Author｜Bao Yonggang

Editor｜Liu Wei

With 50 million yuan on the table, Dr. Zhang finally decided not to start a business.

He had just left an acquired company and caught up with the new wave of chip entrepreneurship. A well-known VC is willing to spend 50 million yuan to let him form a team and become CEO. Dr. Zhang frequently meets with investors and identifies core members, and the project has reached the start-up stage.

After a while, he stopped.

After talking with investors and people in the market, Dr. Zhang felt that his abilities did not fully match what he really needed to start a business.

Placing it in the spring of 2026, this is an abnormal decision. At that time, the valuation of popular chip projects could increase tenfold in three months, and technical backbones of major manufacturers were constantly being persuaded to start their own businesses. Most people were worried about missing the window, but Dr. Zhang chose to stay on the sidelines when the money was already in front of him.

Wang Xiaofeng is already on the court.

Wang Xiaofeng, who has a technical background, has worked for several domestic AI chip companies. What attracted him to starting a business was that he could finally do something according to his own judgment.

What he is doing is one of the hottest directions in 2026, LPU chips. During a financing exchange, an investor asked him: What programming model should LPU use?

For GPUs, the answer is SIMT. There is no ready answer for LPU chips.

Wang Xiaofeng knows very well that this is not a problem that can be solved with a few pages of PPT. If he could calm down and study for two or three months, he felt that he could at least make a preliminary plan.

The problem is, he doesn't have these two or three months. "There is no time at all, and I can't calm down to study it slowly. Everyone in the team is under great pressure."

It’s not because the company has no money. After receiving the first investment in early 2026, it took more than half a year to advance three or four rounds of financing, and its valuation increased from the initial few hundred million yuan to nearly one billion yuan.

Dr. Zhang chose not to start a business when faced with 50 million yuan, but Wang Xiaofeng was pushed forward by rounds of financing.

In the chip entrepreneurship boom of 2026, two sets of clocks with different speeds begin to pull against each other. Capital is clocked in weeks and months, while chip research and development is still clocked in years.

There are more counter-intuitive phenomena in this craze. The less short of money a chip company has, the easier it is to raise money. In the second half of the year, capital began to cool down, but the wages of chip talents were still doubling.

Why is capital willing to push the valuation of a chip company tenfold in a few months? Facing the same craze, why do some people rush to start a business, while others choose to give up after getting money? Add the author on WeChat BENSONEIT to discuss how long can the chip boom last?

The last round of GPU listings has begun to materialize, and the next round of inference chip entrepreneurship is booming

Sarah from Tuochi Headhunting has been working in the chip industry for nearly ten years. Around October 2025, she began to come into contact with more and more new chip companies and investment institutions. By the end of the year, the number of customers in her hands increased significantly.

At that time, many companies that would emerge in 2026 had not yet surfaced, but Sarah had already felt that the market had changed.

Similar changes have also occurred upstream in the industry chain.

Li Wei, who is in the chip IP business, found that starting from the second half of 2025, some projects that had been talked about for a long time but were shelved due to lack of money were restarted. It is not only new companies that will get money, but also some old companies that have been established for four or five years and have survived the trough in 2023 and 2024.

Why did capital start investing in chips again?

An important reason is that the last round of AI chip investment finally allowed investors to see exits again.

The listing and high market value performance of GPU companies such as Moore Thread and Muxi have simultaneously stimulated investors who have invested in domestic GPUs and those who missed this round of opportunities. The former saw considerable paper returns and was willing to continue betting on chips, while the latter began to look for the next batch of companies that might go public.

But this investment boom does not belong to all chip companies.

Senior chip investor Xue Haofei sees that what will really become hot in 2026 is AI, computing power, and the chips and infrastructure surrounding them. The popularity of many non-AI semiconductor assets is even weaker than in the previous two years.

While capital sees exit channels, AI reasoning also really brings new chip entrepreneurial opportunities.

