When a company raises $1B before shipping a commercial unit, the market is buying position, not revenue. Physical Intelligence (PI) is closing a round above $11B — more than doubling from $5.6B in November 2025 and nearly five times the $2.4B it carried in 2024. A 4.6× appreciation in 16 months, no product in market.

What Happened

Founded by ex-DeepMind researchers, PI builds the model layer, not the metal. Their π0 (pi-zero) is a generalist robot foundation model — the OpenAI bet on language, applied to physical manipulation. Investors: Founders Fund, Lightspeed, Thrive Capital, Lux Capital, Jeff Bezos, Alphabet, NVIDIA. Valuation: $2.4B (2024) → $5.6B (Nov 2025) → >$11B (Mar 2026). The $1B raise, if closed, is the largest single check in humanoid-adjacent foundation model development.

Why It Matters for Operators

PI at $11B with no disclosed deployment timeline means operators cannot yet evaluate the model — the bet is entirely on the VC side. What the raise does signal: the foundation model approach (pre-train on human motion data, fine-tune per environment) is attracting enough institutional conviction that hardware integrators need to start asking their robot vendors which foundation model they intend to run. The question is no longer whether foundation models will power robots — it is which model, and who controls the fine-tuning data.

Why It Matters for Investors

Figure AI reached $39B in its September 2025 Series C. PI hits $11B without a shipping product. The market is pricing model layer and hardware integrator as separate assets. If foundation models commoditize robot hardware the way cloud commoditized servers, the highest-margin position is the model — not the machine.

What Could Go Wrong

PI carries three structural risks that $11B does not erase: pre-commercial valuation fragility (no disclosed customers, no deployment timeline — the round is priced on thesis, not evidence); dependency on hardware partners, since PI does not build the robot and integrators have no obligation to adopt π0; and VC consensus risk — when Founders Fund, Lightspeed, Bezos, and NVIDIA all agree, the counterargument has no institutional voice.

Amazon + Fauna Robotics — Acquisition — Consumer humanoid robot — Amazon ships floor to door.
Sprout ($50K, 3'6", 50 lbs); second M&A this month alongside Rivr.

Kraken Robotics + Covelya — $615M — Subsea autonomy platform — Naval and commercial, one stack.
~$365M 2025 revenue, 24% EBITDA; closes Q2 2026.

Renault + Wandercraft — 350-unit deployment — Factory humanoid — Europe's largest humanoid fleet.
Calvin-40 already stacking tires at Douai; 50% fewer interruptions targeted by 2027.

Soma Energy — $7M seed (Category Ventures, Haystack, RRE, Uncork) — Data center power AI — Robot fleets need megawatts first.
2 GW optimized, 5 data center customers live; ex-Amazon energy team.

FANUC America — $90M Michigan facility. 840,000 sq ft in Pontiac, 225 jobs, late 2027; $300M cumulative U.S. investment. FANUC Academy opens in Auburn Hills — the U.S.'s largest robotics skills-development center.

NVIDIA GR00T N2 + IGX Thor — generally available. GR00T N2 tops RoboArena, completing tasks 2× more often than leading VLA models. IGX Thor (ASIL-D certified) ships to J&J, Karl Storz, and Medtronic. RoboCore (arXiv:2603.01517) wires a modified ray-tracing accelerator into GPUs — 14.8× faster than CUDA baseline, full pipeline under 10ms — closing the real-time collision detection gap that IGX Thor's ASIL-D certification addresses at the hardware level.

Figure 03 — autonomy debate unresolved. Zeighami's March 30 post (256,000 views) claimed teleoperation; Adcock's rebuttal: "Autonomous. No human was in the loop for this." No independent verification published. Figure targets 50,000 units/year at BotQ. Second time this question has surfaced.

ABB + NVIDIA RobotStudio HyperReality — 99% simulation accuracy. H2 2026 release to 60,000 RobotStudio customers — directly relevant to Capital Layer Renault/Wandercraft fleet scaling toward 350 units.

Policy radar. FCC spectrum reform (comments due May 1) + FAA eIPP across 26 states (summer 2026) = most significant airspace deregulation since Part 107. NHTSA FMVSS window closes April 10 — exempting purpose-built AVs from human-control requirements opens the path for Cybercab-class robotaxis.

"RoboGPU: Accelerating GPU Collision Detection for Robotics"
arXiv:2603.01517 — Liu et al., University of British Columbia, March 2, 2026

Liu et al. propose RoboGPU — wiring RoboCore, a modified ray-tracing accelerator, directly into GPUs to achieve 14.8× faster collision detection than CUDA baseline, 3.1× over existing accelerators, with full motion-planning pipelines completing under 10ms while remaining reprogrammable as the field evolves. For operators deploying in unstructured environments, sub-10ms collision detection is the boundary between reactive and safe — a deployment bottleneck that IGX Thor's ASIL-D hardware certification and RoboCore's software layer now address in tandem.

If you are evaluating foundation model vendors for your robot fleet this quarter, demand three things before signing: a disclosed deployment timeline (not a roadmap), a named hardware partner who has run the model in production, and benchmark data from an environment that resembles your own — not a lab. The Physical Intelligence raise tells you the VC market has already decided. Your procurement decision cannot run on the same evidence set.

Weekly Question

PI at $11B, no product. Figure AI at $39B, autonomy claims under scrutiny. At what point does the gap between foundation model valuations and verifiable deployment outcomes become systemic — and who closes it first?

Tell us what you think. Reply on X @PhyAIweekly47 or comment on our LinkedIn post.