Physical Intelligence Deep Dive: A $2.1B Bet on the GPT Moment for Robotics, With pi0.7 Self-Reporting 60-80 Percent Success on Unseen Tasks and Neither Pricing nor Revenue Disclosed for the Pi API

Executive Summary

Physical Intelligence is a San Francisco robotics foundation model company founded in 2024 by a seven-person founding group drawn from Google's robotics research groups, UC Berkeley and Stanford. It builds no robots. It builds the model layer that other companies' robots run on, and it describes that layer as a "physical intelligence layer" that is supposed to sit to robotics the way a foundation model API sits to software development. The company came out of stealth in March 2024 with a roughly $70 million seed round led by Thrive Capital and has since raised enough to put cumulative funding in the range of $1.9 billion to $2.12 billion across four rounds, with a post-money valuation most consistently reported between $7.5 billion and $11.2 billion.

The most consequential fact about this company is that the most recent primary technical document it published contradicts the most common characterization of it. The pi0.7 paper of April 16, 2026 is unusually self-critical for a frontier lab. It states plainly that while seen tasks often exceed 90 percent success, unseen tasks or unseen task-robot combinations fall in the 60 to 80 percent range. It states plainly that when training on a dataset this large and this diverse, it is "practically difficult" to definitively determine which tasks are truly seen or unseen, because the dataset contains so many scenes and behaviors that potentially related skills may well be present elsewhere. And it reports that some tasks, such as loading a sweet potato into an air fryer, have no action-level training data at all and require step-by-step human language coaching. Secondary coverage of pi0.7, which is dense in China, rarely mentions any of these three concessions. They are the analysis.

The other fact that matters equally is that the financing record is genuinely unsettled, and not merely in the usual way. Two mutually exclusive Series C observations are on the public record and they cannot be reconciled. Corporate filings tracked by the private-market platform Forge show an authorization of about $1.05 billion of Series C stock on May 21, 2026 at a post-money valuation of $11.19 billion, with Founders Fund, Lightspeed Venture Partners, Lux Capital and Thrive Capital named. A funding roundup published on July 1, 2026 records an $800 million Series C at a $7.5 billion post-money valuation led by Thrive Capital with NVIDIA added as a strategic co-investor. A third tracker records a $1 billion round on July 12, 2026 led by Jeff Bezos with Founders Fund alongside. The July figures are lower than the May figure, which is what makes them hard: a later round cannot normally price below an earlier one. Cumulative funding is variously stated as $1.87 billion, $1.9 billion, about $2.1 billion and $2.12 billion, and the number of publicly announced rounds is three while the number of rounds that actually closed is four.

The commercial direction is real but early and thin. On October 13, 2026 the company opened a Pi API beta in which a customer uploads robot demonstrations, fine-tunes on infrastructure Physical Intelligence operates, and serves the tuned policy through a hosted inference gateway. Access is waitlisted and batch-onboarded. No price, no billing unit, no latency figure, no minimum commitment and no free tier have been published. The named early deployments are commercially meaningful: Actor runs an autonomous wheel loader loading dump trucks in Utah with no operator in the cab, Almond uses a dual-arm Axol robot to deburr spacers for Axol's own assembly after four days of iteration, and partner Weave reports that folding laundry with pi0.6 reduced missed grasp sequences by 42 percent and human interventions by 50 percent once Weave's own data entered pre-training. The company also openly acknowledges network connectivity problems at some sites.

A note on naming. Physical Intelligence (pi) is a distinct company from Physical Superintelligence (PSI), the Cambridge-based virtual-physics laboratory, and from Physical AI as a general term. The two "Physical" companies are routinely conflated in secondary coverage. They share no founders, no investors in common, no products and no funding history. This report treats them as separate entities throughout.

Company Overview

Physical Intelligence was founded in 2024 and is headquartered in San Francisco. Its own careers page describes the location as the Mission District. The company's public self-description is "a group of engineers, scientists, roboticists, and company builders developing learning algorithms to create a model that will control any robot to do any task," and the corporate site enumerates more than eighty named individuals in its team roster, ending with the phrase "and growing."

The founding group is conventionally counted at seven. The names that recur across primary documents are Karol Hausman, Brian Ichter, Quan Vuong and Adnan Esmail from Google's robotics research efforts, Sergey Levine, who holds a faculty position at UC Berkeley, Chelsea Finn, who holds a faculty position at Stanford, and Lachy Groom, an early Stripe employee who later became an investor and is generally described in the founding context as an investor-operator. The company has never published a formal founder-title page, and most secondary coverage names only three of the seven, Hausman, Levine and Finn, which is why the founding-group count is itself one of the recorded discrepancies below.

Karol Hausman is the chief executive and is the figure through whom most of the public record runs. The 2025 paper author list for pi0.5 includes him alongside Sergey Levine, Chelsea Finn, Karol Hausman and Brian Ichter, and the pi0.7 paper of April 2026 carries an author list of roughly eighty-seven names that includes all of them. The pi0 paper was published in Robotics: Science and Systems 2025.