During training, a general-purpose GPU can undertake most core tasks. In the era of inference, latency, throughput, storage bandwidth, power consumption and token cost have become more important. Agents, long context, and complex reasoning continue to amplify these differences, and the industry is beginning to form a consensus that all AI tasks do not necessarily have to be handed over to the same GPU.

Even NVIDIA, the biggest winner of GPUs, has begun to introduce dedicated architectures such as LPU into inference systems.

Chen Junjie, a second-time entrepreneur, summarized the last round of AI chip entrepreneurship as "domestic substitution", making a chip similar to NVIDIA GPU, and then replacing NVIDIA through a compatible ecosystem.

By this round, he judged that the capital story of domestic substitution had entered the second half, and new opportunities began to turn to AI-specific chips and heterogeneous division of labor.

The two forces encountered each other from the end of 2025 to the beginning of 2026.

On the one hand, the listing of companies such as Moore Thread and Muxi has allowed capital to once again see the exit channel for chip investment.

On the other hand, new problems brought by AI reasoning have given new imagination space to new architectures such as LPU, 3D, integrated storage and computing, and near-memory computing.

One opens the capital window, and the other opens the industrial window.

After the Spring Festival, both windows opened at the same time.

The valuation increased 10 times in 3 months. The less money is needed, the easier it is to raise money.

After the Spring Festival in 2026, the primary chip market has become increasingly popular.

Xue Haofei observed that the period after the Spring Festival is the period when this round of project financing and valuation rise fastest. The valuation of some popular projects can double in one or two months. A quantum chip company's angel round valuation after the Spring Festival was about 200 million yuan. By the time the financing news was announced in May, the valuation had reached about 2 billion yuan, a 10-fold increase in two or three months.

This kind of speed of financing and valuation increases is not common in the 2021 wave of AI big chip entrepreneurship.

The speed at which valuations rise reflects the popularity of the market, but along with the rise in valuations comes a phenomenon that seems counterintuitive. The less money-hungry chip companies are, the easier it is for them to raise money.

The reason is related to the long cycle of chip development.

A large chip requires long-term and continuous investment from design, tape-out, tape-out, software adaptation and customer verification. Xue Haofei estimates that some large chip projects require nearly 2 billion yuan in cash reserves in advance.

If a company's valuation is too low, it will be difficult to raise enough money at once. Instead, investors will worry about whether the company has the ability to last until the day the product is actually produced.

A cycle that seems to go against common sense appears.

Teams with good backgrounds and strong capabilities are more likely to obtain high valuations, and high valuations allow companies to raise more cash. The more abundant the cash reserves are, the lower the probability that the company will die in the middle of research and development. Therefore, investors who enter later will feel that the risk is smaller and they will be more willing to give money.

Rocket is responsible for financing related work at an LPU company. Although the company already has sufficient cash reserves on its books, which is enough to last for a long time based on the current team size, he still wants to continue to raise funds.

"The less money I am short of, the more money others will give me. The more money I am short of, others will take advantage of my illness and kill me."

This is Rocket’s years of experience in financing.

Continuing to raise funds is not just to increase the valuation. Experienced entrepreneurs know that the capital window is often much shorter than the chip development cycle, and they should get money as quickly as possible when they can.

For a chip company, the real danger is not that there is little money on the books when the capital market is the hottest, but that when it needs to continue financing two years later, it happens to encounter a market where no investors are willing to sell.

But money doesn't flow equally to all chip companies.

In the first half of 2026, the chip companies that investors are most willing to bet on are concentrated in a few directions.

LPU, 3D, storage and calculation integrated, taking over the domestic replacement

The chip entrepreneurship boom in 2021 has also seen rising valuations, competition for talent, and entrepreneurial ventures by executives of major manufacturers. But in the two waves of boom, the companies investors are willing to bet on are no longer the same.

Five years ago, investors had an overseas reference answer. GPU, CPU, EDA, analog chips, semiconductor equipment and materials, overseas have proven the demand and what the products should roughly look like.

Chinese startups are more answering the question, who can make it?

This is like a fill-in-the-blank question under the logic of domestic substitution.

In 2026, the question is clear, but the answer is not.

