Author | Xu Xiaofei Editor | Liu Wei
"In the past, the AI industry referred to those working in Infra and engineering as the 'civil engineering department in computer science,' doing the dirtiest and most exhausting work for the lowest pay. Who would have thought they are so valuable now?" Liu Li, an executive at a large model unicorn, sighed to Leiphone. He presented a set of salary comparisons: previously, the annual package for AI Infra engineering positions at major tech companies basically could not reach CNY 3 million; now, top Infra talents can command CNY 7 to 8 million. Although there is still a gap compared to algorithm positions (CNY 5 to 14 million), it is already considered a beautiful turnaround.
Lu Bingning, a senior expert in the AI business in South China at CareerInternational, a listed human resources service company, has also witnessed this surge: "In 2024, I met an Infra candidate in Shenzhen whose annual salary was still CNY 300,000; in 2025, it rose to CNY 500,000; this year, after changing jobs, it directly hit CNY 1 million." This is not an isolated case. "There are Infra talents who jumped from unicorns to major internet companies and doubled their salaries; those who jumped from major companies to unicorn startups saw their compensation packages increase by about 3 times, although options make up the bulk of it.
A 3-fold increase is an exception; a 2-fold increase is more common," he added. The trigger for this round of salary hikes was the release of DeepSeek R1. "At the beginning of 2025, R1 went viral, showing everyone the value of Infra improvements. AI Infra talents were immediately snapped up by various model vendors, and this situation has continued until now." Lu Bingning stated. However, doubled salaries have never been evenly distributed. When the market throws real money at this group, what exactly is pricing them? And which type of Infra professionals have eaten the fattest piece of meat in this round of dividends?
01. Why Does the "Civil Engineering Department" Command High Salaries?
What are AI Infra talents? Simply put, AI Infra talents focus on squeezing the highest possible utilization rate out of expensive GPU computing cards, genuinely saving money for the company. Lu Bingning told Leiphone that from 2022 to 2024, the Infra position was basically a "supporting role." After ChatGPT emerged, China immediately entered the "Hundred Models War" and the application trial stage. Capital and teams focused on two things: matching ChatGPT's model performance and running out a Killer App as soon as possible. Attention and high salaries were both prioritized for algorithm and product positions.
Under the anxiety of "getting on the bus before buying a ticket," engineering efficiency was not a high priority, and the underlying issues of computing power waste and scheduling efficiency were temporarily obscured. "As long as the cluster doesn't crash and the bills don't explode, the job is considered done. The value of support teams like Infra engineering was equated to operations and maintenance at the time, and no one took it seriously." Zhang Tong, a computing power scheduling expert at a large model laboratory, admitted frankly. The reversal happened after the emergence of DeepSeek R1.
According to Lu Bingning's recollection, after the release of R1, the industry was pondering one thing: DeepSeek has so few people, yet they made such a good product and earned so much money. How exactly did they do it? Later, part of the answer fell on Infra cost reduction. Everyone realized that large model inference burns real money, and often burns through the budget. At this time, doing Infra cost reduction means every penny saved is solid profit. So various companies started targeting DeepSeek to poach talents, and salaries were basically negotiated with open cards.
Doing Infra Cost Reduction Well Can Save Tens of Millions in Computing Power Expenditures
"The current role of Infra is to save money for vendors. Saving money is also making money," Lu Bingning said. How much can be saved? According to his estimation, doing Infra cost reduction well can save at least tens of millions in computing power expenditures. In terms of specific directions, operator optimization squeezes a single card to the extreme, while inference combines various resource scheduling to optimize costs and improve effective utilization. Among them, optimizations such as PD disaggregation, Expert Parallelism, vLLM, and SGLang achieve the most drastic cost reductions.
It should be known that the most expensive expenses for large model companies are almost entirely pressed on these cards. There is a budget allocation circulating in the industry: for every CNY 100 an AI model company invests in AI training, CNY 40 is spent on computing power, CNY 30 on talents, CNY 20 on advertising, and the remaining CNY 10 on buying data. Training a large model once or running inference for a night, the procurement, leasing, and electricity costs of GPU clusters often eat up the bulk of AI business costs, with some companies even accounting for over 60%. Yet money is spent, but the cards are often idling. Zhang Tong told Leiphone that the actual utilization rate of GPUs in the industry has been low for a long time, averaging only 25% to 27%, meaning more than half of the purchased computing power is not truly running.