Headcount is reported three different ways and the spread is wide enough that any single number should be treated as an estimate of an estimate. One 2026 industry write-up describing the company's March 2026 funding discussions states that the company had fewer than 80 employees and no revenue. A commercial database that tracks employment from public professional-network data recorded 361 employees as of October 5, 2026. A robotics-focused directory records a headcount band of 50 to 100 and lists the founding year as 2023 rather than 2024. A Chinese business-data aggregator lists 1,550 people, a figure that is inconsistent with every other source and with the company's own published roster.

The corporate structure of Physical Intelligence is a Delaware corporation rather than a public benefit corporation, which distinguishes it from several neighbors in this survey including Physical Superintelligence. It has no debt disclosed, no IPO filing, and no ticker.

Common Misreadings: Seven Errors That Change the Analysis

1. Reading the valuation as a single number. There are at least three mutually inconsistent Series C records on the public record and they point in different directions. A research summary that says "Physical Intelligence is valued at $11 billion" has silently selected one of them. The correct statement is that the most defensible figures are a roughly $1.05 billion Series C authorized at an $11.19 billion post-money valuation on May 21, 2026 per private-market certificate records, and a separately reported $800 million Series C at a $7.5 billion post-money valuation on July 1, 2026 with Thrive Capital leading and NVIDIA participating. Both are recorded here without merging them.

2. Assuming the later round must be worth more. The temptation is to prefer the $11.19 billion figure because it is larger and matches the March 2026 Bloomberg reporting of talks at above $11 billion. The obstacle is that the $7.5 billion figure is dated two months later and names a different lead investor set. Either the two observations describe different transactions, or one of them is wrong. It cannot be resolved by preferring the larger number.

3. Treating the pi0.7 success rates as 85 to 90 percent. This figure circulates widely in secondary coverage and does not appear in the paper. The paper's own text states that seen tasks often have success rates in excess of 90 percent while unseen tasks or unseen task-robot combinations have success rates in the 60 to 80 percent range. The 60 to 80 percent band is the number that belongs in a due-diligence model, and it is the one the authors chose to write down.

4. Reading "unseen" as a controlled claim. The pi0.7 authors state that it is "practically difficult" to definitively determine which tasks are truly seen or unseen when training on a dataset this large and diverse, because potentially related skills may well be present elsewhere in the data even when the researchers did not deliberately collect them. The evaluation of 14 instruction-following scenarios across 4 unseen kitchens and 2 unseen bedrooms is real, and it is also weaker than the phrase "unseen environments" implies.

5. Assuming the pi-series is one model. pi0 is a 3.3 billion parameter flow-matching VLA built on a PaliGemma backbone with a 300 million parameter action expert. pi0.5 is a different training recipe on the same family of architecture. pi*0.6 is a roughly 5 billion parameter model post-trained with RECAP, a method combining demonstrations, human corrections and reinforcement learning from autonomous experience. pi0.7 is a roughly 5 billion parameter steerable model on a Gemma3 4B backbone with an 860 million parameter flow-matching action expert. The progression is real and the lineage is coherent, but "pi0" is used loosely in most coverage to cover four materially different systems with different weights-availability status.

6. Treating the open release as the current model. The openpi repository distributes pi0, pi0-FAST and pi0.5 under Apache-2.0, along with fine-tuned checkpoints for DROID, ALOHA and LIBERO. pi0.6, pi-star-0.6 and pi0.7 have no publicly released weights, and the Pi API does not name which model version it serves. Reading "openpi is available" as "the current model is available" overstates by two generations.

7. Assuming pi0.7 was validated in a competitive setting. It was not. The one international competition Physical Intelligence voluntarily entered, Benjie's Humanoid Olympic Games, was won by Beijing's Xingdong Jiyuan on three tasks, breaking records that Physical Intelligence had held, using the closed pi-star-0.6. On the orange-peeling task Physical Intelligence used a peeler and took 2 minutes 46 seconds against Xingdong Jiyuan's tool-free 1 minute 47 seconds; on the lock-opening task 66 seconds against 49 seconds; on the sock-turning task 1 minute 33 seconds using 176 samples against 1 minute 4 seconds using 120 samples. The competition's rule is that a challenger must beat the standing record by at least 25 percent to take the medal.

Technology: The pi Series, Step by Step

pi0, October 2024. The first generalist policy. The architecture is a pre-trained PaliGemma vision-language model backbone with a separately initialized 300 million parameter action expert trained with a flow matching objective, for a total of about 3.3 billion parameters. The paper states that the pre-training mixture consists of the company's own dexterous manipulation datasets collected on 7 different robot configurations for 68 different tasks, plus the entire OXE open dataset, which contains data from 22 robots. The own-data component is about 10,000 hours of demonstrations. The paper quantifies the mixture in timesteps and states that 9.1 percent of the training mixture consists of open-source datasets including OXE, Bridge v2 and DROID. The action chunk horizon is H equals 50, the model controls robots at frequencies of up to 50 Hz for dexterous tasks, and inference integrates the learned vector field over 10 steps.

