Physical Intelligence

Phase A — Understand the business

Company Overview

Physical Intelligence ("π", or "PI") is a San Francisco company founded in 2024 that builds a general-purpose foundation model for robot control — a single "brain" meant to run across many different robot bodies. It is explicitly not a hardware company: it builds no robots, no actuators, no humanoid. It sells (or intends to sell) the policy layer — the Vision-Language-Action (VLA) model that turns a camera feed + a natural-language instruction into motor commands.

The product. Its flagship models are π0 (pi-zero), released/open-sourced December 2024, and π0.5, released April 22, 2025. π0 is a VLA built on Google's PaliGemma VLM, trained on data from 7 robot types across 68 tasks plus the Open X-Embodiment dataset, using flow-matching for continuous action generation; it demonstrated laundry-folding, table-bussing, grocery-bagging, box assembly. π0.5 added open-world generalization — cleaning a kitchen/bedroom in homes never seen in training, with performance scaling as training environments grew from 3 → 104 homes, beating a GPT-4-powered planner on messy real-world tasks.

Business model (reported, not confirmed by the company). Secondary write-ups describe a B2B model — a "$300/month subscription per connected robot," an "Android for robots," with on-prem licensing for latency-sensitive deployments and a π0 fine-tuning API in private beta. Critical caveat: PI itself says it has no commercial product, no revenue, and "no defined timeline" for a launch (co-founder Lachy Groom), running a deliberately research-first posture. So the $300/robot figure should be read as analyst modeling of a future motion, not a live price book — write n/a — pre-revenue, not disclosed for any actual revenue line.

Customers/partners. No named paying customers. Reported "first enterprise partnerships" are fine-tuning beta arrangements / data-sharing deals with select robotics companies rather than license sales. The end-market it is aimed at (logistics/manufacturing manipulation — the GXO/Amazon/BMW deployment class) is real, but PI reaches it through the robot builders, not directly.

Contract structure / payment terms. n/a — private, not disclosed. The open-source release of π0 weights is a deliberate go-to-market choice (seed the ecosystem, harvest fine-tuning + data network effects later), not a monetization event.

Supply Chain

PI sits at one node of the robotics stack — the VLA / foundation-policy layer — and is explicitly listed there alongside Skild, NVIDIA GR00T, and Covariant in the compiled map. Mapping the chain around it with named players:

• Upstream compute (its real "supplier"): training + inference compute. PI is a cloud-compute buyer — Groom: "most spending going toward compute … there's always more compute you can throw at the problem". The on-device inference layer for any deployed robot concentrates on NVIDIA Jetson / Thor. This makes NVIDIA both an arms-dealer to PI and a competitor (GR00T) — a structural tension.
• Upstream model: π0 is built on Google PaliGemma — i.e. PI's base VLM is itself a Google artifact, while Google DeepMind's Gemini Robotics is a direct rival.
• The actual input that matters — teleop/robot data: flagged in the compiled layer as the moat candidate and a chokepoint ("fleet-scale teleop data … controlled by top-5 builders … low substitutability"). PI owns no fleet, so it must acquire diverse robot-hours via partnerships, its own data-collection rigs (ALOHA/DROID-class), and licensing — a dependency the vertically-integrated builders (Tesla, Figure) do not have.
• The company → customer: PI → robot builder/integrator → end deployer (logistics/manufacturing). Named builder/customer pairs in the space: Figure↔BMW, Agility↔Amazon/GXO, Apptronik↔Mercedes.

Chokepoint verdict: PI's single-source dependency is diverse robot interaction data at scale — and it is on the wrong side of that chokepoint versus the fleet owners. Its compensating bet is that a cross-embodiment model trained on many partners' data beats any single fleet's narrow data.

Competitive Advantages (moats)

The one durable moat: talent density + research lead. PI has assembled what the field broadly calls the strongest robot-learning bench in existence — see Lens 9. In a domain where the recipe is not yet settled, being first to the working recipe is the moat, and PI has repeatedly shipped state-of-the-art (π0 = "broadest task generalization of any robot policy to date"; π0.5 = first credible open-world home generalization).