And this is precisely where Infra talents come into play. "With the same batch of cards, a good Infra team can pull the utilization rate up to 40%. This gap of more than ten percentage points is the barrier." Let's do a simple calculation: a set of machines with 8 B cards, domestic spot delivery is CNY 13 million, and futures after 3 months are about CNY 11 million. "As long as 8 cards are saved, over ten million is saved." But saving money is only the most intuitive layer on the books. A deeper layer is that the technological competition of large models has moved from competing in algorithm innovation to the tough battle of competing in underlying system engineering.
The AI Industry Stage Has Changed, and the Value Weight of Infra Has Also Changed
At present, when the algorithm dividend has peaked, underlying system engineering has become a more cost-effective leverage. "Squeezing out a little optimization in Infra brings an efficiency leap that is much more cost-effective than grinding algorithms to death," Zhou Yue, a senior AI Infra engineer at a large model unicorn, admitted frankly.
It is becoming increasingly difficult to grind algorithms now because the technological side of large models has changed. Large model training is currently akin to an open-book exam. Thanks to the popularization of open-source models, foundational architectures like Transformer are maturing and converging. The focus of industry competition has shifted from basic pre-training to engineering implementation and inference optimization.
"In current large model training, there is very little truly 'indispensable' creative work. Most of the work is data cleaning, data ratio experiments, and engineering tuning. These tasks cannot be called high-tech, but they can bring in real money," Zhou Yue added. He also gave an example of a "Level 7 old fitter": an Infra engineer opens the server, tidies up the fiber optics and cables, loosens the cable ties a bit so they are not too tight, and can make a server 10% to 20% faster. "You really can't explain the logic behind this stuff; you have to actually do it to fully grasp it." "The Infra position is indeed no longer the Infra of the past."
Lu Bingning told Leiphone that in the past, Infra efficiency was not high because in 2023 and 2024, everyone was looking at evaluations. Improving by a few points was positive for the company and could also help raise money. But now that algorithms and products are basically mature, companies have shifted to demanding the maximum value each card can generate. At this time, the benefits brought by Infra are easily quantifiable, and the value naturally goes up. But value re-evaluation does not mean everyone gets a raise. When the market throws high salaries at Infra, only a small group of people actually eat the fattest premium.
02. Did Only 5% of People Eat This Round of Premium?
It seems that the four words "doubled salary" have become a catchphrase in the AI Infra recruitment market. But in the view of Guo Quanwei, co-founder and CTO of StartLux, this is more like mistaking a few outliers for a universal norm. His judgment has some data support. Job salaries on Maimai show that the monthly salary of AI Infra talents in mainland China is roughly CNY 73,000, which is already slightly higher than those doing large model algorithms.
However, the average conceals internal differentiation. What really pulls up the salary curve is the group of people who can increase the utilization rate of computing cards; while a large number of people doing "ordinary Infra" are far from doubling. For example, Infra positions that deploy inference frameworks like vLLM and SGLang to products mostly have monthly salaries stopping at CNY 40,000 to 50,000, which has nothing to do with "doubling."
More specifically, whose salaries are doubling? The answers given by several interviewees focus on two directions: inference optimization and model quantization. What these people do is precisely squeeze the computing power of every GPU to the limit. "If you help the company save 30% of training time, or get 30% more Token throughput, costs can be directly reduced by a margin, which means R&D output has increased," Guo Quanwei added.
In his view, it is not excessive for those who can solidify operator optimization, beautifully quantify inference engines, and truly run out throughput to double their salaries. But the problem is that this group of people is not as large as imagined. In Guo Quanwei's view, most of these people are veterans who have been deeply engaged in the Infra industry for 5 to 10 years.
In their early years, they worked on distributed systems, high-performance computing, search and recommendation, and databases at large internet companies. Their starting salaries were not low to begin with, and their experience is hard to replace. "Only with these 5 to 10 years can you handle today's AI Infra game; a fresh graduate or a newcomer who has only worked for a year or two asking for double the salary right off the bat is unscientific." According to Guo Quanwei's estimation, such people who truly eat the high premium basically account for only 5% to 10% of the entire AI Infra population. On the differentiation of salary increases, Lu Bingning also holds a similar view. He divides Infra into three blocks: inference optimization acceleration, training frameworks, and operator development. "The increases in these three blocks are all high, but the degree of attention varies." He gave a popular analogy: those doing operator development are like sharpening kitchen utensils, how to sharpen the knife the sharpest and how to heat the pan the most evenly; those doing training frameworks are building the back kitchen system, who cuts the vegetables, who wields the spatula, and who steps in when mistakes are made; those doing inference are thinking about how to serve the food the fastest. How to serve 10 dishes from one recipe at the same time and feed thousands of guests, which is the monetization link.
