DeepSeek’s latest AI model, V4, was supposed to arrive sooner. Instead, the Chinese AI lab spent months reworking its software stack to run on Huawei’s domestically produced Ascend processors, according to a social media account affiliated with China’s state broadcaster CCTV. The delay, first reported by Bloomberg in late April 2026, points to a deliberate trade-off: sacrificing speed to market in favor of breaking free from restricted Western chips.
The decision lands at a charged moment. U.S. Commerce Department export controls, first imposed in October 2022 and tightened multiple times since, have cut off Chinese companies from Nvidia’s most powerful GPUs, including the A100 and H100. DeepSeek’s earlier models, V3 and R1, already stunned the AI world by delivering frontier-level performance at a fraction of competitors’ costs. V4 raises the stakes further: if it can match that standard on Chinese-made silicon, it would undermine a core assumption behind Washington’s chip restrictions.
What the sources actually say
The chip-migration narrative traces to Yuyuantantian, a social account tied to CCTV’s international commentary operation. Bloomberg’s reporting described “months of deep engineering work” to optimize V4 for Huawei’s Ascend chips, going far beyond basic compatibility to restructure how the model’s training and inference routines interact with the processor’s architecture. Bloomberg framed the effort squarely within Beijing’s campaign for technological self-reliance.
DeepSeek did release a preview of V4 in April 2026, describing it as an open-source update. The Associated Press reported that the model showed notable gains in mathematical reasoning and code generation, though it noted the absence of independent benchmarks to verify those claims against competing systems. The gap between when analysts expected the model and when it actually appeared lines up with the reported optimization timeline.
Georgetown University researcher William Hannas, quoted by The Washington Post, described V4 as “a milestone for China’s AI industry” and its drive toward independence from Western supply chains. That on-the-record assessment reinforces the idea that the delay was strategic rather than accidental, a calculated bet on long-term hardware sovereignty over short-term competitive positioning.
Why the open-source release matters
By publishing V4’s model weights publicly, DeepSeek has effectively invited the global research community to stress-test its work on Huawei hardware. That transparency cuts both ways. If outside developers confirm strong performance on Ascend chips, it validates China’s semiconductor ambitions in a way no government press release could. If testing reveals bottlenecks in speed, memory handling, or energy efficiency compared with Nvidia-based systems, those gaps will be impossible to hide.
For companies and research labs operating under export restrictions, or those simply looking to diversify their chip supply chains, the open-source approach turns V4 into a live experiment. Every benchmark result will carry weight beyond the technical, feeding directly into procurement decisions, policy debates, and investor calculations about the future of AI hardware.
What we still do not know
Several critical gaps remain. Neither DeepSeek nor Huawei has issued a formal statement confirming the scope of the software rework or specifying which Ascend chip variant, such as the 910B or 910C, served as the primary optimization target. The Yuyuantantian account carries institutional weight in China’s media landscape, where state-affiliated channels often signal official positions, but a social media post is not a corporate disclosure or a technical white paper.
No independent performance benchmarks for V4 on Ascend hardware have been published as of early May 2026. DeepSeek’s own claims about reasoning and coding improvements have not been verified against standardized tests or compared head-to-head with models running on Nvidia’s H100 or H200 GPUs. Without that data, it is impossible to say whether the Huawei optimization produced competitive performance or required meaningful sacrifices in speed, accuracy, or cost efficiency.
Beijing’s policymakers have not publicly detailed how V4 fits into national technology strategy. The self-reliance framing comes from analysts, experts, and state-adjacent media rather than from official policy documents tied to this specific release. Whether the Huawei integration was driven by a government directive, by DeepSeek’s own supply-chain calculations, or by some blend of informal guidance and commercial necessity remains an open question.
The preview status of the release adds another layer of uncertainty. Performance characteristics, hardware requirements, and optimization levels can all shift between a preview and a general-availability launch. Conclusions about V4’s viability on domestic chips should be treated as provisional until a full release and independent testing follow.
When benchmarks arrive, the real verdict begins
The strongest evidence supporting this story comes from institutional reporting by Bloomberg, the AP, and The Washington Post, outlets that applied editorial standards before publishing and whose accounts are broadly consistent on timing, the open-source release, and the connection to China’s self-reliance push. The Yuyuantantian account adds a layer of semi-official signaling, directionally useful but not independently verifiable on technical specifics.
For the global AI industry, the practical question is straightforward: can Huawei’s Ascend chips handle frontier model workloads at a level that competes with Nvidia? DeepSeek’s engineering effort to port V4 is real. The open-source release guarantees scrutiny. And the geopolitical stakes ensure that every benchmark, once published, will be parsed for meaning far beyond the technical.
The answer will arrive in the form of data, not declarations. When independent researchers begin publishing results from V4 running on Ascend hardware, the numbers will either validate China’s semiconductor ambitions or expose how far the gap still stretches. Until then, the delay itself is the clearest signal: DeepSeek chose to bet on domestic chips, and it was willing to keep the world waiting to do it.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