Moats that are weaker than the price implies:

• Switching costs / lock-in: the compiled layer notes real switching friction — "teleop data doesn't transfer; locked in per company … actuator+sensor stack changes force re-training of policies". But that friction protects whoever the builder already uses — and the builders increasingly train in-house (Figure's Helix, Tesla's in-house policy), which is the opposite of locking into PI.
• Network effects (the bull's core claim): more partners → more diverse data → better model → more partners. Plausible, but unproven while PI has no scaled deployment generating that data flywheel. Open-sourcing π0 is the attempt to kick-start it.
• IP: the architecture is published and open-sourced — there are already third-party reimplementations (lucidrains/pi-zero-pytorch). The IP moat is therefore thin; the moat is the team's next result, not the last one.

Bargaining power. Weak today. PI needs the builders (for data and distribution) more than they need PI (they can use NVIDIA GR00T's open, commercially-licensed model, or build in-house). Power flips only if π1/π2 is decisively better than every alternative.

Segments

n/a — pre-revenue. No revenue, EBITDA, or earnings by segment or geography exists. PI is a single-product, single-segment research company (one model family, π). The only meaningful "segmentation" is by model generation (π0 → π0.5 → π1) and by robot embodiment supported (5+ platforms, 10+ manipulation tasks from one model) — both capability milestones, not financial segments. Geographic split: HQ + operations in San Francisco / Berkeley; no disclosed international revenue base.

---

Phase B — Measure performance (+private overlay: funding/valuation, cap table, traction)

Funding & Valuation Trajectory

The numbers PI is judged on are rounds, not earnings. Trajectory, seed → latest (all, unaudited):

Cumulative raised: ~$1.07–1.1B across ~3–4 rounds. Conflict flagged: the Series A is reported both as $400M @ $2.4B (The Robot Report) and $330M @ $2.8B (Sacra/roboticscenter) — I do not collapse these; the cumulative ~$1.07B is the more reliable anchor since it's the aggregate multiple sources agree on.

The signal that matters: valuation ~doubled ($5.6B → $11B) in ~4 months with no revenue and no product, on the strength of π0.5 + the team. That is a pure bet-on-research markup. Balance-sheet read: with ~$1B+ raised, "doesn't burn much" (Groom), and spend dominated by compute rather than headcount, runwayRunwayHow long the cash lasts at the current rate of spending. It shortens the moment spending rises, which is why a figure taken from a quiet quarter flatters. is plausibly multi-year — : $1B raised, ~209 staff at, say, $0.4M all-in/head ≈ $80M/yr payroll; even adding $100–200M/yr compute, runway >3 yrs absent a hardware build — caveat: compute can scale without limit by design, so "runwayRunwayHow long the cash lasts at the current rate of spending. It shortens the moment spending rises, which is why a figure taken from a quiet quarter flatters." is a policy choice, not a constraint.

Founder/Leadership Signal

No earnings calls exist; the equivalent signal is what the founders say in interviews/podcasts and how the message holds over time. The consistent, deliberately-held line across 2024→2026: research-first, no commercialization timeline, compute is the lever, talent density is the asset. Groom (ex-Stripe operator) repeatedly frames discipline around not being rushed to product ("no defined timeline," "doesn't burn much") — a confident posture that only a fully-funded lab can afford, and one that will read as either visionary or evasive depending entirely on whether π1 lands. Tone shift to watch: any pivot from "no timeline" → a named commercial product or a flagship customer would be the single most important sentiment change for this name.

Cap Table & Secondary Marks + mechanism comps

Syndicate quality — this is the bull's strongest tangible evidence. The cap table is studded with crossover/strategic capital that signals IPO-proximity and conviction: T. Rowe Price (crossover mutual fund — a classic late-stage-into-public tell), CapitalG (Alphabet growth), Founders Fund, Lightspeed, Index, plus Bezos and Amazon and the OpenAI Startup Fund. A T. Rowe + CapitalG presence is exactly the "crossover funds entering" marker the +private overlay flags as an IPO-proximity tell. Secondary marks / mutual-fund markups: n/a — not disclosed.

Comparables — by category, not by P/E (there are no earnings). Private robot-brain peers, valuation as the comparable:

The most damaging single comp: Skild is valued higher ($14B) AND already has ~$30M of revenue. PI at $11B with $0 revenue is being marked on research lead alone against a peer that has both a higher mark and actual commercial traction. That is the bear case in one row.

