Physical Intelligence
Pre-commercial research lab; intended model: license hardware-agnostic VLA models to robot OEMs and deployment partners (Android-of-robotics platform play), seeded via open-weight releases (openpi/pi0) and a partner program (Weave Robotics, Ultra)
First three points are hard priced-round marks; the Mar 2026 >$11B point is a reported in-talks valuation, not a confirmed close, as of Jul 2026. Total disclosed raised across closed rounds: ~$1.07B.
Earnings, margins, COGS & capex
Private and pre-commercial: Physical Intelligence has disclosed no revenue, margins, or burn. Its financial story is a capital-formation story - roughly $1.07B raised in under two years at a valuation that went ~$400M (seed, Mar 2024) -> $2.4B (Series A, Nov 2024) -> $5.6B (Series B, Nov 2025), with a further ~$1B at >$11B reported in talks in Mar 2026. Spending is research-lab shaped: frontier-scale training compute, multi-embodiment robot fleets, teleoperated data collection, and a small elite team (~80 people per Bloomberg, Mar 2026) drawn from Google DeepMind, Berkeley, and Stanford. Monetization is deferred by design - the bet is that a generalist robot brain, once it works, is licensed across every robot OEM.
Revenue trend
Margins
n/a - no commercial revenue base
Burn likely rising with compute scale-up and real-world deployment programs
COGS structure
Not disclosed. At maturity the cost structure would resemble a foundation-model company: inference compute per robot-hour, plus ongoing data-collection and fine-tuning costs per deployment. Today all spend is effectively R&D: NVIDIA GPU training runs, robot hardware (multi-embodiment arms and mobile manipulators), and human teleoperation labor for demonstration data - the last being a cost line software-only AI labs do not carry.
Capex
Not disclosed. Capital intensity is real but mostly runs through opex/cloud: large training clusters, physical robot fleets across many embodiments, and staffed data-collection sites. The reported ~$1B raise at >$11B was framed around funding compute and scaling real-world data collection.
Latest earnings
n/a
None issued. External reporting (Bloomberg/TechCrunch, Mar 2026) frames industrial pilots in the 2026-2028 window with broader commercial deployment beyond that
- Last priced round
- $600M Series B at $5.6B post (Nov 2025, led by CapitalG; Lux, Thrive, Jeff Bezos returning; Index Ventures and T. Rowe Price new)
- Reported round in talks
- ~$1B at >$11B valuation (Bloomberg, Mar 27 2026; Founders Fund set to participate, Lightspeed in talks - not confirmed closed as of Jul 2026)
- Total raised to date
- ~$1.07B disclosed across three rounds
- Flagship models
- pi0.7 (Apr 2026): steerable, emergent compositional generalization - e.g. operating a never-seen air fryer, folding laundry on a robot it had no folding data for; pi-star-0.6 + RECAP (Nov 2025, arXiv 2511.14759): RL from experience with multi-hour autonomous runs (espresso service, laundry folding, factory box assembly)
- Headcount
- ~80 employees (Bloomberg, Mar 2026)
- Founded
- 2024, San Francisco - Karol Hausman (CEO, ex-Google DeepMind), Sergey Levine (UC Berkeley), Chelsea Finn (Stanford), Brian Ichter (ex-Google), Lachy Groom (ex-Stripe)
Growth drivers
- Model capability compounding — pi0 (Oct 2024, first generalist VLA) -> pi0.5 (Apr 2025, open-world generalization - cleaning unseen kitchens/bedrooms) -> pi-star-0.6 with RECAP (Nov 2025, RL from real-world experience; >2x throughput and roughly half the failure rate on hard tasks) -> pi0.7 (Apr 2026, steerable model with emergent compositional generalization - recombining learned skills to solve tasks it was never taught)
- Partner-led commercialization — deployment partners Weave Robotics (laundry folding - prior commercial deployments in SF Bay Area businesses, now the $7,999 Isaac 0 home unit) and Ultra (industrial workcells; pi0.6 hit 96.4% autonomy on packaging orders over a full shift at a customer site, per Pi's Feb 2026 partner update) - the seed of a licensing funnel
- Open-weight distribution (openpi / pi0 weights released Feb 2025) building developer mindshare and making the pi series a default research substrate, analogous to an early Android play
- Macro tailwind — physical-AI/humanoid funding surge (Figure at $39B, Skild at >$14B, 1X, Apptronik all raising) validates the category and keeps capital available
- Hardware-agnostic positioning lets it ride every robot OEM's volume rather than betting on one body
Bull & bear
Pi is the closest thing to an 'OpenAI of robotics': the best research team in the field, a visible capability compounding curve (pi0 -> pi0.5 -> pi-star-0.6 -> pi0.7), and a hardware-agnostic position that captures value across every robot body. If generalist robot policies follow an LLM-like scaling trajectory, the licensing TAM is measured in trillions of robot-hours and today's $5.6-11B valuation is an early-GPT-2-era entry price.
