap with US frontier AI systems.
Bloomberg explicitly states that DeepSeek V4 "does not appear to close the gap with the US in AI," while Into Asia emphasizes that chip constraints and slower iteration cycles continue to limit China's ability to match US model development speed. Meanwhile, CCTV reporting underscores that domestic chip adoption is accelerating but remains an ecosystem in transition rather than full parity.
Article edited by Jack Wu