Funding & Product Catalysts

The "price-moving events" for a private are rounds + model drops + talent + the rival tape:

• Dec 2024: π0 release + open-source → seed/Series-A momentum.
• Apr 22, 2025: π0.5 open-world generalization paper → the technical proof point that justified the step-up.
• Nov 20, 2025: $600M Series B @ $5.6B (CapitalG, T. Rowe, Index enter).
• Mar 27, 2026: reports of $1B @ $11B (Founders Fund, Lightspeed) — val doubles in 4 months.
• Pattern read: this name re-rates on model releases and round announcements, not on commercial metrics (there are none). The next catalyst that matters is π1 (a reasoning/next-gen model) and/or a first named flagship customer — either would be the re-rating event; their absence through 2026 is the de-rating risk.
• Rival catalysts that move PI's relative narrative: Skild's $14B raise (Jan 2026), NVIDIA GR00T N1.7 (Apr 2026, free + commercial license), Figure's $48B mark — each compresses PI's differentiation.

---

Phase C — Judge people & books

Management

Archetype: founder-led, research-aristocracy. This is the rare case where the team is the thesis.

• Karol Hausman — Co-founder & CEO. Led Google Brain's robot-manipulation moonshot (Robot Manipulation Lead); adjunct professor at Stanford. Operator-researcher hybrid.
• Sergey Levine — Chief Scientist. One of the founding figures of deep RL for robotics; built UC Berkeley's RAIL Lab into a top-tier group. Among the most-cited researchers in the field.
• Chelsea Finn — Research Lead. Authored the landmark MAML meta-learning paper (one of the decade's most influential AI papers); runs Stanford's IRIS Lab.
• Brian Ichter, Quan Vuong — ex-Google DeepMind/Brain robotics researchers, co-authors on the RT-line/Open-X work that defined cross-embodiment learning.
• Lachy Groom — Co-founder. Long-time Stripe executive; the business/capital operator (he is the public voice on funding/runway/discipline).
• Adnan Esmail — Co-founder. Hardware-engineering leader.

(1) Track record: elite on the research axis (MAML, deep-RL-for-robotics, the RT/Open-X cross-embodiment lineage are genuinely field-defining). Unproven on the commercial axis — none of them has built a robotics business to scale; Groom's Stripe pedigree is the only at-scale-company DNA, and Stripe is software-payments, not robotics. (2) Skin in the game: founder-held, well-funded; specific insider ownership n/a — private, not disclosed. (3) Capital allocation: raising aggressively and routing capital to compute + talent rather than a premature product — defensible if the research compounds, value-destroying if the recipe commoditizes (NVIDIA/Google give the model away). (4) Red flags: the "no timeline" posture is a double-edged signal — confident research discipline, or the absence of a business model dressed as virtue. The base VLM dependency on Google (PaliGemma) while Google competes (Gemini Robotics) is a strategic exposure. (5) Founder vs professional manager: decisively founder/research-led — correct for this stage (pre-recipe frontier research), but the company will eventually need a commercial operator layer it does not visibly have yet.

Forensic Red Flags

No audited financials exist — so classic forensic accounting (revenue-recognition, receivables/inventory, SBC, goodwill) is not applicable. The "forensic" lens for a private re-points to valuation integrity and structural risk:

• Valuation-to-fundamentals gap: $11B on $0 revenue, doubling in 4 months. Multiple outlets explicitly invoke the 2023–24 genAI-style markup pattern (Anthropic/Perplexity priced on potential, not metrics) and warn PI "risks becoming the most expensive robotics research project in history" if π1/π2 doesn't demonstrate human-level zero-shot transfer. This is the single biggest red flag — not fraud, but price disconnected from realized fundamentals.
• Capability-vs-claim gap: the founders themselves concede π0.5 "often makes mistakes" in both semantic reasoning and motor commands, and the literature flags VLA models broadly as too slow for on-board real-time, data-starved relative to web-scale, and weak on real-world physics/corner-cases. Demo-to-deployment is a wide gap (the compiled layer flags "real deployment hours vs demo hours" as the key unknown).
• Single-product concentration: the entire enterprise is one model family with no revenue diversification.