- Capability lead is documented, not vaporware: pi-star-0.6/RECAP (arXiv 2511.14759) showed multi-hour uninterrupted real-world task runs, and pi0.7 (Apr 2026) added emergent compositional generalization - operating a never-seen air fryer and folding laundry on a robot with no folding training data - a public bar competitors have not matched
- RECAP unlocks the data flywheel bulls need: robots improving from their own fleet experience (RL) breaks the teleoperation-cost bottleneck that capped prior robot-learning economics
- Hardware-agnostic = index bet on the whole robot market rather than a single-body bet; every OEM is a potential customer, none is a channel conflict
- Capital access is proven and deepening: $70M -> $400M -> $600M closed, ~$1B more reportedly in talks within four months of the Series B, with Founders Fund and Lightspeed leaning in alongside CapitalG, Thrive, Lux, Bezos
- Quantified commercial toeholds already exist: Pi's Feb 2026 partner update reports Ultra at 96.4% full-shift autonomy on packaging at a customer site and Weave's pi-powered laundry robots (prior Bay Area business deployments, now the Isaac 0 home unit) - earlier than the 'pure research lab' framing implies
- Open-weight strategy is quietly building the developer standard: openpi/pi0 is the default academic and startup VLA substrate, the same wedge that made Android and Linux platform winners
Pi is a pre-revenue research lab priced like a decacorn on the strength of demos and a narrative. The robot-brain layer may be commoditized from above (DeepMind, NVIDIA give models away to sell other things) and disintermediated from below (vertically integrated OEMs keep the brain in-house), leaving a licensing business with no pricing power - and the physical-data moat is far more expensive to build than the LLM analogy suggests.
- Zero disclosed revenue at a $5.6B (reportedly $11B) valuation: every dollar of value is a bet on milestones that remain years from industrial-grade reliability; even Pi-friendly reporting frames pilots at 2026-2028 and scale deployment beyond
- The 'Android of robotics' analogy cuts both ways - Android won by being free; NVIDIA (GR00T) and Google DeepMind (Gemini Robotics) can price the robot-brain layer at zero to sell chips and cloud, destroying licensing economics
- The direct strategic twin is out-executing on commercial proof: Skild AI raised $1.4B led by NVIDIA and SoftBank at >$14B - above Pi's mark - and took the NVIDIA/Foxconn Houston Blackwell-line deployment, while Figure, Tesla and 1X keep brains proprietary to their bodies, shrinking Pi's addressable OEM pool
- Robot data does not scale like web text: each new task/embodiment needs physical fleets, teleoperators, and site operations - linear (at best) cost per capability gain, unlike the near-zero marginal data cost that made LLM scaling economics work
- Demo-to-dependability gap: folding laundry for hours in a curated video is not 99.9% success across millions of uncontrolled sites; liability, safety certification, and insurance for physical AI have no settled framework
- Valuation re-rated ~2x in four months (Nov 2025 $5.6B -> Mar 2026 >$11B talks) on category heat, not company-specific commercial proof - classic late-cycle private-market momentum, and the round was not confirmed closed as of Jul 2026
- Key-person concentration: the thesis is inseparable from a handful of researchers (Levine, Finn, Hausman); frontier labs are aggressively poaching embodied-AI talent from a team of only ~80 people
What it is worth
Last-priced-round mark plus reported primary-round signals; no fundamentals-based method is honest for a pre-revenue lab - this is a milestone-and-optionality valuation, comparable-round-anchored (Figure $39B, Skild >$14B, humanoid cohort)
Funding cycle turns or NVIDIA/DeepMind commoditize the model layer: down-round toward or below the $5.6B Series B mark, with terminal downside being an acqui-hire by a hyperscaler at research-team value (low single-digit billions) if licensing economics never materialize
