Robotics

Early-stage physical AI rounds are priced like growth deals

Large checks are going to robotics and physical AI at the earliest stages (Axios Pro Rata, signal 26). A $90M seed led by B Capital and Kindred Ventures was reported by three outlets, and a $150M Series A from a16z and F-Prime values its company at $1.45B. Seed rounds in the 90-day window total $6.66B across just 15 deals, more than any other stage, and Series A adds $2.37B across 20 deals. Top investors by deal count are a16z (6, plus 4 under the a16z name), Eclipse, Sequoia and SoftBank, which shows that frontier labs and brand-name funds are driving the pace.

Weekly capital swings from $120M (week of 9/21) to $6.87B (week of 9/28), but deal counts are rising from 1-4 per week in August to 13-15 per week in October. The last four full weeks held 10, 14, 15 and 13 deals. Late-stage rounds still appear, including a $400M Series E from Boyu Capital and a $200M Series D with Tiger Global, Baillie Gifford and BlackRock, so breadth is growing alongside the headline-grabbing outliers.

Why it matters · Investors should read the deal-count trend rather than single-week dollar totals, which a handful of outliers distort.

Mecka (backed by Sequoia) licenses egocentric human data collected with caps and tactile gloves, and says it holds hundreds of petabytes. Axonnic (YC F26) launched muscle-sensing sleeves, and Ropedia released W0. Mecka's CEO argues robotics lacks the internet cold start LLMs had, and the industry is converging on a data pyramid of internet video, egocentric, UMI and teleoperation data.

Why it matters · Data supply is turning into a separately investable layer, and its value depends on whether model labs license data or collect it themselves.

Generalist announced GEN-1 and GEN-1.5 with one-shot learning from a single demonstration, and the firm frames dexterity as a horizontal capability like coding. Interpretability work on VLAs, including π0.5 and MolmoAct2, finds that the vision-language backbone conditions action generation through a narrow middle-to-late layer region. Ropedia W0 reportedly beats NVIDIA's Cosmos3 despite Cosmos3's up-to-64B parameters.

Why it matters · Architecture is still contested, so no single model recipe has locked in an advantage.

Lightspeed backed Warp, which builds AI-driven robotic labs for drug-discovery chemistry. On the consumer side, Matic is cited as an early deployable cleaning bot, and Gecko Robotics is held up as a strong vertical company. Meanwhile, iRobot has failed, and Chinese vacuum-robot makers Ecovacs, Roborock and Dreame are fighting over a mature market.

Why it matters · Vertical deployments with clear ROI look better placed than consumer hardware commoditization.