Inference requires lower latency, higher bandwidth, and lower cost. Advanced processes and high-end storage are still limited, and models are constantly changing. What kind of chip is needed?

Continuing to upgrade the GPU is one answer, and LPU is also an answer. Integrated storage and computing, near-memory computing, dedicated ASIC chips, combined with 3D stacking technology, may all be the answer.

These new structures, which have no standard answers, have also become the direction in which capital is most willing to bet in 2026.

LPU is one of the most typical ones. Companies such as Yuanchuan Micro, Mingxin Qirui, Xinsen Future, Hanxu Technology, and Fangqing Technology have attracted a lot of attention from LPU.

But in fact, LPU has not yet become a clearly defined chip category like GPU, and the LPUs that different startups want to do are not even exactly the same. Startups are also willing to use the LPU label and are happy to see more peers join.

"When the industry has not yet formed a solid understanding, competitors continue to emerge, which is a good thing to some extent. Letting the market and investors form a consensus together is a good thing for the entire track." Rocket's attitude is very clear, "First, let investors believe that LPU is a new track worth betting on, and then it will be everyone's turn to compare whose team is stronger and whose product is faster."

3D chips and integrated storage and computing chips also have a similar aspect.

AI inference is increasingly constrained by memory bandwidth, and domestically it faces limitations such as advanced manufacturing processes, high-end packaging and HBM. Putting storage and computing units closer together in 3D and using higher bandwidth to solve the bottleneck in the Decode stage is a technical route with real industry needs.

But there is still a long distance between "the technical logic is established" and "the company is worth tens of billions of dollars."

At this time, startups will actively look for scenarios that big companies really care about. If a PC, server or host manufacturer is preparing to make a new product, the startup company will think about whether its own 3D chip can be embedded in it. Once recognized by major manufacturers, this solution will gain both customer endorsement and capital narrative.

"Sometimes startups, investment institutions and ideal customers will 'make up stories' together." A 3D chip entrepreneur described it this way.

"Story" here is not a derogatory term. All early-stage technology investments are essentially pricing a future that has not yet happened.

The real question is how far the story goes and how far the engineering verification follows.

No matter how many stories about LPU, 3D, integrated storage and computing, and high bandwidth are told, in the end someone has to make the chip and prove in real models and customer scenarios that it can indeed reduce costs and increase token output.

In the entrepreneurship and financing boom of 2026, many proofs have not yet happened, and capital has already set a high price for these new structures.

After money enters these companies, it quickly flows to the next scarce resource, chip talents.

NPU engineer with two years of experience, annual salary of one million

After capital enters a chip company, it will soon turn into a recruitment budget.

In the summer of 2026, there are already signs of cooling down on the investment side, and the HR and headhunters of chip companies are getting busier and busier. Investment institutions invest money from March to June, and companies only start to expand their teams after receiving the money. There is a natural time lag of several months between the capital market and the talent market.

Salaries have become as intuitive a price signal as valuations in this boom.

A senior chip headhunter from Qianheng Consulting recently encountered a case that impressed him. An engineer who only had a master's degree for two years and worked as an NPU Core for two years received an annual salary of more than one million yuan.

NPU R&D engineers may be one of the most sought-after talents in this boom.

Sarah said the reason is straightforward. Traditional AI chip companies want NPU talents, emerging LPU companies want them, and embodied smart chip companies are also vying for them. Different architectural stories ended up in the same NPU talent pool.

There are still a few cases of annual salary of one million yuan. The more common market situation is that two years of NPU experience can earn about 700,000 yuan, and three years of experience can earn 900,000 yuan, which is significantly higher than the previous trough.

NPU isn't the only one being robbed, either.

As super nodes become a hot spot in computing infrastructure, related talents are also scarce. Sarah is helping a related team recruit people, and the other party is interested in a senior engineer with a salary of over one million.

Sarah suggested that if you really wanted to grab people, you might have to double your salary. The other party responded simply: "We can negotiate for doubling the salary."

If you only look at these numbers, it is easy to think of the battle for chip companies to grab people around 2021, but the two waves of craze are not exactly the same.

