Zhu Xiaohu: this may be the last year of foundation models
The managing partner of GSR Ventures says the industry has no bubble — but the model layer's story is running out of road.
Zhu Xiaohu, managing partner at GSR Ventures, says 2026 may be the last year the foundation-model layer gets to be the story at all. In an interview with Tencent News' Deep Web published this weekend, the investor who called humanoid robots a bubble at the start of the year repeated his long position: models end up as a commodity and a public utility, and the clock on paying frontier prices for them is nearly out. His evidence is historical rather than technical — IBM's Unix earned real money bound to hardware, then Linux took thirty years to replace it. Fully open models like DeepSeek, he argues, get to industry standard within a decade; within a year, open weights should handle "99.9 percent" of tasks without difficulty, at which point there is no reason to rent a closed frontier model. The API business carries a specific hazard he thinks is underrated: if someone ships an equally good model at 5 to 10 percent of your price, they take your customers in a single night. That, he says, is why Anthropic is rushing toward a listing while coding margins are still 60 to 70 percent.
The bubble call is narrower than the headline suggests. Zhu says there is no bubble at the industry level — compute centers sell out wherever they get built, and storage and memory absorb whatever capacity arrives, because video and agent workloads keep growing. The bubble is in individual companies: a trillion-dollar valuation "is definitely a bubble." His break signal is a revenue inflection — growth flattening or turning down — and he puts the market well short of it. What has already flipped is how public markets treat capex: announcing more of it now depresses a share price instead of lifting it, while software names have recovered to prior highs on improving margins. On the China-US split he is consistent with his past framing — the US does the expensive frontier science, China industrializes and compresses cost, catching up "in two or three months at 1 or 2 percent of the cost."
Where he puts money instead is the useful part of the interview. Office agents are the next market after coding, bigger and harder, and the moat is not the model — it is the office entry point. The fight is a continuation of WeCom versus Feishu versus DingTalk, and a startup without that layer "has almost no possibility," leaving vertical agents to work inside one of the three platforms. He names WorkBuddy as the current leader and sets the ecosystem test as whether partners can earn hundreds of millions, with more than a hundred partners above 100 million yuan. Robotics gets the reverse treatment: he counts close to 200 Chinese humanoid companies with almost no differentiation, revenue that is mostly government and research orders, and points at Unitree halving within two weeks of listing as the market agreeing with him — we covered that repricing when it happened. His alternative is industrial automation, where the same money buys a working commercial case: a hull-cleaning robot he backed is now being approached by customs to inspect ship bottoms for smuggled cargo.
Worth separating the two claims. The commoditization argument is a thesis about the next ten years and cannot be settled this year; the market-structure argument — capex punished, software rewarded — is already visible in the tape and is checkable every quarter.
What to watch: whether open weights close the last gap on hard tasks inside the next year, and whether office-agent revenue shows up in the software names Zhu says are already re-rating. If it does not, the "last year" line is just a good quote.
If open models match frontier models on 99.9 percent of tasks, what exactly are the labs selling? Tell us in the comments.
