Epoch AI scraped 1,604 public job listings from six Chinese AI labs, including ByteDance and DeepSeek, and found the same companies still hiring for Nvidia while staffing up on domestic alternatives like Huawei's Ascend.
ByteDance, the Chinese internet company behind TikTok, is hiring for two AI infrastructure roles that look like they belong to different eras. One posting seeks an "Inference GPU Performance Optimization Expert" with experience in Nvidia's CUDA programming software and TensorRT-LLM, the toolkit Nvidia ships for running large language models efficiently on its own chips. A separate posting from ByteDance's AI research lab, Seed, looks for an "AI heterogeneous computing optimization expert" familiar with Ascend and Cambricon, the AI accelerators designed by Huawei and by Cambricon, a separate Chinese chip designer.
The two postings sit side by side on ByteDance's careers page, and that coexistence is the spine of a new analysis from Epoch AI, a research outfit that studies frontier AI development. Epoch AI scraped 1,604 job listings from six Chinese AI labs: DeepSeek, MiniMax, Moonshot, Z.ai, ByteDance, and Alibaba. It then treated each description as a window into what those labs intend to build next, not just what they need today.
The piece is part of Epoch AI's "Gradient Updates" Substack, which the publication itself describes as "more opinionated or informal takes" than its formal research papers. The translations of Chinese-language postings were done by the post's author or by Anthropic's Claude Opus 4.8, with no independent bilingual validator. Both caveats travel with the findings. This is one research team's interpretation, not a Chinese-side primary source.
The framing matters, because the temptation is to read job postings as a roadmap. Epoch AI is careful not to. A posting asking for Ascend expertise is not proof that Ascend is running production inference at scale inside ByteDance's data centers. It is evidence the company wants to hire people who know how to make Ascend work, a leading indicator of intent rather than a snapshot of deployment.
Across the 1,604 postings, the pattern is concrete and lab-specific. Some roles cluster around Nvidia's CUDA and TensorRT-LLM ecosystem. Others lean on domestic chips. ByteDance sits on both tracks. So do several of its peers.
The technical anchors deserve a doorway. CUDA is Nvidia's programming software, used to run computations on its graphics processors. TensorRT-LLM is a layer built on top of CUDA that makes large language models run efficiently on Nvidia hardware. Ascend, designed by Huawei, and Cambricon are part of China's effort to build a domestic alternative to Nvidia's GPUs. Hiring for all four at once means a company is keeping the Nvidia pipeline warm while staffing the team that would run inference, the act of using an already-trained model to answer questions, on something else.
The six labs in Epoch AI's panel are not a monolith. DeepSeek, MiniMax, Moonshot, Z.ai (the consumer-facing brand of Zhipu AI), ByteDance, and Alibaba each have different stacks and different hiring profiles. The list also leaves out Baidu, Tencent Hunyuan, iFlytek, and JD Explore, which limits how far any "Chinese AI" claim can stretch.
The financial backdrop sharpens the picture. Zhipu AI reported 132% revenue growth in its debut earnings report, with a 73.7% gross margin in its 2025 HKEX annual results filing. A separate company using the English name MiniMax filed its own 2025 HKEX results, showing a 73% gross margin and a 9.8% operating margin; the MiniMax in Epoch AI's panel may be a different legal entity, and neither the HKEX filing nor Epoch AI confirms entity-level equivalence. Both Zhipu AI and MiniMax are generating strong revenue growth while posting substantial operating losses — each dollar of revenue comes with significant capital burn, which means any infrastructure hire is a bet on whether revenue can sustain the pace.
Epoch AI's methodology is not new. The group previously ran a similar scrape across Western AI labs in the US and UK, which makes the Chinese panel a direct comparison rather than a standalone curiosity. The cross-panel structure is what gives the analysis its leverage. A reader can hold Nvidia hiring and domestic chip hiring up against each other and watch the gap, or the absence of one, lab by lab.
The honest read is that the leading Chinese AI labs are not choosing between Nvidia and domestic silicon. They are doing both, and the public job market is where the dual-track strategy is most legible. Watch ByteDance's careers page, and the Seed team's postings in particular, as the cleanest single-company window into whether the Ascend and Cambricon tracks are scaling toward production or staying as a hedge. If roles like "AI heterogeneous computing optimization expert" keep appearing, and start asking for production-scale deployment rather than optimization research, that is the next phase of the story.