Sarah remembers that the last round of salary increases was more like a general increase. Some candidates without fully matching experience might get a 50% or even double increase.

Chip companies in 2026 will still be willing to charge high prices, but they will obviously be more selective.

An engineer with five or six years of experience may only earn an annual salary of four to five million yuan if his ability and direction are not outstanding. On the other hand, even if you only have two or three years of experience, as long as you happen to be doing the NPU or chip interconnection that is most lacking in the market, you can still earn an annual salary of one million.

The last round of capital was more about paying a premium for “big factory experience” and resume. In this round, more and more companies are beginning to pay a premium for specific capabilities that can solve immediate problems.

Sarah judged that this round of high premiums for NPU and Internet talents will be difficult to maintain in the long term.

In addition to changes in talent prices, Sarah also noticed another deeper change. Chip entrepreneurs are getting younger.

In the past, a chip entrepreneur born in 1988 was already considered young. In 2026, she began to frequently meet doctors born in 1995 and 1996 as CEOs or co-CEOs of start-up companies.

These young Ph.D.s may not have more than ten years of complete chip mass production experience, but they can judge the technical direction by themselves and recruit people with five, ten or even more years of engineering experience to supplement their design, verification and engineering capabilities.

A capital boom has changed more than just how many companies are founded. It is also re-pricing capabilities in the chip industry and changing who can sit in the No. 1 position in advance.

Money is on the table, but some people still choose not to start a business

In mid-2026, Jie Shu, who had been in the chip industry for more than ten years, found that more people were trying to persuade him to start a business, and someone came to his door almost every few weeks.

“In March, four groups of people came to me, and they even came to negotiate with angel and Series A money.” He said that in May, when the market was at its hottest, someone even told him directly: “You don’t have to worry about money. I can help you bring in all the remaining money in 2026.”

His answer is: "You cannot start a business when a certain field is the hottest. Instead, you should make bets more than two years in advance in fields that no one is paying attention to."

In Jieshu's view, if we start focusing on this field before 2024 and survive the two years that have not yet been noticed, by 2026 when the trend comes and a consensus is formed in the field, we will already be the leader in the field. This is a more reliable path, but it requires entrepreneurs to have industry insight and predictability.

But if we see that the market is hot by the end of 2025, we set up a company in March 2026, and expect to raise several billion yuan in a few months, this path will be difficult to copy.

For an emerging field, the entrepreneurial window and the financing window are not the same window.

This is also where Jieshu is most wary of this wave of craze. “Sometimes the development of the story depends not only on the person telling the story, but on what kind of story the person who believes it wants.”

When a technical direction suddenly becomes valuable, it is not necessarily because the entrepreneurs are better at telling stories, or it may just be that capital needs a target that can carry a new narrative at this point in time.

Jieshu chose not to quit and continue to settle down with an excellent team. There are still some people who are not faced with "should they start a business", but whether what they already have is worth giving up.

Dong Yong is a middle-level manager of a listed domestic GPU company. In 2026, several groups of people also approached him, hoping that he would join a startup company to create joint ventures, but he rejected them all.

"It will take a few years for the company's options to be realized, and I will definitely not leave my job until they are realized."

In addition to options, his health is also a concern for him. Starting a business means bearing high-intensity work and uncertainty again. For a technical manager who has been working in the industry for many years, this account is completely different from that of young people.

The same capital window, for a doctor in his twenties, may mean becoming a CEO ten years earlier.

For a technical leader who has worked in a large factory for more than ten or twenty years, it means giving up stocks, income, organizational resources and occupational security, and taking another gamble on a financing window that may only last half a year.

50 million was placed on the table. Dr. Zhang tried to walk in, but then backed out.

Jieshu did not end up when the field was booming and capital was at its most enthusiastic, but kept a close eye on the development direction of the industry.

Dong Yong chose to stay in an already listed company.

The chip entrepreneurship boom in 2026 is hot enough, but it does not allow everyone to come to the same answer. Are you afraid of missing out or choose to miss out? Add the author on WeChat BENSONEIT to learn more about startup companies.