The industry generally considers operators and inference to be more important, and the popularity is divided accordingly. Leiphone learned that in the recruitment market, the popularity of training framework talents is relatively average. One of the reasons is that this work burns too much money; only top internet and cloud vendors can afford dedicated teams, while some unicorns simply cut the entire training group, leaving only inference and operator teams.
03. Who Is Paying Heavily: Five Types of Buyers, Five Accounts
Behind the differentiation of job popularity are the private calculations and anxieties of different employers. Looking at the entire industry, the "buyers" willing to pay the premium for Infra roughly fall into the following five categories, each with different needs. "ByteDance has a massive business volume, and they want people who can stabilize training and inference scheduling. Whether it is Seedance or Doubao, if inference engineers can accelerate it by 10%, the computing power expenditures saved in a year are quite considerable, and the return on investment is shockingly high," said Liu Li, the aforementioned executive at the large model unicorn.
Tencent is taking a different path: a few years ago, their foundation models were weak, and in the past two years, they have obviously been making up for foundation models and complete Infra, bringing in Yao Shunyu, establishing a separate AI Infra department, and relying on poaching mature teams from the outside to fill the gaps.
Cloud vendors like Alibaba and Baidu are on a different rhythm. During the previous bonus period of cloud computing and big data, they accumulated a large number of Infra engineers working on distributed systems and infrastructure. When the AI wave came, many could directly transfer to fill the gaps. Coupled with the fact that both have self-developed chips, the focus is more on heterogeneity, and recruitment has become a daily routine, steady and continuous, rarely showing the frenzy of exploding salaries. Those with a real surge in demand for Infra talents are ByteDance, Tencent, plus unicorns like DeepSeek, Kimi, and MiniMax, which want to squeeze the benefits of every card to the extreme.
Generally speaking, the deployment of giants emphasizes regular army corps operations. An Infra department usually starts with at least 50 to 60 people, with operators, training, inference, and networks divided as finely as a surgical operating room. Unicorns, on the other hand, often adopt a "lean special forces" model. Their Infra teams usually have about 20 people, favoring all-rounders who understand both the underlying operators and can operate inference engines. The architecture is also relatively flat, with an "all-around" technical leader taking the lead, bringing two or three top experts in subdivided directions, and leading about 20 senior fighters, relying on extremely high individual combat capabilities to fight head-to-head with major companies. Finally, there is a group of domestic chip manufacturers who are "betting on national fortune."
"Many domestic chip manufacturers have already reached the IPO stage, and options are expected to be cashed out. Some Infra talents will jump ship with the idea of 'taking a gamble and achieving financial freedom,'" Liu Li told Leiphone. Although the cash increase offered by chip factories is generally around 45%, and it is rare to see the doubled offers from major companies, the stock that has the potential to be cashed out is itself an extremely tempting chip. At the same time, looking at the underlying technology stack, the premium gap between the two routes of NVIDIA NV cards and domestic cards is equally obvious.
The aforementioned senior headhunter Lu Bingning told Leiphone that familiarity with the NVIDIA CUDA ecosystem remains "hard currency," and the salary increase for such talents changing jobs is guaranteed to be over 80%; domestic cards rely more on software-hardware collaboration and complex ecosystem adaptation, with increases usually around 50%, not as fierce as NV talents. However, limited by external sales restriction policies, there are not many talents on the NV side to begin with; and policies are pushing various companies to purchase domestic chips, so people on the domestic card line actually have more room for imagination in their prospects. It is worth mentioning that the North American market is also eager for AI Infra talents. The positions recently released by OpenAI are almost entirely occupied by inference engines, GPU underlying operators, and performance acceleration, and Anthropic also has a dozen corresponding positions open.
04. Conclusion
How much longer can this wave of salary increases for AI Infra talents last? This is also a concern for many people. Lu Bingning's judgment is that this wave of salary increases can last for at least two or three more years. "We are still in a state of supply falling short of demand. Basically, all vendors are recruiting, and salaries are posted very high." The reason it will continue to rise is that Infra has spawned many new business scenarios.
In the area of general dialogue foundation models, many domestic vendors have basically completed their layout; but the implementation of complex vertical scenarios has just begun to unfold. Whether it is the AI Coding track that ravenously consumes computing power, or the high-definition video generation track that has extremely high requirements for throughput and latency, the full-chain commercial closed loop has not been completed for many.
These new tracks are engaged in fierce battles, and the war among vendors to snatch talents is far from over. As long as the commercialization of large models continues to be stung by expensive computing power bills, those Infra talents who can squeeze profits out of computing cards will continue to receive ever-rising salary premiums.