DeepSeek hiring ramp-up

DeepSeek is hiring data center staff in Inner Mongolia as it prepares to launch its next flagship artificial intelligence model, V4. (scmp.com) The new postings are for server maintenance engineers and delivery managers in Ulanqab, and South China Morning Post reported on April 11 that they are DeepSeek’s first public on-site roles tied directly to computing infrastructure. Earlier recruiting had focused on software engineers and artificial intelligence researchers in Hangzhou and Beijing. (scmp.com) DeepSeek also added “instant” and “expert” modes to its chatbot on April 8, days before the hiring report, as it pointed users toward a V4 release expected in April 2026. The company said instant mode is for daily conversations, while expert mode is for more complex problems and may be busy at peak times. (scmp.com) Training and running a large model depends on compute, the clusters of chips and servers that do the math behind each answer. DeepSeek’s hiring in Ulanqab points to the physical side of that work: keeping machines online, opening facilities, and adding capacity before a new model goes live. (scmp.com) Inner Mongolia has become a draw for that build-out because electricity is cheaper and power is easier to secure than in China’s coastal tech hubs. China’s “Eastern Data Western Computing” plan, launched in 2022, was built around moving heavy data-processing loads from the east to energy-rich western regions. (scmp.com) The chip question is hanging over the launch. South China Morning Post said The Information reported that V4 would run on Huawei’s Ascend 950PR processor, while Reuters reported in March that the model had been trained on Nvidia Blackwell chips; DeepSeek did not respond to South China Morning Post’s request for comment. (scmp.com) That scrutiny intensified after the United States tightened rules on Nvidia’s China business in April 2025. Nvidia said on April 15, 2025 that Washington had told it on April 9 that exports of H20 chips to China would require a license, and the company said it would take a roughly $5.5 billion quarterly charge. (cnbc.com) DeepSeek’s last flagship generation helps explain why infrastructure now matters so much. In its public repository, the company said DeepSeek-V3 has 671 billion total parameters, activates 37 billion for each token, and used 2.788 million H800 graphics processing unit hours for full training. (github.com) DeepSeek kept updating that older line even as V4 slipped. On March 25, 2025, the company released DeepSeek-V3-0324 and said the update improved reasoning, front-end development, and tool use, while keeping application programming interface usage unchanged. (deepseek.com) The immediate test is whether DeepSeek can turn new racks, new power, and whatever chips it has secured into a V4 launch after more than a year of anticipation. For now, the clearest public signal is not a benchmark score but a hiring notice in Ulanqab. (scmp.com)