As part of Epoch AI’s Gradient Updates, researchers examined 1,604 job postings from six notable Chinese AI organisations—DeepSeek, MiniMax, Moonshot, Z.ai, ByteDance, and Alibaba—to infer hiring priorities and strategic directions. The postings provide direct signals about what skills firms are seeking and, by extension, the constraints and plans shaping their work. Below are the main findings drawn from the cited job descriptions.
Continued Nvidia use, alongside exploration of domestic chips
Many postings indicate that Chinese labs still rely on Nvidia hardware and software for inference workloads. For example, a ByteDance role titled "Inference GPU Performance Optimization Expert" specifies responsibility for the company’s LLM inference framework and explicitly references CUDA and TensorRT-LLM: "Primarily through GPU and CUDA performance optimization techniques... build an industry-leading, high-performance LLM inference engine." That points to Nvidia-based inference in practice.
At the same time, other roles show interest in domestic accelerators. ByteDance Seed advertised an "AI heterogeneous computing optimization expert" role that prioritises knowledge of Ascend and Cambricon optimisations. Z.ai’s posting notes that in early 2026 the team developed and open-sourced GLM-Image, "trained entirely on domestic chips."
From these signals the authors infer that domestic chips are likely used more for inference and for training smaller or post-training models, whereas large-scale pretraining of the biggest models probably still relies on other hardware. They note GLM-Image has 16 billion parameters, which is likely 10–100× smaller than the largest models referenced by these organisations.
Startups rent cloud compute and are increasingly building data centers
Compute sourcing matters for AI progress. The postings show Chinese startups both renting cloud capacity from domestic cloud providers (e.g., Alibaba, ByteDance) and increasingly investing in self-built data centers. Moonshot and other startups list roles for procuring cloud compute resources, while MiniMax seeks engineers to "participate in building company-level self-built data centers and the SRE & DevOps system." DeepSeek advertised a data center role in Inner Mongolia, and Moonshot has positions explicitly coordinating hybrid deployments combining public cloud and self-built intelligent computing centers.
The overall pattern suggests a hybrid approach: renting cloud capacity for scalability while also building owned infrastructure.
Diverse commercial strategies: B2B focus vs consumer products
The job ads show variation in commercial focus among startups. Z.ai’s go-to-market roles skew towards B2B sales, whereas MiniMax and Moonshot list more marketing and consumer-oriented positions. Financial context mentioned in the analysis: about 70% of MiniMax’s 2024–2025 revenue came from AI-native consumer products (e.g., the Talkie companion AI app and Hailuo video generation), with 30% from enterprise sales. By contrast, 73.7% of Z.ai’s 2025 revenue came from running models on customers’ infrastructure—an intensive B2B approach.
Z.ai’s postings highlight government institutions and state-owned enterprises in energy and finance as target clients and include roles aimed at US market expansion and Fortune Global 500 customers. MiniMax, by contrast, displayed a stronger international hiring footprint (16 active postings in San Francisco and others across global cities), and 73% of MiniMax’s 2025 revenue came from international markets compared to 9.8% for Z.ai.
Startups remain model-centric; platform companies take broader research bets
Job descriptions indicate that startups such as DeepSeek and Moonshot focus narrowly on improving LLMs or building products atop them—examples include roles like "Agent Harness R&D Engineer." Z.ai is similarly LLM-focused but also posted two robotics roles within its X-Lab research unit, which explores new model architectures and research frontiers.
By contrast, larger platform companies ByteDance and Alibaba advertise roles in robotics and wearable hardware. Alibaba’s Qwen team has a role centered on automotive applications and the technical deployment of a "Qwen cockpit voice-assistant AI Agent." The authors attribute this to established platform firms having stronger supply-chain access and more latitude for riskier, hardware-oriented research bets, while startups concentrate resources on software and model development.
Geographic distribution: several major hubs rather than a single cluster
The postings are concentrated in a few Chinese tech hubs—Beijing, Hangzhou, and Shanghai. According to the dataset, 93% of postings with stated locations were in at least one of those three cities, and Beijing appeared in 63% of postings. This is less geographically concentrated than the US AI ecosystem, where roughly 85% of frontier AI job listings are in San Francisco.
Possible explanations include provincial competition (local subsidies attracting companies to multiple cities) and the geographic distribution of top university talent across Shanghai, Zhejiang, and Beijing.
Lower prior-experience requirements compared with US labs
A striking difference is the amount of prior work experience requested. The US labs in a comparable dataset required on average 5.5 years of prior experience, whereas Chinese lab postings averaged 1.6 years. Even when restricting to postings that explicitly state a number, the gap persists (5.5 years for US vs 3.4 years for Chinese labs).
Part of this is institutional: Chinese authorities encourage campus recruitment as a primary channel for graduate employment, and the Ministry of Education runs campaigns to promote this. Companies mirror that approach—DeepSeek’s CEO Liang Wenfeng has been quoted as saying the firm "hires on ability, not experience," and ByteDance runs a "Top Seed" recruitment programme aimed at students and recent graduates. As a result, campus postings account for nearly 20% of the engineering roles in the Chinese dataset.
Conclusion: a heterogeneous, context-dependent ecosystem
The 1,604 job postings suggest Chinese AI firms are not following a single playbook. Some, like Z.ai, prioritise B2B sales and government or state-owned clients; others, like MiniMax, emphasise consumer products and international expansion. Big platforms such as ByteDance and Alibaba make broader research bets that include robotics and hardware, while startups focus on models and software.
The ads point to ongoing Nvidia reliance for inference, a growing interest in domestic accelerators and owned data centers, and a hiring ecosystem that leans heavily on university recruitment—resulting in lower experience thresholds compared with US counterparts. The authors note a caveat: absence of public job postings in an area does not prove a firm is not already active there internally.
Acknowledgements: the original post thanked Isabel Juniewicz, Lynette Bye, Stefania Guerra, and Campbell Hutcheson for feedback and support.
