Podcast Episode
DeepSeek V4: China's Trillion-Parameter Bet on Homegrown Chips
April 13, 2026
0:00
3:18
DeepSeek founder Liang Wenfeng has confirmed that the company's highly anticipated V4 artificial intelligence model will launch in late April 2026. The trillion-parameter model will run entirely on Huawei's Ascend chips, marking a major milestone in China's push to build frontier AI without American silicon.
China's Biggest AI Model Yet
DeepSeek's founder Liang Wenfeng has confirmed that the company's next-generation V4 model will arrive in late April, ending months of speculation and missed deadlines. The model is a one-trillion-parameter Mixture-of-Experts system, though only around thirty-seven billion parameters activate per response, keeping it fast and efficient in practice.Built on Huawei Silicon
In a move with major geopolitical implications, V4 has been built to run entirely on Huawei's Ascend 950PR processors and chips from Cambricon Technologies. DeepSeek reportedly spent months rewriting its model stack to work with the domestic hardware, deliberately withholding early access from American chipmakers including Nvidia and AMD. Major Chinese technology firms including Alibaba, ByteDance, and Tencent have placed orders for hundreds of thousands of the 950PR chips, with Huawei planning to deliver roughly 750,000 units this year.A Grey Release Already Underway
A limited grey release of V4 has begun rolling out to select users, featuring a one-million-token context window and new layered modes for speed and expert reasoning. A notable service outage in late March is widely believed to have been a stealth infrastructure test. Prediction markets put the probability of a full public release before the end of April at ninety-four percent.The Export Control Question
The launch arrives against the backdrop of intensifying US-China tensions over chip exports. Washington's export controls, first imposed in late 2022 and repeatedly expanded, have forced Chinese AI labs to seek alternatives to American hardware. DeepSeek CEO Liang has described the chip ban as the greatest challenge facing his company, noting that Chinese firms must use two to four times as much computing power to match results achievable on restricted US hardware. If V4 performs as advertised, it could validate Beijing's parallel AI infrastructure and reshape the debate over whether export controls are containing China's AI progress or merely redirecting it.Published April 13, 2026 at 8:13am