AI Infra Pay in China Surges to 8 Million Yuan as DeepSeek R1 Reshapes Demand

Liu Li, an executive at a large-model unicorn, told Leiphone that AI Infra engineering packages at big tech companies used to stay below 3 million yuan a year. Now, good Infra people can be offered 7 million to 8 million yuan. That remains below algorithm roles, which pay 5 million to 14 million yuan, but Liu called it a turnaround. Lu Bingning, a senior AI expert at listed human-resources firm Career International, said he met an Infra candidate in Shenzhen in 2024 whose annual salary was 300,000 yuan. It rose to 500,000 yuan in 2025, and after a job change this year, it reached 1 million yuan. Lu said moves from unicorns to big internet companies can double pay, while moves from big companies to unicorn startups can raise packages about threefold, though options make up a large share. He said threefold is an exception; twofold is more common.

The trigger, according to Lu, was the release and popularity of DeepSeek R1 in early 2025. The model's success showed the value of Infra improvements, and AI Infra talent was quickly fought over by model makers. That competition has continued. AI Infra engineers work on extracting higher utilization from costly GPU cards and saving money. From 2022 to 2024, Infra was a supporting role. After ChatGPT appeared, Chinese companies entered a race to match its model performance and find a killer app. Attention and high pay went first to algorithm and product roles. Engineering efficiency had a lower priority, and problems of compute waste and scheduling efficiency were temporarily hidden. Zhang Tong, a compute scheduling expert at a large-model lab, said that as long as the cluster did not crash and the bill did not explode, the team was considered done. Support teams like Infra engineering were equated with operations, he said, and no one took them seriously.

The reversal came after DeepSeek R1. Lu said the industry asked how DeepSeek, with so few people, could build such a good product and make so much money. Part of the answer was Infra cost reduction. Large-model inference burns real money and often runs up huge bills. Cost reduction through Infra turns every yuan saved into profit. Lu estimated that good Infra cost reduction can save at least tens of millions of yuan in compute expenses. The key areas are operators, inference, and scheduling. PD separation, expert parallelism, VLLM, and SGLang are among the optimizations that cut costs most sharply.

Compute is often the largest expense. A budget ratio circulating in the industry says that for every 100 yuan an AI model company invests in AI training, 40 yuan goes to compute, 30 yuan to talent, 20 yuan to advertising, and 10 yuan to data. GPU procurement, rental, and electricity often take the largest share of AI business costs, sometimes more than 60%. Yet cards often sit idle. Zhang said GPU utilization in the industry has long been low, averaging only 25% to 27%, meaning more than half of purchased compute is not fully used. Good Infra teams can raise utilization to 40%, and that gap of more than ten percentage points is a barrier. An eight-card B-series machine has a domestic spot price of 13 million yuan and a three-month futures price of about 11 million yuan, so saving eight cards saves more than 10 million yuan.

Zhou Yue, a senior AI Infra engineer at a large-model unicorn, said that squeezing a little optimization out of Infra brings an efficiency leap far more cost-effective than desperately grinding on algorithms. Large-model training has become almost an open-book exam. With open-source models spreading and architectures such as Transformer maturing, competition has shifted from basic pre-training to engineering implementation and inference optimization. Zhou said there is little creative work that truly requires one person and no one else; most work is data cleaning, data ratio experiments, and engineering tuning. He gave the example of a 'level-seven old fitter': an Infra engineer opens a server, straightens the fiber and cables, and loosens the cable ties a bit so they are not too tight, and can make a server 10% to 20% faster. Lu said the Infra role is no longer the old Infra. In 2023 and 2024, companies looked at benchmarks, and a few points of improvement could help them raise money. Now algorithms and products are more mature, and companies want the maximum value from each card. Infra returns are easier to quantify, so its value has risen.