Regulatory findings (required). Per the Stage-1 pre-fetch: Physical Intelligence has no CIK and is not an SEC filer — zero SEC Litigation Releases or AAERs are possible/found. Non-SEC web search ("Physical Intelligence" (FTC OR DOJ OR FDA OR consent decree OR settlement OR penalty)) and an IP/litigation search returned no company-specific enforcement, lawsuit, or patent dispute. Material macro/regulatory backdrop (not company-specific): the US Commerce/BIS opened a Section 232 national-security investigation into robotics & industrial-machinery imports (Sep 24, 2025), with possible tariffs/restrictions by spring 2026 — relevant to PI's builder customers and the hardware layer it rides on, not to PI directly. AI-training-data copyright litigation is escalating industry-wide (≈47 suits by mid-2025) — a latent risk for any model trained on web images, though PI is not named. Net: No material regulatory or legal findings against the company — verified via SEC EDGAR EFTS (no CIK), web enforcement search, and IP-litigation search as of 2026-06-18.

---

Phase D — Project & stress-test (+private overlay: IPO-readiness)

IPO-Readiness & Path-to-Tradeable

No EPS to forecast (pre-revenue). The +private question is what unlocks a tradeable security, and when.

• Current stage (vs private-watch.json readiness scale 1–5, where 4 = pre-IPO/secondary-active, 5 = S-1/IPO imminent): PI is a 3 — late-stage, trending toward 4. Evidence for "approaching 4": crossover capital on the cap table (T. Rowe, CapitalG), $11B mark, marquee syndicate (Founders Fund/Lightspeed). Evidence it is not yet 4–5: no revenue, no product, no commercialization timeline, no S-1 signal, no reported secondary-tender activity. (Note: physical-intelligence is listed in research/private-watch.jsonwatch but as a bare slug with no structured stage/ipo_readiness/catalyst fields — this dossier supplies that grounding; no JSON write performed per wave boundaries.)
• Milestones that would unlock an S-1 (the real watch-list):

1. A revenue line at all — even Skild-scale (~$30M) — to move from "research project" to "company." This is the gating item.
2. π1 (a next-gen reasoning/commercial-grade model) shipping with reliable zero-shot transfer across many embodiments at deployment (not demo) reliability.
3. A named flagship customer / fleet generating the data flywheel the moat depends on.
4. A commercial operator/CRO layer + a stated business model the company will actually stand behind.

• Estimated window: IPO is unlikely before 2028–2029. The more probable near-term liquidity is secondary/tender or a strategic acquisition (Google/NVIDIA/Amazon/Tesla are all logical acquirers) rather than an IPO.
• No Brier forecast logged — per --watchlist rules (no our model create in the breadth loop; and the natural binary, "π1 ships with deployment-grade zero-shot transfer by FY-end," is not yet crisply resolvable).

Bull vs Bear

Bull case. Robotics is having its "GPT moment," and PI has the best research team on the planet for the exact problem that defines the category (a general cross-embodiment policy). If the foundation-model-for-robots thesis is right, value accrues to whoever owns the best brain, and a horizontal brain that runs on everyone's hardware is a bigger prize than any single humanoid builder. π0/π0.5 already shipped genuine SOTA (open-world home generalization); the open-source flywheel + crossover-investor conviction (T. Rowe, CapitalG, Founders Fund) suggest sophisticated money sees an IPO-scale outcome. If π1 demonstrates deployment-grade zero-shot transfer across 10+ platforms, the $11B mark looks cheap.

Bear case (2–3 permanent-impairment risks).

1. Commoditization by the platform owners. NVIDIA gives GR00T away (open, commercially-licensed, 3B params, improving fast) and Google has Gemini Robotics. If a "good-enough" robot brain is free from the company that also sells the inference chip, PI's paid model has no floor. The architecture is open-sourced and already reimplemented — the IP doesn't protect the price.
2. The builders go in-house. Tesla (in-house policy), Figure (Helix) train their own brains. PI's addressable market shrinks to builders too small to train their own — precisely the customers least able to pay.
3. The recipe doesn't reach reliability. VLA models are slow on-board, data-starved, and weak on physics/corner-cases; π0.5 "often makes mistakes" by the founders' own account. If the demo→deployment gap proves structural (not just a data-scale problem), PI is "the most expensive research project in history".

