Zhang Yaqin, head of Tsinghua's Institute for AI Industry Research, draws a layered map of the 2026 cycle: AI infrastructure is in a generational build out, but the startup layer above it is overcapitalized.
Five years into running Tsinghua's flagship AI commercialization lab, Zhang Yaqin is drawing a sharp line between two layers of the current cycle. The technology, he argues, is in a generational build-out. The companies raising on top of it are in a bubble. "Valuation figures like 150 billion yuan don't really mean much," he said in a 36kr interview timed to the third annual Taihu Dialogue in Wuxi. "In the end, companies have to commercialize the technology, generate revenue and profit, and build real competitiveness."
The line matters because Zhang is not an outside skeptic. He is one of China's most credentialed AI insiders: the country's youngest IEEE Fellow, a Chinese Academy of Engineering academician, the former head of Microsoft and Baidu's China operations, and the founding dean of Tsinghua's Institute for AI Industry Research (AIR), a Beijing-based lab that pairs academic AI research with company incubation. He is also, by his own count, the steward of roughly 10 portfolio companies that have together raised about 15 billion yuan and now carry an aggregate valuation of about 150 billion yuan.
The framework Zhang laid out in the 36kr conversation splits the 2026 AI cycle into two distinct bets. The infrastructure layer, meaning power, compute, and algorithms, is in a 1998-to-1999 internet-style build-out, where the question is not whether the underlying technology will mature but how fast capacity can come online. The company layer above it, where most of the current capital is flowing, is a different story. "AI has no bubble, but early-stage AI companies do have bubbles," he said. The trillion-dollar AI leaders of the next decade, in his view, are unlikely to be today's marquee startups.
The forecast Zhang offers is concrete enough to be falsified. He expects the domestic large-model market to consolidate to three or four winners. The robotics and embodied-AI sector, where more than 300 startups are chasing capital, should compress to roughly 20 viable players within three to four years. AIR's own portfolio leans into that second theme. Two of the institute's better-known spinouts, Huashen Zhiyao (华深智药), an AI drug-discovery firm, and Tashi Zhihang (它石智航), an embodied-intelligence company, are both described as unicorns with strong capital-market reception. A separate Sina Finance report describes Earendil Labs, an overseas AI drug-discovery venture that the 36kr interview and Sina coverage pair with the Huashen Zhiyao founder team, closing multiple 2026 funding rounds totaling USD 787 million, a figure the report calls a 2026 global biotech funding record. Named investors include Dimension Capital, DST Global, INCE Capital, Luminous Ventures, Miracle Capital, and Sanofi. Neither the 36kr interview nor the Sina report states the corporate relationship explicitly, so the linkage should be read as an industry pairing rather than a confirmed fact.
Zhang's most specific warning is reserved for the professor-founder archetype that supplies many of these startups. In his own taxonomy, AI academics can spin out companies in three ways: go all in as a full-time founder, with Qualcomm's two MIT-affiliated founders and AIR professor Peng Jian (who left to run Huashen Zhiyao full time) as his reference cases for the first pattern; keep the academic role but pair with a professional operating team; or stay on as a managing professor overseeing multiple ventures. Only the first two, he argues, work. The third almost never does, because a CEO who is still teaching and grading cannot give the company the full operating attention it needs through the difficult early years.
What Zhang says separates a real company from a frothy one, as described in the 36kr interview, is the discipline of commercial fundamentals: knowing the problem being solved, the AI solution being offered, the customer paying for it, and the path from there to revenue and profit. Reflexive over-raising in a hot market, he warns, tends to burn capital without converting it into operating performance. That gap is what he expects the next three to four years to expose.
The institute's own research direction tracks the same bet on physical AI over the model layer. An arXiv preprint from AIR on a "Real-Sim-Real" loop for transferring robotic policies from simulation to physical hardware sits in the same embodied-intelligence lane as the institute's better-known spinouts, alongside faculty work on simulation-to-real robotics.
What to watch over the next twelve months: whether Chinese AI funding visibly concentrates around the three or four large-model survivors Zhang expects, and whether the robotics and embodied-AI cohort thins from several hundred to roughly twenty. The third annual Taihu Dialogue convenes AIR's portfolio and outside investors in Wuxi against that backdrop, and Zhang's published bottom line, that revenue and profit, not aggregate valuation, decide which startups survive, gives readers a clean test for which of the current marquee names pass it.