Unitree loses half its value as robot hype meets its GPT-2-era reality
A reality check rippled through the hottest corner of AI investing this week, and it came from the public markets rather than a lab. China's flagship robot maker saw its valuation roughly halve days after a euphoric listing — at the same moment the builders of robot "brains" were privately admitting their field is still stuck in a pre-ChatGPT phase.
Unitree lost roughly half its market value in a week, the sharpest signal yet that investors are starting to ask what all those humanoid robots can actually do. The Chinese robot maker, valued around $66 billion after roaring onto a mainland listing that spiked through a 460% debut, has since given nearly all of it back — one tracker pegs the slide at 45% as technology investors flee, and trade press reports the stock halved in a week amid figures showing only about 9% of robot sales reach real end customers. Analysts framing the rout as a warning about both China's tech-bubble risk and shoddy IPO mechanics blame a now familiar gap: robot bodies keep improving, but the intelligence to put them to value-creating work still lags badly.
The timing is brutal because the builders themselves agree with the market. At last week's Actuate conference — which tripled to 1,500 attendees since 2023 — creators of physical-AI models described their field as being in its "GPT-2 era," the pre-ChatGPT moment when more data and compute were the only obvious way forward. A booth for infrastructure startup Avala promised to solve "the robotics data crisis," and the honest responses were striking: Wayve CEO Alex Kendall likened manipulation robotics to "self-driving five years ago," while Genesis AI's Théophile Gervet noted that "no customer cares about the general purpose robot that works at 80% success rate." The deployments that work today are narrow — Gritt building solar farms, Agility in industrial settings, Bedrock running excavators — while the general-purpose humanoids everyone is funding stay in the lab.
Why it matters: this is the first real stress test of the largest funding cycle in embodied-AI history. Billions have flowed into robot startups on the promise that physical AI is on the verge of its ChatGPT moment; the public market is now demanding proof rather than the promise. The productive read of the rout isn't that the category is a bubble — it's that the general-purpose bet is premature while the vertical specialists quietly generating revenue and real-world data are the ones with staying power. Sharing that data and tooling (itself a spillover from autonomous driving, via Wayve, Uber, and Foxglove founders) is what nudges the field down the road to a useful robot. We covered the mainland listing rush in Unitree prices $9B humanoid IPO, first in mainland China — worth revisiting as the froth deflates.
AWS is acquiring DuckLabs, the company behind the wildly popular DuckDB analytical database — and pledging to keep it, and its "Duck Stack" of projects, open source. The bootstrapped Amsterdam team of founders Mark Raasveldt and Hannes Mühleisen, together just over 30 people, will join Amazon Web Services effective early September, with the promise to continue their years-long work on DuckDB, DuckLake, and Quack. The core projects — which see more than a million downloads a day — remain free under the MIT license, the nonprofit DuckDB Foundation keeps stewardship, and the team stays put in Amsterdam; AWS says it's committed to the project long term, while the foundation gains a technical advisory board and the extension stack opens up.
Why it matters: hyperscaler takeovers of beloved open-source projects ordinarily trigger an open-source anxiety alarm; this one leans the other way. AWS is explicitly buying the team's reach and infrastructure rather than the code, and it framed the deal as a way to bring DuckDB's embeddable, in-process analytics to "more enterprises" — which is exactly the property that has made DuckDB the quiet engine inside a huge share of modern data and AI products, from RAG pipelines to agent-facing analytics. Endorsements from MotherDuck and Fivetran signal the ecosystem reads it as momentum, not a power grab. For an AI-focused reader, it's a reminder that the analytical layer under the agent stack is consolidating fast, and AWS is betting its scale can grow the DuckDB "revolution by another couple of orders of magnitude."
What to watch: whether Unsloth, vLLM, and the broader open-weight serving stack keep pace as frontier labs ship earlier architecture previews — and whether DuckDB's foundation structure becomes the template other beloved open-source projects point to when the acquisition calls come in.
Does an AWS-owned, foundation-stewarded DuckDB feel safer to build on than a VC-backed one — and have the robot rout changed your mind about investing in physical AI?