pi0-FAST, January 2025. An autoregressive rewrite that introduced FAST, Frequency-space Action Sequence Tokenization. The reasoning is specific: standard per-dimension binning of robot actions fails on high-frequency dexterous tasks because consecutive action vectors are highly correlated and binning each dimension independently produces an enormously redundant token stream. FAST applies a per-dimension discrete cosine transform to action chunks, prunes low-magnitude high-frequency coefficients, and applies byte-pair encoding to the resulting integer sequence. The result is a compact and effectively lossless tokenization governed by only two hyperparameters. A pretrained universal tokenizer, FAST+, was trained on one million action sequences from single-arm, bimanual and mobile manipulation platforms. The principal measured benefit is training speed, up to five times faster than the flow-matching version on the same data, at the cost of an inference cost roughly four to five times higher. Weights and code for pi0 and pi0-FAST were released on February 4, 2025.

pi0.5, April 2025. The open-world generalization model, and the point at which the company moved from architecture to data recipe. The paper describes co-training on heterogeneous tasks using data from multiple robots, high-level semantic prediction, web data and other sources, with hybrid multi-modal examples that combine image observations, language commands, object detections, semantic subtask prediction and low-level actions in one sequence. The directly relevant mobile manipulation slice is about 400 hours of data collected with mobile manipulators in about 100 different home environments, and the paper states that after only about 100 training environments the model approaches a baseline trained directly on the test environments. The measured language-following results are: in-distribution follow rate 86 percent and success rate 83 percent; out-of-distribution follow rate 94 percent and success rate 94 percent. The ablation table is the more useful part of the paper. Removing web data leaves in-distribution nearly unchanged at 86 and 82 percent but drops out-of-distribution follow to 80 percent and success to 74 percent. Removing cross-embodiment data drops the four numbers to 74, 67, 67 and 49 percent. Removing multi-environment non-mobile data drops them to 66, 57, 33 and 31 percent. The company's own conclusion is that web data makes the biggest difference for generalizing to out-of-distribution objects while data from other robots matters across all evaluation conditions. The out-of-distribution figure exceeding the in-distribution figure is not a typo but a consequence of the evaluation design, and it is worth noting that a reader who takes 94 percent as "the" pi0.5 number will be reading the easiest of the two splits.

pi-star-0.6, November 2025. The reinforcement-learning model, and the one that changed the reliability story. The method is RECAP, RL with Experience and Corrections via Advantage-conditioned Policies. The stated motivation is that imitation alone fails because small errors compound: a gripper in the wrong spot, a missed grasp or a knocked-over object produces a slightly different situation, in which the model is more likely to make a bigger mistake. RECAP adds two signals on top of demonstrations: expert teleoperator corrections that fix the mistakes the robot actually makes, and reinforcement learning from autonomous experience, with the policy conditioned on an advantage value so it can distinguish better from worse behavior. The model is built on a 5 billion parameter vision-language model with an action expert and supports heterogeneous prompts including desired execution quality and desired action advantage. The measured result is that on the hardest tasks, throughput more than doubles and failure rates fall by a factor of two or more. The demonstrations are long-duration and continuous: making espresso drinks from 5:30 in the morning to 11:30 at night, folding 50 different novel laundry items in a new home, and assembling and labeling 59 boxes used for real packaging in a factory. Weights and training code were not released.

MEM and RLT, March 2026. Two shorter posts. Multi-Scale Embodied Memory gives the models both long-term and short-term memory and enables tasks longer than ten minutes. The following post, on precise manipulation with efficient online RL, extracts an RL token from VLA models to enable fast online RL and improve throughput on precise tasks with only a few hours of data. The RL token idea matters because it addresses the central cost objection to RL post-training in robotics, which is that on-robot data is slow and expensive to collect.

pi0.7, April 2026. The current model, and the most carefully self-qualified document in the company's output. The architecture is roughly 5 billion parameters: a vision-language backbone initialized from the Gemma3 4 billion parameter VLM with a 400 million parameter vision encoder, a MEM-style video history encoder with temporal and spatial compression over history observations, and an 860 million parameter flow-matching action expert, with knowledge insulation applied so that action-expert gradients do not flow into the VLM. The novel element is diverse context conditioning. The prompt may carry a language task description, a detailed subtask instruction, episode metadata describing data quality and strategy, a control-mode tag, and up to three generated subgoal images. The subgoal images come from a lightweight world model initialized from BAGEL, a 14 billion parameter mixture-of-transformers image understanding, editing and generation model, with web-scale pre-training and augmentation from egocentric human video. The dropout design is deliberate and documented: subgoal images appear in only 25 percent of examples in each batch, the subtask instruction is dropped another 30 percent of the time when subgoal images are present, episode metadata is dropped entirely 15 percent of the time, and each metadata component is dropped with 5 percent probability individually. The point is to prevent the model from learning to rely on any single context channel.