Not everyone benefits. Guo Quanwei, co-founder and CTO of StartLux, said the phrase 'salary doubling' mistakes a few outliers for the norm. Maimai salary data shows that AI Infra talent in mainland China earns about 73,000 yuan a month, slightly above large-model algorithm workers. But the average hides internal division. The people who pull the salary curve up are those who can raise card utilization. Many doing ordinary Infra are far from doubling. Infra roles that deploy inference frameworks such as VLLM and SGLang into products mostly pay 40,000 to 50,000 yuan a month, not double. Those who are doubling are concentrated in inference optimization and model quantization. Guo said if an engineer helps a company save 30% of training time or increase token throughput by 30%, costs drop and R&D output rises. He said people who can do solid operator optimization, elegant quantization of inference engines, and real throughput gains deserve doubled pay. But such people are not numerous. Most have five to ten years of experience in Infra, having worked on distributed systems, high-performance computing, search and recommendation, or databases at large internet companies. Guo estimated that those who truly capture the high premium are only 5% to 10% of the AI Infra population. Lu divided Infra into inference optimization and acceleration, training frameworks, and operator development. All three have high increases, but they receive different levels of attention. He compared operator development to sharpening kitchen tools, training frameworks to building a kitchen system, and inference to serving food fastest. The industry generally considers operators and inference more important. Training framework talent is in relatively ordinary demand. The work is expensive, and only top internet and cloud companies can afford dedicated teams. Some unicorns have cut entire training groups and kept only inference and operator teams.

Different employers have different calculations. Liu said ByteDance has a large business and needs people who can make training and inference scheduling stable. Whether for Seedance or Doubao, if inference engineers can speed things up 10%, the annual compute savings are considerable, and the return on investment is very high. Tencent has taken another path. Its base models were weaker in earlier years, and in the past two years it has clearly been catching up on base models and full Infra, bringing in Yao Shunyu and setting up a separate AI Infra department. Alibaba and Baidu, as cloud vendors, built up large numbers of distributed and foundation Infra engineers during the cloud computing and big data boom, and many can transfer to AI roles. Both have self-developed chips and focus more on heterogeneous computing, so hiring is steady rather than explosive. Demand has surged most at ByteDance, Tencent, and unicorns such as DeepSeek, Kimi, and MiniMax that want to extract maximum value from every card. Big tech companies usually field regular armies: an Infra department often starts at 50 or 60 people, with operators, training, inference, and networking divided as finely as operating rooms. Unicorns often use a lean special-forces model. Their Infra teams usually have about 20 people and prefer all-rounders who understand both operator fundamentals and inference engines. The structure is flat, with one all-round technical leader, two or three top experts in subfields, and about 20 senior fighters relying on individual capability to compete with big tech.

Domestic chip makers are another buyer. Liu said many have reached the IPO stage, and options are about to be cashed in, so some Infra talent will jump to them with the idea of taking a gamble for financial freedom. Cash increases at chip makers are generally around 45%, rarely matching the doubled offers at big tech, but the stock is attractive. In the underlying technology stack, Nvidia cards and domestic cards also show a clear premium gap. Lu said familiarity with Nvidia's CUDA ecosystem remains hard currency, with job-change increases of at least 80%. Domestic cards rely more on software-hardware co-design and complex ecosystem adaptation, with increases usually around 50%. Nvidia-side talent is limited by export restrictions, while policy is pushing companies to buy domestic chips, giving the domestic card line more room. North America is also hungry for AI Infra talent. OpenAI's recent openings are almost all for inference engines, low-level GPU operators, and performance acceleration, and Anthropic has more than a dozen corresponding positions.

Lu expects the salary increase to last at least two to three years. Supply remains short, companies are hiring, and salaries are high. Infra is also creating new business scenarios. Many Chinese companies have mostly laid out general chat base models, but complex vertical scenarios are just beginning. AI Coding consumes large amounts of compute, and high-definition video generation demands high throughput and low latency. Full commercial loops in many of these areas are not complete. As long as commercialization is stung by expensive compute bills, Infra talent who can squeeze profit from cards will continue to receive rising pay premiums.