Capital is starting to cool down, but chip talent prices are still rising.

Han Peng, who has been working in chips for more than 20 years and experienced his second entrepreneurial experience, originally judged that this round of capital enthusiasm could last until at least mid-2027.

But by the second half of 2026, he had clearly felt that the market was cooling down.

对于一个做了二十多年芯片的人来说，几个月原本并不算长。但在2026年，几个月已经足够跨过一整个资本周期。

上半年，创业公司担心的是会不会错过估值快速上涨的窗口。到了下半年，担心的开始变成，两年以后还会不会有人愿意给钱。

摩尔线程、沐曦等公司的上市和高市值表现，曾经让一级市场重新看到芯片投资的退出机会，也迅速放大了投资人的风险偏好。到了下半年，二级市场转弱，一级市场也开始重新计算风险。

春天的融资是进攻，到了秋天，融资开始变成储备弹药。

但资本开始降温的同时，芯片人才还在涨价。

Sarah明显感受到，6月之后一级市场已经不如上半年，一些机构甚至已经在前半年消耗掉了相当一部分全年投资额度。但她手里的招聘需求还在继续增加。

春天投资人把钱打进创业公司，公司拿到钱后组建团队、启动招聘，到了夏天和秋天，这些钱才开始集中流向人才市场。

2026年下半年出现了一幅看起来矛盾的景象，芯片投资已经开始降温，芯片人才的价格还在上涨。

Sarah判断，到2027年三四月以后，才更容易看清这一轮招聘热潮真正的拐点。

资本先热，也先冷。

人才后热，也后冷。

而芯片研发，比资本和人才都更慢。

上市可以是资本的终点，但不是芯片的终点

资本热潮给了LPU、3D、存算一体这些新架构一次难得的机会，但资本从来不是为了技术理想而来。它要的是回报和退出。

上一批国产GPU公司上市之后，投资人才重新愿意把钱投进芯片。对创业者来说，上市或者被并购也意味着另一件很现实的事情，个人财富可能因此发生巨大变化。

一位资深一级市场投资人接触过的一些创业者很直接，从公司成立时就已经在考虑未来如何上市和退出，“他们也想为了套现”。

这很容易理解，工程师放弃大厂工资、股票和确定性去创业，希望承担更高风险换来更高回报，是商业世界的一部分。

真正值得警惕的是，当投资人和创业者的回报都越来越依赖下一轮融资和最终上市，公司的时间表也会越来越接近资本的时间表。

产品还没有验证，下一轮融资已经开始。

客户还没有规模使用，估值可以先涨几倍。

产业投资人黄致远观察这一轮LPU项目时看到，一些公司估值从约1亿元迅速涨到二三十亿元。如果每一轮估值都在上涨，产品、业务和客户却没有同步往前走，最后只会变成“纯涨故事”。

融资可以跳跃，产业不会随融资节奏跳跃发展。

上市可以让投资机构退出，也可以让创业者实现财富跃升，但芯片最终还是要进入客户的真实业务。

LPU要证明Token成本真的更低，3D芯片要证明高带宽的优势在实际场景发挥出来，客户也不会因为一家公司刚完成高估值融资，就把核心业务交给一颗没有经过验证的芯片。

王晓峰和团队已经越来越知道资本市场应该怎么玩。

但投资人最早问他的那个问题，LPU芯片到底应该采用什么样的编程模型，并没有因为公司估值上涨而消失。

在2026年的资本市场里，两三个月足够一家热门芯片公司完成一轮融资。

在芯片研发里，两三个月可能只是解决一个工程问题的开始。

等上市、并购和财富兑现以后，那个最朴素的问题还是会留下来：这颗芯片，到底有没有人用？

注，张博士、王晓峰、李伟、陈俊杰、Rocket、揭曙、董勇、韩鹏、黄致远均为化名 雷峰网(公众号：雷峰网)雷峰网

雷峰网原创文章，未经授权禁止转载。详情见转载须知。

编辑

重发邮箱修改邮箱