The three headline results are worth separating from the caveats. First, out-of-the-box dexterity: pi0.7 matches the RL specialists released with pi-star-0.6 on all tasks considered in the paper, and on the difficult laundry and box-building tasks it exceeds them in throughput, while also closely matching SFT specialists trained on top of pi0.6 for other dexterous tasks. Second, broad instruction following: across 14 instruction-following scenarios, each a sequence of 3 to 6 open-ended instructions, spanning 4 unseen kitchens and 2 unseen bedrooms, pi0.7 significantly outperforms pi0.5 and pi0.6 and attains high absolute success rates, with the subgoal-image variant further ahead still. Third, zero-shot cross-embodiment transfer: a bimanual UR5e setup folds a t-shirt with no laundry-folding data collected for that robot, and the paper reports that task progress matches the zero-shot performance of the most experienced human teleoperators, based on a study of 10 operators averaging 375 hours of teleoperation experience across all robots, all of whom rank in the top 2 percent of the operator pool by experience and none of whom had performed the shirt-folding task on the UR5e before.

The caveats are in the same paper. Unseen tasks or unseen task-robot combinations land in the 60 to 80 percent success range against above 90 percent for seen tasks. The authors concede it is "practically difficult" to determine definitively which tasks are truly seen or unseen given dataset scale and diversity, noting that some generalization experiments used tasks for which they did not deliberately collect data while the dataset may well contain potentially related skills elsewhere. And coaching is a genuine dependency for some tasks: for loading a sweet potato into an air fryer, the model has no action-level training data and must follow live step-by-step human language instructions, after which the resulting verbal coaching data can be used to fine-tune pi0.7 into a high-level policy that maps observations, task specification and subtask history to the next subtask instruction. The autonomy of the system therefore ranges from near-total on trained task families to near-zero on untrained ones, with a human in the loop bridging the gap.

Product and Commercial Model: The Pi API Beta and Partner Economics

On October 13, 2026, under the title "Frontier Models for Every Robot," Physical Intelligence opened a beta of the Pi API. The workflow is: upload demonstrations recorded on your own robot, run fine-tuning on compute and infrastructure that Physical Intelligence operates, then serve the tuned model through a hosted inference gateway the robot calls during operation; collect failures and evaluation data to drive the next iteration. The framing is that robotics teams should not have to build a machine-learning infrastructure team before they can put a learned skill on their robot, the same way an application developer calls a language model API.

What is not published is as important as what is. Access is a waitlist with onboarding in batches, and the application form asks about cloud, local or both without establishing local availability. There is no public training rate, no inference rate, no billing unit, no reserved-capacity pricing, no failed-run billing policy, no storage or transfer charge, no minimum commitment and no free tier. No latency figures and no supported-hardware list have been published. The announcement references pi0 and pi0.5 as foundations but does not name the version the beta serves, and the openpi checkpoints are explicitly not proof of the beta endpoint's identity. The company's own material claims more data-efficient fine-tuning and improved partner capability, with no published figures behind those claims.

The named early deployments do establish that this is more than a laboratory demonstration. Actor, a Utah heavy-machinery firm, runs an autonomous wheel loader that scoops material and loads a dump truck in what the company calls a commercial deployment with no operator in the machine. Almond, a San Francisco robot maker, uses its dual-arm Axol robot to deburr spacers used in Axol's own assembly, with a model built after four days of iterating with the API. Also named are AgRobotics, CloudChef, Ultra, Weave, Chef Robotics, Deft, Formic, General Instinct, Argon Robotics, Dream Machines, Foundry, Frontier Machines, Intuition.dev, Omakase, robot.com, Yondu and an unnamed dairy partner, though for most of these no task detail is given. Deployments span Japan, Switzerland and multiple sites in the United States, most running cloud-hosted inference.

One operational detail is disclosed openly and is a real deployment risk: Physical Intelligence states that some partners hit network connectivity problems. For a factory, warehouse or farm with patchy coverage, how the robot behaves when the link drops is a deciding question, and it is a question the company has not yet answered publicly.

The most useful commercial evidence in the company's possession is not from the API but from two partners who published their own results in the Physical Intelligence Layer post of February 24, 2026. Weave, which builds laundry-folding robots and has them deployed at customer laundromats in the San Francisco Bay Area, reports that pi0.6 significantly increases autonomy compared to pi0.5 and that adding Weave's own data into pre-training reduced missed grasp sequences by 42 percent and human interventions by 50 percent per full laundry load, with folding a full load taking 30 to 90 minutes. Ultra, which builds industrial robots for e-commerce order packaging, reports a full customer shift at 96.4 percent autonomy with pi0.6 and a further throughput increase after adding Ultra's own pre-training data, plus a qualitative note that pi0.6 decomposes complex workflows into smaller autonomous subtasks and draws on a more diverse pool of recovery strategies in edge cases. Ultra separately disclosed a $50 million Series A led by Framework Ventures in October 2026, totaling $62 million, leasing warehouses picking robots on a robots-as-a-service model built on Physical Intelligence's models. These are partner-written figures on Physical Intelligence's own page, which is stronger than vendor self-reporting but still not independent verification.

Competitive Landscape

Physical Intelligence sits in the model layer, and its argument is that the model layer should be hardware-agnostic while the hardware layer is crowded and consolidating. The competitors split into four groups.

