DeepSeek V4 Sparks Debate in the U.S.: Think Tank Claims Lagging Behind by Six Months Due to Banned Chips, Silicon Valley CEO Advocates for Open Innovation

According to monitoring by Dongcha Beating, Chris McGuire, a senior researcher on China and emerging technologies at the Council on Foreign Relations (CFR) and former member of the White House National Security Council and Department of Defense, stated that V4 has not changed the competitive landscape of AI between China and the U.S. He cited the original V4 report, which indicates that DeepSeek itself admits its reasoning capabilities are ‘approximately 3 to 6 months behind leading models,’ referring to the GPT-5.2 and Gemini 3.0 Pro released six months ago. He also questioned why the V4 report disclosed compatibility with NVIDIA GPUs and Huawei Ascend NPUs for inference but did not reveal the specific GPU models and costs used for training (V3 had claimed to use 2000 H800s at a cost of $5.57 million), suggesting that the silence implies the use of export-controlled NVIDIA Blackwell chips. Previously, U.S. government officials anonymously made similar claims in February, which NVIDIA called ‘far-fetched’; DeepSeek denied using Blackwell, stating that the model was trained on NVIDIA H800 and Huawei Ascend 910C. In contrast, Replit CEO Amjad Masad argued that while U.S. politicians and lobbyists are stoking fears of ‘Chinese distillation,’ Chinese scientists are openly sharing genuine AI breakthroughs. He referenced structural innovations listed in DeepSeek’s official tweets, including token-level attention compression (DeepSeek Sparse Attention) and significant improvements in long-context computational efficiency, pointing out that V4-Pro’s single token reasoning power and KV cache usage at 1M context are far lower than those of V3.2. Masad believes that such architectural innovations are entirely unrelated to training data distillation, and that everyone can benefit from open-source developments, including labs of all sizes in the U.S.