Round closes near the reported >$11B; capability milestones continue (pi0.7-class releases), partner deployments (Weave/Ultra-class) multiply but revenue stays immaterial into 2027 - value accrues as an option on the 2028+ deployment wave, marked by each successive round rather than fundamentals
If pi-series models hit industrial-grade reliability and 2-3 anchor OEM licensing deals land by 2027-2028, the horizontal-brain analogy supports $25-40B+ private marks (Figure-cohort parity or better, justified by superior research output) - with genuine long-run optionality far above that if robot fleets scale like smartphone units
Hard marks: $2.4B (Series A, Nov 2024), $5.6B (Series B, Nov 2025, CapitalG-led), and a reported ~$1B raise in talks at >$11B (Bloomberg, Mar 27 2026 - not confirmed closed as of Jul 2026). The ~2x re-rate in four months tracks category beta more than company-specific commercial proof; direct comp Skild AI was marked >$14B in the same window. Any 'fair value' claim beyond these marks would be fabrication; the honest frame is scenario-based.
SWOT
Strengths
- Arguably the densest robot-learning research team in the world — founders wrote much of the canonical VLA/robot-RL literature (SayCan, RT-series lineage via Hausman/Ichter; Levine and Finn labs)
- Demonstrated generalization lead — pi0.5 cleaning wholly unseen homes, pi-star-0.6's RECAP (RL from autonomous experience), and pi0.7's compositional generalization (solving tasks it was never taught) are peer-recognized firsts, published with real multi-hour uninterrupted task videos
- Hardware-agnostic strategy avoids the capital sinkhole of building robot bodies and lets it partner with every OEM without channel conflict
- Blue-chip capital and cap table — Bezos, CapitalG (Alphabet), Thrive, Lux, OpenAI, Sequoia, Khosla, Index, T. Rowe Price - deep pockets for a long pre-revenue runway
- Open-weight pi0 release created developer ecosystem gravity and recruiting pull that closed rivals lack
Weaknesses
- No disclosed revenue two-plus years in; the entire valuation rests on research milestones, not unit economics
- Real-world robot data is scarce and expensive — unlike LLM labs, Pi must pay for teleoperation and physical fleets to generate every marginal training token
- No control of distribution — hardware-agnostic means dependent on OEMs and integrators (many of them better-capitalized or vertically integrated rivals) to reach end customers
- Reliability gap between impressive demos and the 99.9%+ task-success rates industrial customers require; even Ultra's showcase 96.4% shift autonomy implies interventions commercial operators must staff for
- Compute and talent cost structure of a frontier lab without a frontier lab's revenue (no consumer subscription analog yet)
Opportunities
- Become the horizontal software layer ('Android of robotics') across a robot install base that NVIDIA, Morgan Stanley and others project as a multi-trillion-dollar market by 2040-2050
- Labor-shortage-driven demand in logistics, light manufacturing, hospitality, and eldercare as demographics tighten in the US, Japan, and Europe
- License or co-develop with humanoid OEMs that have bodies but weaker brains; RECAP-style fleet learning creates a data flywheel where every deployed robot improves the shared model
- US reindustrialization and supply-chain reshoring policy tailwinds favor US-built automation stacks
- Potential strategic acquirer/partner interest from hyperscalers (Alphabet already invested via CapitalG) if a capability breakthrough lands
Threats
- Google DeepMind (Gemini Robotics), NVIDIA (Isaac GR00T), and Tesla can cross-subsidize robot AI indefinitely from core businesses; NVIDIA is simultaneously Pi's critical supplier and a platform competitor
- Vertically integrated players (Figure AI at a $39B post-money valuation, Tesla Optimus, 1X) may prove the brain and body must co-develop, stranding a software-only layer
- Skild AI - now valued above Pi at >$14B after a $1.4B round led by NVIDIA and SoftBank - landed the marquee NVIDIA/Foxconn Blackwell-line deployment in Houston (Mar 2026): direct evidence the hardware-agnostic-brain niche is contested and better-funded elsewhere