Direct generalist-model peers. Skild AI, incubated from Carnegie Mellon University's robotics institute in Pittsburgh, is the closest analogue and the furthest ahead on disclosed commercial metrics. Skild reported crossing $100 million in annual recurring revenue ten months after its first commercial deployment, with more than 60 paying customers and about $50 million already recognized, with roughly 90 percent of revenue from manipulation and 10 percent from mobility including the Zebra Technologies robotics arm it acquired in April 2026. Skild's January 2026 Series C of $1.4 billion was led by SoftBank at a valuation above $14 billion, and its S1 model of August 2026 introduced in-context learning from a single video example without weight updates. Skild's own framing is worth noting because it is aimed directly at Physical Intelligence's position: it argues that deployment is the hidden pillar of robotics research and that watching demo videos is not belief. Physical Intelligence has no comparable revenue disclosure.

Platform and hyperscale competitors. Google DeepMind's Gemini Robotics is positioned, according to October 2026 reporting, as a licensable Android-style platform for many robot makers rather than a vertically integrated hardware effort. NVIDIA's Isaac GR00T N1.7 is the leading open option among current models, with roughly 3 billion parameters, Apache-2.0 code and weights under the NVIDIA Open Model License, alongside a large installed base of hardware partners. These two have the advantage of distribution and the disadvantage of being attached to their own platforms.

Lab-adjacent entrants with conflicting positions. OpenAI is simultaneously an investor in Physical Intelligence, having backed it since at least March 2024, and a competitor, having formally announced OpenAI Robotics in late May 2026. Anthropic, which has made four known acquisitions in 2026 and confidentially filed for an IPO on June 1, 2026, is not a robotics competitor but has been reported to have held acquisition talks with Physical Intelligence in the spring of 2026. The optics of a frontier lab holding talks to buy the leading robotics model company while its chief rival owns a stake in the same company is worth stating plainly: it is a live tension in the cap table and in the strategy, not a hypothetical.

Hardware-side and adjacent competitors. Figure AI at roughly $39 billion, 1X at roughly $10 billion, AgiBot with GO-2, Generalist with GEN-1.5, Sunday Robotics at a $1.15 billion valuation on $165 million raised, FieldAI with more than $135 million in revenue and contracts from over 30 customers, and Unitree at roughly $28 billion. FieldAI is the more direct comparable to Physical Intelligence on the software-only business model and reports a substantially larger commercial footprint than Physical Intelligence has disclosed.

The benchmark problem. There is no generally accepted standard benchmark for robot foundation models. Each vendor evaluates on tasks it has selected, and the pi0.7 paper's own admission that it cannot definitively separate seen from unseen tasks in its own dataset is the sharpest available statement of the underlying difficulty. This is why third-party registries conclude that pi-star-0.6 and pi0.7 headline figures are self-reported and not independently reproduced, and why the reported lack of any verified third-party production deployment of these models matters more than it would for a software product.

Financing

Four rounds are recorded across three public announcements plus one set of certificate filings.

The seed round closed on March 12, 2024. The certificate record shows $69.65 million raised at a $424.99 million post-money valuation, $34.63 per share, with Anti Fund, Flat Capital, Jeff Bezos, Khosla Ventures, Lux Capital, OpenAI, Outset Capital, Sequoia Capital and Thrive Capital listed. Thrive Capital led. Reported media coverage rounds this to $70 million at roughly $400 million.

The Series A closed on November 4, 2024, eight months after pi0 was released. The certificate record shows $398.68 million raised at a $2.4 billion post-money valuation, $134.70 per share, alongside a separate Series A-1 tranche of $4.67 million at $35.00 per share. The investor list is long and includes Amazon Industrial Innovation Fund, Bezos Expeditions, Bond Capital, OpenAI, Redpoint Ventures, Samsung Electronics and Sequoia Capital, with Jeff Bezos, Thrive Capital and Lux Capital leading. CNBC reported the $2.4 billion post-money valuation while Reuters put it at $2 billion, and that gap has never been reconciled.

The Series B closed on November 20, 2025, the same week as pi-star-0.6. The certificate record shows $599.93 million raised at a $5.6 billion post-money valuation, $271.56 per share, with CapitalG leading and a long syndicate including Index Ventures, T. Rowe Price, Lux Capital, Thrive Capital and Jeff Bezos returning, and new entrants including 137 Ventures, Abstract Ventures, Alkeon Capital Management, Alpen Capital, Astro Capital, Black Box Ventures, Emergence Equity Management, Factorial Funds, G Squared, Hanwha Impact Partners, Mangusta Capital, Positive Sum and Trust Ventures. One Chinese business-data aggregator records the same $600 million round with T. Rowe Price, Lux Capital and CapitalG named, but dates it November 21, 2025 and classifies it as a B round, which is correct on the round letter and off by one day on the date.

