A venture firm put RMB 50 million on the table for Dr Zhang to build a chip company and be its chief executive. He recruited, met investors, got the project to launch stage, and then stopped. After talking to enough people he concluded that his abilities and the job did not match.
In the spring of 2026 that was an odd decision. Hot chip projects were seeing valuations multiply tenfold in three months, and engineers at large companies were being talked out of their jobs at a steady rate. Zhang walked away with the money already in front of him.
Two clocks running at different speeds
Wang Xiaofeng is inside the window. A technical founder who has worked at several domestic AI chip companies, he picked one of the hottest directions of the year: an LPU. In a funding conversation an investor asked him which programming model an LPU should use. For a GPU the answer is SIMT. For an LPU there is no ready answer.
Wang reckons two or three quiet months would produce a preliminary design. He does not have two or three months. After a first round in early 2026 he raised three or four times in half a year, and the valuation moved from a few hundred million renminbi to close to RMB 1 billion.
That is the structure of the boom in one comparison. Capital counts in weeks and months. Chip development still counts in years. The two clocks are pulling against each other, and the companies in the middle are being asked to answer questions that only time can answer.
Tenfold in a quarter, and money chasing the funded
Heat rose sharply after the Spring Festival. Investor Xue Haofei saw projects double in a month or two. A quantum chip company that raised an angel round at a valuation of about RMB 200 million after the holiday was worth about RMB 2 billion when it announced in May.
The counterintuitive part is that the companies needing money least raise it most easily. A large chip requires design, tape-out, silicon bring-up, software adaptation and customer qualification, and Xue estimates some projects need close to RMB 2 billion of cash in reserve. A company with a low valuation cannot raise enough in one go, and investors worry it will not survive to the product. High valuations attract capital, capital reduces the chance of dying mid-development, and lower risk attracts more capital.
One founder working on fundraising at an LPU company describes the logic without decoration: the less he needs money, the more people offer it, and the more he needs it, the more they smell blood. The real danger is not a thin balance sheet in a hot market. It is needing to raise in two years and finding nobody willing to write.
What the money is actually buying
The two forces that met at the end of 2025 were an exit window and a technical one. Listings by GPU makers Moore Threads and Muxi gave investors visible returns, which pulled money back into the sector. At the same time inference changed the requirements: latency, throughput, memory bandwidth, power and cost per token now matter more than raw training throughput, and agents with long context amplify every one of those differences.
Nvidia itself is introducing LPU-style architectures into inference systems, which legitimises the category without defining it. Domestic startups use the LPU label loosely, and several are building quite different things. One founder argues that a crowded field helps, because the first job is persuading investors that the category deserves money at all.
3D stacking and compute-in-memory follow the same pattern. Memory bandwidth is the binding constraint on inference and access to advanced process, advanced packaging and HBM is restricted, so putting memory physically closer to compute has a real industrial rationale. Getting from a plausible architecture to a valuation of tens of billions of renminbi is a different distance. As one 3D chip founder put it, startups, investors and the imagined customer sometimes write the story together. Early-stage technology investment is always pricing a future that has not happened. The question is how far the story has run ahead of the engineering validation.
The price of a two-year engineer
Money that reaches a chip company becomes a recruiting budget within weeks, which is why hiring stayed hot while investment cooled in the summer. One engineer with a master’s degree and two years on NPU cores was offered more than RMB 1 million a year. The common range for two years of NPU experience is around RMB 700,000, and three years reaches RMB 900,000.
The bidding is more selective than in 2021. Back then, pay rose broadly and candidates without matching experience could still get 50 per cent. Now a capable engineer with five or six years but no standout direction may earn RMB 400,000 to 500,000, while two or three years in exactly the right specialism clears a million. Recruiters expect the premium on NPU and interconnect talent to fade, because premiums tied to a specific shortage usually do.
The founders are getting younger as well. Where a founder born in 1988 once seemed young, recruiters now meet chief executives born in 1995 and 1996, doctors who cannot claim a decade of volume production but can pick a technical direction and hire the engineers who can build it.
The ones who said no
Not everyone took the money. Jie Shu, who has spent more than a decade in the industry, had four approaches in March alone, some carrying angel and Series A capital, and was told in May that funding was not his problem. His answer was that you should not start a company at the peak of a sector’s fashion, but two years before anyone is looking. For a new field, the startup window and the funding window are not the same window.
Dong Yong, a mid-level manager at a listed domestic GPU company, turned down co-founder offers because his options need several more years to vest and because a technical manager in his forties is pricing a different set of risks than a twenty-six year old doctor. The same capital window that gives one person a decade of head start asks another to give up stock, salary and organisational resources to bet on a funding window that may last six months.
Dr Zhang tried the door and stepped back out. What the boom has actually repriced is not the technology. It is how much a specific shortage is worth per year, and which people are allowed to sit in the first chair before they have shipped anything.
Editor’s note: This is an adapted translation of the original Leiphone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/chipdesign/6mjibhXfWFoLrabg.html.