The release of DeepSeek V4 has prompted a reassessment of the artificial intelligence balance of power between the United States and China. A new analysis published by the Council on Foreign Relations (CFR) argues that while the model does not close the capability gap with the American frontier, its architectural innovations and aggressive pricing strategy signal a dangerous new phase in the geopolitical rivalry.
The CFR report, published on May 1, 2026, provides a nuanced view of DeepSeek’s latest offering. It acknowledges that V4 produced a “more muted” market reaction compared to the explosive debut of the R1 model in January 2025, which briefly tanked US tech stocks. However, the analysis warns against complacency, noting that China is successfully exploiting access to US technology to remain highly competitive.
The Capability Gap and the Distillation Threat
The CFR analysis aligns with the recent CAISI evaluation of DeepSeek V4, confirming that the Chinese model trails state-of-the-art US systems by approximately three to six months. DeepSeek’s own technical paper concedes that V4’s reasoning and agentic capabilities are comparable to models released by OpenAI and Anthropic roughly half a year ago.
However, the CFR report highlights a critical vulnerability in the US lead: the threat of “illicit distillation attacks.” The White House and leading US AI labs have accused Chinese firms, including DeepSeek, of training their models using data generated by interacting with advanced American models. This practice allows Chinese developers to replicate the capabilities of US models at a fraction of the original research and development cost.
“DeepSeek’s aggressive use of distillation attacks against US models… saves it hundreds of millions or even billions of dollars in R&D costs,” the CFR analysis notes. This dynamic allows Chinese firms to offer highly capable models at rock-bottom prices, undercutting American competitors in global markets.
Hardware Constraints and Architectural Innovation
The CFR report also delves into the hardware dynamics shaping the rivalry. DeepSeek V4 is notable for being optimized for inference on Huawei’s Ascend chips, reportedly at the direction of the Chinese government. This aligns with broader trends, such as Tencent Cloud deploying DeepSeek V4 on Huawei silicon.
Yet, DeepSeek remains heavily dependent on US technology. US government officials assert that V4 was trained on smuggled Nvidia Blackwell chips, which are banned for export to China. Tellingly, the V4 technical report is silent on the specific hardware used for training.
Despite these constraints, DeepSeek has achieved significant engineering breakthroughs. The CFR highlights V4’s hybrid attention architecture, which enables a one-million-token context window at a fraction of the inference compute cost of its predecessor. “This indicates that Chinese AI labs are largely keeping pace with advancements in algorithmic efficiency that are also occurring in closed-source US AI labs,” the report states.
Policy Recommendations
The CFR analysis concludes with a stark warning and a call to action for US policymakers. “DeepSeek V4 is a reminder that in the race to develop the most advanced capabilities, the United States is ahead, but China is not far behind,” the report asserts.
To widen the lead, the CFR recommends that the United States “go on offense against adversarial distillation” and close existing loopholes in export controls. By fully restricting China’s access to US AI chips, models, and chipmaking tools, the report argues, the US could open up a lead measured in years rather than months. As the technological decoupling accelerates, the battle over distillation and hardware access will define the next chapter of the AI cold war.
The CFR analysis also notes that DeepSeek V4 Pro is not even the most cost-competitive Chinese model available. Kimi K2.6 from Moonshot AI and GLM 5.1 from Zhipu AI perform comparably to V4 on most public benchmarks and are priced lower. DeepSeek V4 Pro is more expensive than either. This competitive dynamic within the Chinese AI ecosystem, where multiple labs are producing frontier-class models at commodity prices, is itself a strategic challenge for the United States, regardless of the precise capability gap at the absolute frontier. The CFR concludes that the US government must treat the AI rivalry not as a sprint to a single finish line but as a sustained competition across multiple dimensions: frontier capability, cost efficiency, hardware independence, and global market penetration, all of which China is actively contesting.