The Series C is where the record splits. The certificate record dated May 21, 2026 shows $1.05 billion raised at a $11.19 billion post-money valuation, $502.90 per share, with Founders Fund, Lightspeed Venture Partners, Lux Capital and Thrive Capital listed. The company has not announced this round and has not named a lead. This connects directly to the March 27, 2026 Bloomberg reporting, carried by TechCrunch and others, that Physical Intelligence was in early talks to raise about $1 billion at a valuation above $11 billion, with Founders Fund expected to join and Lightspeed in talks, and that terms could still change. Separately, a funding roundup published July 1, 2026 records an $800 million Series C at a $7.5 billion post-money valuation led by Thrive Capital, with Jeff Bezos, Sequoia, OpenAI Startup Fund, Redpoint, Bond Capital, Lux, Khosla and NVIDIA listed. A third tracker records a $1 billion round on July 12, 2026 led by Jeff Bezos with Founders Fund alongside.

Cumulative funding is therefore reported four ways: $1.87 billion in the July 1 roundup, $1.9 billion on the same tracker's company page, about $2.1 billion in a research summary, and $2.12 billion on the certificate platform's own page. The distinction that matters most is between total raised and publicly announced: approximately $1.07 billion was publicly announced across three rounds, and the fourth round of about $1.05 billion appears only in filings.

The valuation trajectory across the certificate records runs $424.99 million, $2.4 billion, $5.6 billion and $11.19 billion, roughly a 26-fold increase in 26 months, with no commercial product in place at any point in that sequence.

DISCREPANCIES

1. Series C size and valuation: $1.05 billion at $11.19 billion versus $800 million at $7.5 billion versus $1 billion led by Bezos. The certificate platform records an authorization of about $1.05 billion of Series C stock on May 21, 2026 at a $11.19 billion post-money valuation with Founders Fund, Lightspeed, Lux and Thrive. The July 1 roundup records $800 million at $7.5 billion led by Thrive with NVIDIA added as a strategic co-investor. A third tracker records $1 billion on July 12, 2026 led by Jeff Bezos with Founders Fund. These are three different observations with different dates, different lead investors and different valuations. Ruling: three separate observations, not merged. Source level: filings-derived records and funding trackers, all unconfirmed by the company, which has not announced any Series C.

1. The July valuation is lower than the May valuation. $7.5 billion on July 1 sits 33 percent below $11.19 billion on May 21, roughly two months earlier. A later priced round normally does not sit below an earlier one. Ruling: this internal ordering inconsistency is the strongest argument that the two records describe different transactions or that one is erroneous. It cannot be resolved by preferring the larger figure. Source level: both records are unconfirmed.

1. Lead investor conflict: Thrive Capital versus Founders Fund and Lightspeed. The certificate record lists Founders Fund, Lightspeed Venture Partners, Lux Capital and Thrive Capital for the May 21 round. The July 1 roundup names Thrive Capital as lead, a different syndicate including NVIDIA, and a different total. Ruling: not reconcilable; both are recorded. Source level: certificate records versus funding-roundup database.

1. Cumulative funding: $1.87 billion versus $1.9 billion versus about $2.1 billion versus $2.12 billion. The July 1 roundup sums the four rounds it records to $1.87 billion, and the same tracker's company page shows $1.9 billion total funding. A research summary states about $2.1 billion, and the certificate platform states $2.12 billion. Ruling: the certificate-derived sum of $2.12 billion and the announcement-only sum of about $1.07 billion are both defensible for different questions, and the intermediate figures differ mainly in which Series C they assume. Source level: all database-derived; no company-confirmed total exists.

1. Number of rounds closed versus number of rounds announced. Four rounds are documented; three were announced publicly, and the fourth appears only in filings. A research summary draws the distinction explicitly as about $2.1 billion raised against about $1.07 billion publicly announced. One commercial database shows one funding round of $400 million and labels it a Series A, which is simply wrong. Ruling: four rounds closed, three announced.

1. Series A valuation: $2.4 billion versus $2 billion. The certificate record shows $2.4 billion post-money on November 4, 2024; CNBC reported $2.4 billion while Reuters put it at $2 billion. Ruling: the certificate figure is the primary record; the Reuters figure is a contemporaneous report with no stated derivation. Source level: filings versus media.

1. Series B date and valuation: November 20, 2025 at $5.6 billion versus October 14, 2025 at $5.5 billion versus November 21, 2025 at $600 million. The certificate platform and multiple outlets give November 20 and $5.6 billion. A pre-IPO valuation tracker gives October 14, 2025 and $5.5 billion. A Chinese business-data aggregator gives November 21, 2025 and $600 million without a valuation. Ruling: the certificate date and valuation are the primary record; the tracker figure is a derived estimate. Source level: filings versus derived databases.

1. Seed valuation: $424.99 million versus approximately $400 million versus not recorded. The certificate record shows $424.99 million post-money at $34.63 per share on March 12, 2024. Media coverage rounds to approximately $400 million. One pre-IPO tracker records the seed amount with no valuation at all. Ruling: use the certificate figure.

