Robotics venture has already eaten 2025 by lunchtime

Global robotics startups raised $18.8bn in the first roughly five and a half months of 2026, according to a Crunchbase sector snapshot published this morning. That is more than the $15.0bn the sector raised in all of 2025, and more than the previous annual peak of $14.1bn set in 2021. We are not yet at the summer solstice and the year has already broken the record.

The interesting question is not whether this is a lot of money. It is. The interesting question is what kind of money it is, and what the structure of the flow tells us about where the marginal AI venture dollar now goes.

What the snapshot actually says. Crunchbase reports $18.8bn of disclosed rounds into robotics startups year-to-date, with Figure AI (humanoid robots) and Skild AI (an embodied-AI foundation model, meaning a general-purpose model trained to control physical systems) named as anchor deals. Chinese embodied-AI companies are described as co-leading by deal count. The piece frames robotics as "the next concentration of AI venture dollars." F-Prime's parallel read, cited in industry summaries this week, puts the YTD figure closer to $16bn and projects roughly $20bn for the full year — a more conservative count that nonetheless agrees on the direction.

The handoff is now visible in dollars. For most of the last three years the "physical AI is next" story lived in keynotes and slide decks. The capital didn't move; it stayed in frontier model rounds, where OpenAI, Anthropic and xAI absorbed almost everything the LLM lane could carry. What changed in H1 2026 is that the rounds cleared. Figure, Skild, and a handful of Chinese peers raised at sizes that used to belong to language-model labs. If the 2024–25 wave was capital betting on tokens, the 2026 wave is capital betting on actuators.

The shape of the flow is familiar. A small number of mega-rounds are doing most of the dollar work. Crunchbase doesn't publish a median, but the named deals are large enough that the arithmetic only closes if the distribution is heavily top-weighted. This is the same structural shape as the LLM lane: a handful of capitalised-to-the-teeth labs at the top, a long tail of smaller companies that collectively raise a fraction of what the leaders do in a single round. The frame that applies is winner-take-most concentration under capability uncertainty — investors who don't know which architecture wins write very large cheques to the two or three names they think might.

Whether physical AI actually has LLM-style winner-take-most dynamics is a separate question, and I am not sure it does. Robots are hardware. Hardware has supply chains, regulatory surfaces, and per-unit margins that don't compress to zero the way inference does. A humanoid platform that wins logistics may have nothing to say about surgical robotics. The capital is behaving as if one or two foundation-model-for-robots companies will eat the category; the underlying economics suggest a wider distribution of value is at least plausible.

The defence slice is larger than the framing admits. New Market Pitch's June summary flags defence-adjacent robotics as a "quietly large share" of the run-rate, and this matches what is visible round-by-round: ground autonomy, ISR (intelligence, surveillance and reconnaissance) platforms, and logistics autonomy are taking material capital under the robotics banner. This matters for two reasons. First, defence-adjacent rounds have customer-concentration and exit profiles that look nothing like consumer or enterprise robotics — one buyer, long procurement cycles, programme-of-record rather than product-market fit. Second, it means a non-trivial portion of the $18.8bn is effectively dual-use defence capex moving through a venture wrapper, which has implications for how durable the run-rate is once primes start writing their own cheques.

China co-leading by deal count is the geopolitical tell. The Crunchbase framing, Chinese firms co-leading by count rather than dollar volume, is important to read carefully. It almost certainly means the US still leads on aggregate dollars while losing the count, because Chinese median round sizes in embodied AI run smaller than US humanoid mega-rounds. But the count is the leading indicator. Deal count tracks ecosystem breadth: number of teams capitalised, number of architectures being tried, number of shots on goal. Two parallel embodied-AI ecosystems capitalising heavily in the same six-month window, with export controls limiting compute and talent flow between them, is the picture an intelligence-explosion-signals frame would predict. It is also the picture you would expect to produce divergent technology stacks within 24 months.

Is this signal or performativity? The fair contrarian read is that $18.8bn into pre-revenue or early-revenue robotics companies looks a lot like 2021's autonomous-vehicle wave, which compressed hard when unit economics arrived. The capital is real; whether the underlying businesses ship at margin is a different question. Figure and Skild are not, today, revenue stories. The structural case for the round sizes is that embodied foundation models, once they work, command rents similar to LLM foundation models, and the time-to-deployment is short enough that a 2026 cheque sees revenue by 2028. If either half of that breaks — the models don't generalise across embodiments, or the deployment cycle is the usual hardware seven years rather than the hoped-for two — the run-rate compresses.

What to watch. Three things. Whether any Q3 round prints at a valuation above the Figure or Skild marks, which would signal the concentration is still tightening rather than peaking. Whether US primes (Anduril, Shield AI, Palantir-adjacent) start absorbing the defence-adjacent slice through acquisition, which would pull a meaningful chunk of the venture run-rate out of the count. And whether any of the Chinese embodied-AI leaders disclose deployment numbers, not orders, not LOIs, deployment, which is the only data point that distinguishes this wave from the previous two robotics surges that promised the same thing.

The $18.8bn is real. What it buys is still being negotiated.

Footnotes

¶