- Chinese ecosystem speed (Unitree, AgiBot, Galbot — context only, not investment calls) compresses hardware costs and could commoditize embodied AI faster than Western licensing models mature
- A capital-markets turn against pre-revenue AI mega-rounds would hit an $11B pre-commercial valuation hardest; safety incidents in physical deployments carry outsized regulatory and reputational tail risk
Moats, dependencies & bottlenecks
Moats
Founders authored the field's canonical work, but frontier labs (DeepMind, OpenAI, Meta) are recruiting the same tiny talent pool aggressively
Real-world robot interaction data + RECAP fleet-learning flywheel compounding Potentially high Proprietary cross-embodiment demonstration and RL-experience data is genuinely hard to replicate; becomes a strong moat only once deployed fleets scale
Documented lead via pi0/pi0.5/pi-star-0.6/pi0.7, but DeepMind's Gemini Robotics and NVIDIA GR00T are converging on the same target with more compute
Standard-setting wedge, but open weights also arm competitors and cap direct monetization of older models
Cycle-dependent ~$1.07B raised, blue-chip cap table; evaporates as a moat if the AI funding cycle turns
Dependencies
Compute supplier / partial competitor Training and inference run on NVIDIA GPUs while NVIDIA's Isaac GR00T competes for the same robot-brain layer - and NVIDIA led rival Skild AI's $1.4B round
Ultra, arm makers e.g. Franka, Trossen, ARX) Distribution channel Hardware-agnostic strategy means Pi reaches end customers only through third-party bodies and integrators it does not control
Pre-revenue lab with frontier-lab burn; multi-year runway requires repeated mega-rounds (next: reported ~$1B at >$11B, unconfirmed as of Jul 2026)
Demonstration data is human-labor-intensive; RECAP reduces but does not remove this dependency
Strategic investor Alphabet is simultaneously an investor (CapitalG led Series B) and the parent of chief rival Google DeepMind - alignment could shift
Frontier-scale training depends on leased GPU cluster availability and pricing; Pi has not disclosed its cloud provider
Advantages
- State-of-the-art published generalization results (pi0.5 unseen-home cleaning; pi-star-0.6 multi-hour autonomous task runs; pi0.7 compositional generalization on never-taught tasks)
- RECAP: first credible production recipe for robots improving from their own experience via RL, with >2x throughput and roughly half the failure rate on hard tasks
- Hardware-agnostic neutrality - every robot maker is a potential customer, none a channel conflict
- Founder team with unmatched academic and industrial robot-learning pedigree (DeepMind, Berkeley, Stanford, Stripe operator DNA)
- Open-weight ecosystem strategy building the default research substrate for VLAs
- Deep-pocketed, strategically diverse cap table spanning Bezos, Alphabet/CapitalG, OpenAI, Thrive, Lux, Sequoia, Khosla, Index, T. Rowe Price
Weaknesses
- No disclosed revenue or commercial contracts of scale as of mid-2026
- Valuation momentum (~2x in four months, per the reported Mar 2026 talks) outpacing commercial proof points
- Dependent on rivals' silicon (NVIDIA) and rivals' parent capital (Alphabet)
- No owned robot body or end-customer channel; monetization model still unproven category-wide
- Physical-data cost curve scales with atoms, not bits - burn stays high even as models improve
- Key-person risk concentrated in a handful of founding researchers on a ~80-person team
Bottlenecks
- Real-world robot data acquisition cost and speed - the binding constraint the LLM playbook never faced
- Task-success reliability — bridging from impressive demos (and 96.4% shift autonomy at Ultra) to the 99.9%+ consistency industrial and consumer deployments demand
- GPU training-compute availability and cost at frontier scale
- OEM adoption pace — no owned distribution means commercialization moves at partners' hardware and deployment cadence
- Safety certification, liability, and insurance frameworks for autonomous physical systems remain unsettled