1. Company founding year: 2024 versus 2023 versus November 2023. The company came out of stealth in March 2024 and most sources give 2024. A robotics directory records 2023. The same directory describes the seed as November 2023, Series A as November 2024 at $70 million and Series B as November 2025 at $600 million, which scrambles the round-to-year mapping. Ruling: founded 2024, seed March 2024. Source level: that directory's record is internally inconsistent with the certificate records on every line.

1. Headcount: fewer than 80 versus 361 versus 50 to 100 versus 1,550. A March 2026 funding write-up says fewer than 80 employees and no revenue. A commercial database tracking public professional-network data recorded 361 as of October 5, 2026. A robotics directory records a band of 50 to 100. A Chinese aggregator lists 1,550. The company's own site roster enumerates more than eighty named individuals. Ruling: the company's own named roster and the professional-network count are the two meaningful figures, and the gap between them is most plausibly the difference between named contributors and total staff. The 1,550 figure is not credible. Source level: company site, database estimate, and one aggregator record that contradicts all others.

1. Founding group size: seven versus three. The pi0.5 and pi0.7 author lists, the company site and multiple write-ups support seven co-founders: Hausman, Ichter, Vuong, Esmail, Levine, Finn and Groom. Most secondary coverage names only Hausman, Levine and Finn. Ruling: seven co-founders, of whom three are cited in most coverage. Source level: primary company documents.

1. pi0.7 unseen-task success rate: 60 to 80 percent versus 85 to 90 percent. The paper states that unseen tasks or unseen task-robot combinations have success rates in the 60 to 80 percent range, while seen tasks often exceed 90 percent. The 85 to 90 percent figure appears in secondary commentary but not in the paper. Ruling: 60 to 80 percent is the paper's own number and is the figure to use. Source level: primary paper.

1. The strength of the "unseen" claim. The paper's own discussion states it is practically difficult to definitively determine which tasks are truly seen or unseen given the scale and diversity of the training data, and that potentially related skills may well be present elsewhere in the data even for tasks the researchers did not deliberately collect. Secondary coverage presents the 4 unseen kitchens and 2 unseen bedrooms evaluation as a clean generalization result. Ruling: the evaluation is real and its scope is narrower than the framing. Source level: primary paper, self-qualified.

1. Open-weights availability versus current model. The openpi repository distributes pi0, pi0-FAST and pi0.5 under Apache-2.0 with DROID, ALOHA and LIBERO fine-tuned checkpoints. pi0.6, pi-star-0.6 and pi0.7 have no released weights, and the Pi API does not name the version it serves. One source describes pi0's open release as beginning February 4, 2025, matching the company announcement. Ruling: the open line ends at pi0.5; two generations are paper-only. Source level: repository documentation and multiple secondary registries agree.

Risks and Due Diligence Gaps

No revenue is disclosed anywhere, and the absence is structural. The company sells no product in the conventional sense, has never published revenue, ARR, customer count, contract value or backlog, and the Pi API has published no price at all. The closest available proxy is partner-reported deployment metrics: Weave at 96.4 percent autonomy for a full Ultra shift, and the 42 percent and 50 percent improvements at Weave. These are operating metrics of a partner's business, not the model company's revenue.

The competitive benchmark cannot be trusted to be settled. There is no accepted standard benchmark in robot foundation models. Each vendor selects its own tasks, and the pi0.7 authors themselves cannot rigorously separate seen from unseen tasks in their own data. This makes every cross-vendor capability comparison, including the claims in this report's competitive section, structurally weak.

The company has already lost a public capability contest. Benjie's Humanoid Olympic Games is the one international event Physical Intelligence entered voluntarily, and it lost three records on April 10, 2026 to Xingdong Jiyuan, which became the only Chinese embodied company on the leaderboard. The margin requirement was 25 percent and the winning margins were comparable or larger. Any narrative that treats Physical Intelligence as the unchallenged leader of the model layer has to contend with this.

The open-versus-closed line has moved against the company. The models with the most impressive published results, pi-star-0.6 and pi0.7, are exactly the ones that cannot be inspected, run or reproduced by outsiders. The Pi API compounds this by not naming which model it serves. A customer buying managed fine-tuning and hosted inference is buying an opaque artifact on an unpublished price.

Network dependency is an unresolved production risk. The company itself discloses that some partners hit network connectivity problems, while most early deployments run cloud-hosted inference. Latency-sensitive scenarios and sites with patchy coverage are precisely where a cloud-hosted robot policy is hardest to run.

The financing record is not clean enough to price. An unannounced round with no named lead, a second and third observation that conflict with it on both date and valuation, and four different cumulative totals mean any valuation-based comparison with Skild AI at above $14 billion, Figure AI at roughly $39 billion or 1X at roughly $10 billion rests on which record is chosen.

The Anthropic episode left the strategic question open rather than closed. The Information reported that Anthropic and Physical Intelligence did hold acquisition talks in the spring of 2026; the viral rumor that Anthropic was buying the company was denied by chief executive Karol Hausman to employees via a Slack message containing a shaking-head GIF. The two statements are compatible: a denial of an active acquisition does not rule out past conversations, and OpenAI's existing stake, with whatever information or refusal rights may attach to it, complicates any sale to a competitor. No reason for the collapse of the talks has been reported.