Pre-mortem (18 months out, thesis broke): It's late 2027. π1 slipped or underwhelmed on reliability (not benchmarks); NVIDIA GR00T N3 + Gemini Robotics made a free brain "good enough" for most builders; the serious builders all trained in-house; PI still has trivial revenue while Skild and Figure converted; the $11B round becomes the high-water mark and a down-round or acqui-hire follows. The cause was not bad research — it was a horizontal-software bet in a market that integrated vertically and commoditized the horizontal layer.

Are multiples too high? There are no multiples — but $11B on $0 revenue, +97% in 4 months, against a revenue-positive peer (Skild) marked higher, is priced for flawless execution. Any slip de-rates hard.

Contrarian view (what the market refuses to see): The crowd is debating which robot brain wins. The likelier outcome the market under-weights is that the brain layer commoditizes (free from NVIDIA/Google) and the durable value sits one layer down (compute/NVIDIA) and one layer up (the builders + their proprietary fleet data) — leaving the pure-play model labs (PI and Skild) as brilliant research orgs that struggle to capture the value they create. The contrarian trade isn't "PI vs Skild"; it's own the index/arms-dealer, not the model lab.

Devil's Advocate (short-seller)

Dismantling the bull: What breaks the money model? PI has no money model yet — that is the short. The thing it sells (a robot foundation policy) is being given away for free by the two richest, most strategic players in the stack (NVIDIA GR00T, Google Gemini Robotics), and PI's own π0 is open-sourced and reimplemented. Where is revenue concentrated? It isn't — it's zero; the concentration risk is existential (one model, one bet). Why is the moat weaker than bulls think? The moat is "best team + best last result," both of which are flow (must be re-won each generation), not stock; the architecture is public; the data flywheel requires deployments PI doesn't have. Most dangerous competitor bulls underestimate: not Skild — NVIDIA, because it can subsidize a free brain indefinitely to sell chips, structurally undercutting any standalone model price. Worst capital-allocation / incentive flags: raising at a doubling valuation every few months with no revenue and "no timeline" puts maximal pressure on a single future result; the Google PaliGemma base dependency sits under a Google competitor. What must hold for $11B? That a paid horizontal robot brain becomes the industry standard before free/in-house alternatives reach "good enough" — a narrow, time-boxed window. If growth disappoints 20–30%? There is no growth to disappoint — the analog is the next round; a flat or down round (vs the +97% pace) would itself be the de-rating event. Single scenario that permanently impairs: GR00T/Gemini reach deployment-grade and free → PI's TAM collapses to sub-scale builders → revenue never compounds → acqui-hire at a fraction of $11B. Plausibility: moderate-to-high on a 2–3 year horizon — this is the central risk, not a tail.

Management Questions (ordered by information value)

1. What is the paying revenue today, and what is the path to your first $10M of recurring revenue — with named customers and dates?
2. Given NVIDIA GR00T and Google Gemini Robotics are free/near-free, what exactly will customers pay you for, and at what price, that they can't get for $0?
3. When does π1 ship, and what is the deployment-reliability (not benchmark) bar you're holding it to before a customer runs it in production?
4. You own no robot fleet — how do you win the data flywheel against vertically-integrated builders (Tesla, Figure) who do?
5. Your base VLM (PaliGemma) is Google's, and Google competes via Gemini Robotics — how exposed are you, and what's the mitigation?
6. What share of robot builders will train in-house vs license a horizontal brain, and how does that bound your TAM?
7. "No commercialization timeline" — at what point does that flip from research discipline to a missing business model, and what's the internal trigger?
8. What is your actual burn and runway at current compute intensity, and how does the $1B/$11B round change the spend plan?
9. How do you defend a moat when the architecture is open-sourced and already reimplemented?
10. Which is the more likely liquidity path — IPO, strategic acquisition, or staying private long — and what milestones gate each?
11. What's the gross-margin profile of the $300/robot model once on-prem inference and support costs are loaded?
12. How do you keep this research bench from being poached by NVIDIA/Google/Tesla as the race intensifies?
13. What's your plan for real-time on-board inference given VLA latency/compute constraints today?
14. How do you think about regulatory/safety liability when a learned policy controls a physical machine around people?
15. If a free, in-house, or Chinese low-cost brain reaches "good enough" first, what is the plan B for the business?

---