- Elite embodied-AI researcher supply — a global pool of perhaps a few hundred people, contested by every frontier lab
Top signals & trends
Top signals
Deep-pocket validation and runway; but a ~2x re-rate in four months also flags category froth - close not confirmed as of Jul 2026
TechCrunch-covered capability milestone; verbal coaching lifted a never-seen air-fryer task from 5% to 95% success in Pi's demo
Genuine capability milestone: RL from autonomous experience, >2x throughput, roughly half the failure rate on hardest tasks
Small in revenue terms but the first quantified public proof of the partner-licensing motion
The most visible industrial robot-brain contract and the biggest war chest in the niche both went to Pi's most direct competitor
Strategic validation and information channel to Pi's most capable rival simultaneously
Ecosystem gravity compounding; also arms fast followers
The gap between research milestones and commercial traction is the central open question
Trends
Keeps capital flowing to Pi and validates the category, at the cost of valuation froth and talent-price inflation
Pi defined much of the paradigm (pi0 lineage) and benefits as the field standardizes on its home turf
Pi's RECAP is the flagship public result of this shift; it directly attacks the data-cost bottleneck
Risk of the model layer being given away to sell compute, compressing licensing economics
Structural pull for manipulation-capable robots in logistics and manufacturing this decade
Cheap bodies expand the addressable fleet a hardware-agnostic brain could ride, but also accelerate rival full-stack ecosystems and raise export-control friction
Slows deployment timelines but favors well-capitalized labs with rigorous evaluation practices
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Training/inference GPUs; also a platform competitor via Isaac GR00T and an investor in rival Skild AI
Leased frontier-scale training capacity; Pi has not named its cloud provider publicly
Trossen Robotics, ARX, UFactory) Multi-embodiment hardware fleet for training and evaluation data collection (embodiments supported in openpi)
Deployment partner running Pi models for laundry folding - earlier commercial units in SF Bay Area businesses, now the $7,999 Isaac 0 home unit; Pi reports Weave data cut missed grasps 42% and interventions 50%
Deployment partner running Pi models in industrial workcells; pi0.6 achieved 96.4% autonomy on packaging orders over a full shift at a customer site (Pi partner update, Feb 2026)
The intended licensing base - manufacturers who license the pi brain rather than build their own; no anchor OEM contract publicly announced as of Jul 2026
The founders' alma mater; frontier-scale compute and the RT-series legacy aimed at the same generalist robot-brain layer
Supplies Pi's compute while shipping a competing robot foundation model bundled with the dominant robotics silicon/sim stack - and bankrolling rival Skild
Vertically integrated humanoid program with in-house AI, manufacturing scale, and fleet-data ambitions
Private, $39B post-money (Series C, Sep 2025, Parkway-led with NVIDIA, Intel Capital, Qualcomm Ventures); Helix in-house VLA plus its own humanoid body - the flagship vertical-integration counter-thesis
Private hardware-agnostic robot-brain rival, >$14B valuation after a $1.4B NVIDIA/SoftBank-led round; won the NVIDIA/Foxconn Houston Blackwell-line deployment (Mar 2026) - Pi's most direct strategic twin, now larger
Private; early Pi investor now rebuilding an in-house robotics effort - could pivot from backer to frontal competitor
Absorbed Covariant's foundation-model team in 2024; warehouse-scale deployment surface and in-house demand
Private humanoid maker (NEO home robot) with in-house models; OpenAI-backed
Private humanoid OEM (Apollo) partnered with Google DeepMind - potential customer captured by a rival brain
Private; Digit humanoid in warehouse pilots with its own autonomy stack
Context only, not investment calls: fast-cycling low-cost hardware plus increasingly capable in-house models compress the global cost curve