Documentation gaps worth naming. The company publishes no team page or founder-title page; the official website's team route returns a 404 and the careers page is the usable substitute. There is no published pricing, latency, SLA or supported-hardware list. There is no disclosed revenue, customer list or deployment count of its own. There is no independent reproduction of any headline result.

Overall Assessment

Physical Intelligence is the strongest research organization in the robot foundation model space by paper quality alone, and that is a genuinely high bar. Four model generations in twenty-two months, each with a specific technical thesis that addresses a specific failure of the last one; a training corpus of about 10,000 hours across 7 robot configurations and 68 tasks as the disclosed scale of the first release; an ablation table in pi0.5 that isolates the contribution of web data versus cross-embodiment data versus multi-environment data; a reinforcement-learning recipe that addresses the compounding-error problem that imitation cannot solve; and a current model that is simultaneously more capable and more honest about its own limits than any comparable publication in the field.

The commercial picture is the opposite. No revenue, no price, no SLA, no supported-hardware list, no customer list, and two generations of the model line with no weights. The Pi API is a real product with real partner deployments in construction, earthmoving, manufacturing, laundry and packaging, and the partner-reported operating metrics are genuinely encouraging, but they belong to the partners. The closest analogue in the space, Skild AI, has published $100 million in ARR, more than 60 paying customers and about $50 million in recognized revenue. Physical Intelligence has published none of these. If Physical Intelligence is the better research lab, Skild AI currently has the better business, and the gap between a $7.5 billion and an $11.19 billion valuation and a $14 billion valuation does not close on research alone.

What would change this assessment. A published rate card with a stated billing unit for the Pi API, together with the named model version it serves. A named model version at open weights, which would restore external verifiability. A third-party reproduction of the pi0.7 generalization results, or at minimum a standard benchmark that pi0.7 is evaluated on by someone other than Physical Intelligence. A confirmed Series C with a named lead, which would settle the valuation record. And a competitive result in a setting Physical Intelligence chooses, after the Benjie's Olympics result.

References

1. Physical Intelligence, "pi0: A Vision-Language-Action Flow Model for General Robot Control", arXiv:2410.24164, October 2024, revised January 2026, published in Robotics: Science and Systems 2025.
2. Physical Intelligence, "pi0.5: a Vision-Language-Action Model with Open-World Generalization", arXiv:2504.16054, April 2025.
3. Pertsch, Stachowicz, Ichter, Driess, Nair, Vuong, Mees, Finn and Levine, "FAST: Efficient Action Tokenization for Vision-Language-Action Models", arXiv:2501.09747, January 2025.
4. Physical Intelligence, "pi*0.6: a VLA that Learns from Experience", arXiv:2511.14759 and company blog post, November 17, 2025.
5. Physical Intelligence, "pi0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities", company technical report, April 16, 2026.
6. Physical Intelligence, "VLAs with Long and Short-Term Memory", company blog post, March 3, 2026.
7. Physical Intelligence, "Precise Manipulation with Efficient Online RL", company blog post, March 19, 2026.
8. Physical Intelligence, "The Physical Intelligence Layer", company blog post with partner-written sections by Weave and Ultra, February 24, 2026.
9. Physical Intelligence, "Frontier Models for Every Robot", Pi API beta announcement, October 13, 2026.
10. Physical Intelligence, "Open Sourcing pi0", company blog post and openpi repository documentation, February 4, 2025.
11. Forge Global, Physical Intelligence pre-IPO valuation and funding round certificate records for the March 2024 seed, November 2024 Series A and A-1, November 2025 Series B and May 2026 Series C.
12. Revenue Memo, "Physical Intelligence funding: Every round, investor, and valuation", 2026.
13. FundBat, "Daily Funding Roundup: July 1, 2026" and the Physical Intelligence company page.
14. Bloomberg reporting of a roughly $1 billion raise at above $11 billion, carried by TechCrunch, Tech in Asia and others, March 27, 2026.
15. The Information, reported by TechCrunch, on Anthropic and Physical Intelligence holding acquisition talks in spring 2026 and on Karol Hausman's denial, July 21, 2026.
16. Beijing Economic Daily, Beijing News and other Chinese coverage on Xingdong Jiyuan taking three global firsts at Benjie's Humanoid Olympic Games and breaking Physical Intelligence's records, April 10 to 11, 2026.
17. Skild AI, "The Hidden Pillar of Robotics", company blog post on crossing $100 million ARR, September 10, 2026.
18. EmbodiedWire, "Robot foundation models compared: GR00T, Gemini Robotics, pi, Helix and the rest", September 28, 2026.
19. American Society of Mechanical Engineers, Benjie's Humanoid Olympic Games competition rules and task descriptions.

Disclaimer: This report is compiled from public sources including company papers, official blog posts, corporate filing records, partner-written case studies, competition results, and media reporting. Figures cited under DISCREPANCIES are recorded as separate observations and are not merged into single values. This report is for reference only and does not constitute investment advice.