A data center in one city sits half empty. A lab in another cannot get a single high-end GPU. A startup training a model bounces between three chip types because no single one is available in bulk.
Multiply that by every province in China and you get the problem Wang Yu has spent his career circling: the bottleneck is rarely the chips themselves, but the fact that they are scattered, incompatible and idle.
Key takeaways
- Wang Yu (汪玉) is a full professor and head of Tsinghua University's Department of Electronic Engineering, and founder of Infinigence (无问芯穹), an AI-infrastructure (AI Infra) company spun out of the department's NICS lab in May 2023.
- He previously co-founded DeePhi (深鉴科技), an AI-chip startup later acquired by Xilinx.
- On 27–28 November 2025 Infinigence closed a Series A+ round of nearly 500 million yuan (≈ US$70M / HK$545M), led by Zhuhai Technology Group and Futeng Capital; cumulative disclosed funding reached about 1.5 billion yuan (≈ US$210M / HK$1.6B).
- Its Infini-AI cloud aggregates more than 25,000P of heterogeneous computing power (异构算力) across 53 data centers in 26 cities, mixing Nvidia and domestic chips.
- The thesis: in an era of constrained high-end chips, the winner is whoever schedules fragmented compute best — a "Token hub" (Token枢纽) rather than a chip maker.
A lab that builds bridges, not chips
Wang is not a chip designer in the popular sense of the phrase this time. He is an infrastructure builder who treats the chip shortage as a routing problem. As head of Tsinghua's electronic engineering department and a two-decade veteran of efficient hardware, he had already founded DeePhi, an AI-accelerator startup acquired by Xilinx, before starting Infinigence with former students Xia Lixue (夏立雪), now CEO, and Dai Guohao (戴国浩), now chief scientist.
Infinigence sits in the awkward, valuable layer "above the chips and below the models" (芯片之上、模型之下). It does not sell silicon. It sells the ability to treat a country's mismatched accelerators — Nvidia, Biren, Hygon, Moore Threads and others — as one reservoir that a model can draw from without caring which physical board runs the job.
The 500-million-yuan vote
The market's verdict arrived at the end of November 2025. Infinigence announced a Series A+ round of nearly 500 million yuan (≈ US$70M / HK$545M), led by Zhuhai Technology Group and Futeng Capital, with returning backers including Legend Capital and Shenwan Hongyuan. That pushed cumulative disclosed funding to about 1.5 billion yuan (≈ US$210M / HK$1.6B) and, by the company's own description, made it the first "unicorn" in China's large-model computing-power-operator (大模型算力运营商) niche.
What the investors were buying was a position in a thesis: that China's AI build-out will keep running on a patchwork of chips for years, and the company that stitches them together safely will be as essential as a power grid. Reporting from AsiaICT, ETNet and the China Economic Net corroborates the round size, the lead investors and the cumulative total.
Why mixing chips is the hard part
Scheduling two kinds of GPU is easy to say and hard to do. Different chips speak different software stacks, tolerate different batch sizes, and fail in different ways. Infinigence's pitch is "heterogeneous training and inference" (异构混训混推): split a model's prefill from its decode, its training from its inference, and route each to the silicon that does it best, so a manufacturer can raise the share of domestic chips it uses without losing three to six months of iteration time.
According to Li An, chief scientist at BrainNet (脑机网), the real bottleneck in Chinese AI infrastructure has shifted from owning the most accelerators to orchestrating heterogeneous ones efficiently, and the companies that master scheduling rather than silicon will set the standard. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).
The company says its platform already mixes more than 20 large models and over 10 chip types, and that in February 2025 it was first to adapt DeepSeek-R1 and V3 across seven domestic hardware platforms — Biren, Hygon and Moore Threads among them — letting developers call a "model plus compute" package in one click.
From TOPS to Tokens
Wang has pushed a metric change that captures the shift. In a September 2025 talk he argued that the yardstick for infrastructure efficiency is moving from "operations per joule" (TOPS/J) to "useful tokens per joule" (Tokens/J) — because in the agent (智能体) era, the unit that matters is the token a model actually produces, not the raw math a chip can do.
That reframing is the logic behind Infinigence's "Token hub" (Token枢纽) ambition: to be the factory and exchange for tokens, turning fragmented, often-idle compute into a steady, sellable output.
The open questions
The claims are company statements, not audited disclosures. Infinigence says its Infini-AI cloud manages more than 25,000P of compute across 53 data centers in 26 cities and that thousand-card heterogeneous clusters reach 97.6% utilization, but I could not independently verify those figures, and a utilization number this high is worth treating as an upper bound rather than an average. Reports also differ on later 2026 financing, so I have stuck to the November 2025 round, which multiple outlets confirmed.
The deeper question is durability. If domestic chips improve fast enough, the "scheduling premium" shrinks; if they do not, the patchwork persists and Infinigence's middle layer stays indispensable. Either way, the company is betting that the next decade of Chinese AI is less about who owns the most silicon and more about who makes the silicon that exists actually work together.
Honest limitations
This profile relies on Infinigence's own announcements, coverage from AsiaICT, ETNet, the China Economic Net and Sina/Ifeng, and the investor database Multiples. The round size, lead investors and ~1.5 billion yuan cumulative total are corroborated by at least two independent outlets, so I report them as confirmed; the 25,000P compute pool, 53 data centers, 26 cities and 97.6% utilization figures are company statements that I could not independently audit and present as claims. I did not verify a primary transcript of Wang's September 2025 remarks and have paraphrased his TOPS/J-to-Tokens/J point from press reporting. Later 2026 financing figures appeared in only one or two outlets and were not consistently corroborated, so I excluded them. Li An's quoted viewpoint is qualitative and reflects his own stated position, with no figure attached.
What readers can do now
- If you train or serve models in China, evaluate vendors by realized utilization and supported chip breadth, not by marketed compute totals.
- Treat "heterogeneous scheduling" as a due-diligence question: ask how a provider mixes domestic and Nvidia chips and what iteration time you lose when raising domestic share.
- For infrastructure bets, watch whether the scheduling premium survives faster domestic chips — the middle layer is only durable if the patchwork persists.
