On April 24, DeepSeek released a preview of V4, its long-awaited flagship model. Two variants shipped: Pro and Flash. The Pro variant is the one claiming world-class reasoning and the best open-source agentic coding. The Flash variant is the smaller, cheaper one. What makes V4 a policy story rather than a release story is what it runs on. The model was adapted for Huawei’s Ascend AI processors, the Chinese chipmaker’s advanced accelerators, rather than the Nvidia hardware the entire global AI industry defaults to.
That adaptation is not cosmetic. Chinese AI firms have been cut off from the highest-performing Nvidia and AMD chips since the Biden-era export controls, a posture the Trump administration has maintained and expanded. DeepSeek’s choice to build V4 around Huawei silicon is the direct consequence of that policy. It is also the first serious test of whether the policy did what it was designed to do.
The answer is in DeepSeek’s own technical report. V4, by the company’s acknowledgment, “falls marginally short of GPT-5.4 and Gemini 3.1-Pro,” with a developmental trajectory the company says trails leading frontier models by approximately three to six months. That sentence is the US chip export control scorecard in plain language. Three to six months of lag, achieved without access to Nvidia’s top-end hardware. Not a decade. Not a generation. A quarter or two.
For context on what that means: last year’s DeepSeek V3 drop collapsed the assumption that Chinese frontier AI was structurally behind. V4 shows the pattern holding and tightening. The policy has bought time, not built a moat. Our prior coverage on the state-level AI regulation fight and the federal preemption framework the Trump administration released in March treats US AI policy as if the core variable is domestic. V4 is a reminder that the binding constraint on US AI leadership is not going to be set in any state capitol or in Washington. It is going to be set by how fast Huawei can scale Ascend production and how aggressively Beijing will subsidize the chip stack.
Price matters too. DeepSeek undercut on API pricing with V3 and has done it again with V4. Flash is explicitly positioned as a cheap variant. If the V4-Pro-versus-GPT-5.4 gap is three to six months and the V4-Flash-versus-everything-else gap is closer to parity at a fraction of the cost, the commercial layer of the frontier is being redrawn regardless of what Washington does about chips. Developers in emerging markets, research institutions in resource-constrained universities, and governments that cannot justify US hyperscaler pricing now have a credible alternative they did not have eighteen months ago.
The counter-argument worth stating is that three to six months of model lag matters more than the raw benchmark gap suggests. In frontier AI, a two-quarter advantage compounds. The most advanced users deploy the newest models first, they build the infrastructure around those capabilities, and the next generation arrives before the competition closes the previous gap. US policy is not trying to stop China from building frontier AI. It is trying to preserve a rolling lead that only works if it is continuously reinforced. By that narrower standard, the policy may still be working. The question is whether the political will and the downstream supply-chain investment will survive the next four years.
There is also a harder read of V4. If a Chinese lab can release a frontier-adjacent model trained on Huawei silicon at competitive prices, the US advantage is not the chip stack. It is the ability to turn that stack into real-world products that ship. Google, Anthropic, OpenAI, and the hyperscalers serving them are still ahead on deployment, integration, enterprise distribution, and the capital stack required to keep training at ten-figure cluster sizes. Those advantages are real. They are also orthogonal to anything the Commerce Department can control through export licensing.
Here is the short read and the long read, side by side. The short read is that US chip export policy has reached its ceiling. Three to six months of frontier lag, held under the tightest export regime the West has ever imposed on China, is a speed bump. It is not a moat.
The long read matters more. Every Western strategic-advantage play through export controls has historically had a twelve-to-twenty-four-month adaptation cycle before the target catches up. CoCom on the Soviet Union, Wassenaar on dual-use tech, even the post-Huawei 5G campaign. Twelve to twenty-four months, and then the counter-adaptation arrives. V4 is the evidence that we are now well inside that cycle on AI, and that the Chinese counter-adaptation is further along than anyone who drafted the original policy assumed.
What that means for the AI race is that the next US response cannot be about chips. The chip lever is spent. The question Washington has to answer in the next twelve months is whether industrial policy can do what trade restriction no longer can. Packaging capacity. Cluster-scale deployment subsidies. A coherent federal regime that does not trip over itself on preemption. An actual workforce pipeline. The playbook for that kind of response exists. Whether the current political coalition is capable of running it is a separate problem.
DeepSeek has a preview out. The full V4 release is coming. So is V5. The scoreboard updates on its own schedule, not Washington’s.
