PI Raises $600M at $5.6B Valuation — The Highest for a “General-Purpose Robot Brain”
Physical Intelligence (PI), a San Francisco-based startup founded less than two years ago, has closed a $600 million Series A round, catapulting its valuation from $2 billion to $5.6 billion.
The round was led by CapitalG (Alphabet’s growth fund), with continued participation from Lux Capital, Thrive Capital, Jeff Bezos, and new investors Index Ventures and T. Rowe Price.
Notably, OpenAI is an early backer, having participated in PI’s $70 million seed round (March 2024) and $400 million angel round (November 2024).
PI is not building robots.
It is building the first general-purpose AI “brain” for physical agents — a software platform that can control diverse robot bodies, from humanoids to quadrupeds to industrial arms, using a single, unified model.
This positions PI not as a robotics company, but as an embodied AI infrastructure provider — valued more like an LLM developer than a hardware OEM.
Founding Team: DeepMind, Berkeley, and Stripe — the Tech Mafia
PI’s co-founding team reflects its ambition:
- Karol Hausman (CEO): Former Google DeepMind researcher; led teams in robotic perception and real-world reinforcement learning
- Sergey Levine (Co-founder): Professor at UC Berkeley; pioneer in end-to-end robot learning and offline RL
- Groom (Co-founder): Former Stripe executive; led growth and capital strategy
The broader team includes engineers from Google, Tesla, and Stanford, creating a rare blend of academic depth and product execution.
Their thesis is clear:
The bottleneck in robotics is not hardware — it’s intelligence that generalizes across tasks, environments, and bodies.
Technical Breakthrough: π*0.6 — The First Brain to Generalize Across Real-World Tasks
On November 18, PI released π*0.6, its latest foundation model for embodied AI.
Unlike demo-only systems, π*0.6 has demonstrated sustained, multi-hour performance in unstructured environments:
- Espresso making — 8+ hours of consistent operation, adjusting for bean grind, water pressure, and cup placement
- Clothing folding — handling cotton, silk, denim with >90% success
- Industrial box assembly — precise insertion of components into cardboard frames
These are not scripted motions.
They are language-instructed, vision-guided, force-aware tasks executed in real homes and factories.
The key innovation is RECAP — a training method that prioritizes high-quality, real-world interaction data over simulation scale.
“We believe general intelligence must be learned in the real world,” said Hausman.
“Simulation can’t replicate the friction, compliance, and chaos of physical reality.”
Two Paths to Embodied AI: PI vs. Skild AI
Skild AI, backed by SoftBank, Bezos, Sequoia, and CMU, claims to have simulated “1,000 years” of robot experience.
The global race for the “robot brain” has split into two philosophies:
| Dimension | PI | Skild AI |
|---|---|---|
| Training Data | Real-world interactions (high quality, low volume) | Massive simulation + internet video (high volume) |
| Core Belief | “World diversity is overrated — focus on representative tasks” | “Scale laws apply — more data = more robustness” |
| Model Strength | Open-world generalization (e.g., clean unknown rooms) | Extreme robustness (e.g., walk with broken joints) |
| Commercial Strategy | Pure platform — license brain to all hardware makers | Vertical-first — target industrial inspection and security |
| Valuation | $5.6B | $4.5B (as of June 2025) |
But PI argues:
You can’t learn to pour coffee by watching videos or crashing in simulation. You have to feel the cup.
China’s Answer: QJ Robots — The “PI of China”
In China, QJ Robots is pursuing a near-identical strategy.
- Origin: Spun out of Tsinghua University’s Brain-Inspired Intelligence Lab
- Focus: General-purpose embodied AI model, no robot body
- Architecture: “Brain-region” design — mimicking human functional specialization
- Commercialization: Licensing its “Brain Dock” edge compute module to hardware makers
Like PI, QJ targets existing platforms first — adding autonomy to robot vacuums, quadruped dogs, and delivery bots — to “remove the remote control.”
Its brain already runs on 7 robot form factors, with partnerships across consumer electronics and logistics.
“We’re not waiting for the perfect humanoid,” said CEO Gao Haichuan.
“We’re making today’s robots truly autonomous — one task at a time.”
Investors like Inno Angel Fund see QJ as a “self-evolving, milestone-hitting company” — a rare blend of academic rigor and commercial pragmatism.
The Core Debate: Must “Brain” and “Body” Be Tied Together?
Soft-Hardware Integrated Camp (Tesla, Figure AI, Unitree):
- Argue that actuator dynamics, kinematics, and sensor placement are inseparable from control
- Data collected on one body cannot transfer to another
- Vertical integration is the only path to reliability
Pure-Software Camp (PI, QJ, Skild):
- Believe that higher-level cognition (task planning, object understanding, intent) can be generalized
- Use modular architectures: a “brain” for reasoning, a “cerebellum” for motion
- Aim to become the Android of robotics — a platform that runs on any qualified hardware
Emerging Consensus:
The “dual-brain architecture” is becoming standard:
- Large brain: Handles language, planning, world modeling (can be general)
- Small brain: Controls balance, gait, reflexes (tightly coupled to hardware)
But pioneers like Peking University’s Wang Yuan are pushing further — toward a unified “embodied neural intelligence” that merges perception, cognition, and control in one model.
Investment Takeaway: The Real War Is Over the Stack, Not the Shell
The humanoid robot narrative has matured.
- 2023–2024: Battle was about who can walk
- 2025: Battle is about who can work
- 2026–2027: Battle will be about who owns the intelligence layer
PI’s $5.6B valuation reflects investor belief that:
The winner won’t be the company with the best legs.
It will be the one with the best mind — and the ability to license it to everyone.
For global investors, the landscape is now clear:
| Play | Example | Risk | Reward |
|---|---|---|---|
| Hardware OEM | Tesla, Unitree, UBTECH | High capex, supply chain risk | Near-term revenue |
| Vertical AI | Skild AI | Narrow TAM, commoditization | Fast commercialization |
| General Platform | PI, QJ Robots | Long runway, no revenue yet | Monopoly potential |
The NDRC’s recent warning about “overheating” in China’s robot sector signals that hardware duplication is overvalued.
The real premium now belongs to scalable intelligence.
And in that race, PI — and its Chinese counterpart QJ — are the two horses to watch.
📌 Appendix: Official Funding Tables
Physical Intelligence — Funding History
| Funding Round | Date | Amount | Investors |
|---|---|---|---|
| Series A | 2025-11-21 | $600 million | Led by CapitalG. Index Ventures, Lux Capital, Thrive Capital, T. Rowe Price, Jeff Bezos |
| Seed Round | 2024-11-04 | $400 million | Led by Lux Capital and Jeff Bezos and Redpoint. Bond, OpenAI |
| Pre-Seed | 2024-03-12 | $70 million | Led by Thrive Capital. Sequoia Capital, Khosla Ventures, Lux Capital, Greenoaks, Outset Capital, OpenAI |
| A1 Round | 2022-04-01 | RMB 30 million | CCB Beijing |
Skild AI — Funding History
| Funding Round | Date | Amount | Investors |
|---|---|---|---|
| Series B | 2025-06-14 | $135 million | Led by NVIDIA and Samsung and SoftBank Group |
| Series A | 2024-07-11 | $300 million | Led by Bezos Expeditions and Coatue Management and Lightspeed Venture Partners and SoftBank Group. Sequoia Capital, SVAngel, General Catalyst, Menlo Ventures, CRV, Felicis, Amazon, Carnegie Mellon University, Amazon Alexa Fund, Amazon Industrial Innovation Fund |
