Physicl Seed

Physicl Deep Dive: A Physical-AI Synthetic-Training-Data Infrastructure Spun from E-commerce 3D Rendering, Closing the Sim-to-Real Gap with Physics-Accurate Edge-Case Data, on a $10M Inovia-led Seed

Physicl, founded in 2026 by nfinite's Alex de Vigan, is physics-accurate synthetic-data infrastructure for physical AI; Studio generates edge-case 3D scenes with per-frame ground truth, on a $10M Inovia-led seed, with customers including SpAItial, World Labs and Meta.

Executive Summary

Physicl is a physical-AI synthetic-training-data company founded in 2026 by serial entrepreneur Alex de Vigan, who previously founded nfinite in 2017, a 3D product-visualization company for e-commerce and retail that raised a $100M Series B led by Insight Partners and over $120M in total. Physicl's core thesis targets the long tail of physical AI: language and image models benefited from internet text and billions of captioned photos, but robots, autonomous systems and world models that must act in real space have no comparable corpus, and the rare edge cases where they fail (a cup half hidden by a bag, a shadow crossing a counter at sunset, a slightly misstacked box) are almost impossible to capture, replay or vary in the real world. Physicl's answer is that physical worlds cannot be scraped; they must be built from known geometry and physics, so every label is exact and every condition can be changed on purpose. Its platform lets researchers describe the scene a model gets wrong, then generate thousands of variations of it, each carrying per-frame ground truth (depth, surface normals, segmentation and instance IDs, camera pose, 2D and 3D boxes) derived from the geometry itself. The first product, Physicl Studio, entered private beta on 6 November 2026 alongside a $10M seed round led by Inovia Capital with US Venture Partners, and ships with 10,000+ 3D environments and 100,000+ placeable assets built from real geometry. The company has disclosed work with SpAItial, World Labs, Meta and Epic Games; SpAItial, a world-model company founded by TUM professor Matthias Niessner, reported cutting its 3D reconstruction error on real rooms by more than half after training on Physicl data, in an evaluation led by Niessner's team. Physicl positions itself as the data-infrastructure layer for physical AI rather than a dataset vendor, and envisions an agentic workflow that turns massive image or LiDAR capture into high-fidelity 3D and then generates training data on the fly as a model's training state demands, drawing the comparison to "the Scale AI of physical AI." This report synthesizes public information on Physicl's technology, go-to-market, capital structure and the several conflicting public claims around founding year, headquarters, total funding and customers.

Company Overview

Physicl was founded and is led by Alex de Vigan as chief executive. Before Physicl, de Vigan practised mergers-and-acquisitions law in Paris, founded the property-acquisition platform Matchimmo in 2014, and around 2017 founded Hubstairs and then nfinite, which grew from real-estate and home-retail visualization into a 3D product-visualization and visual-intelligence platform for global retailers and brands, raising a $15M Series A and a $100M Series B in 2022. As transformer-based generative AI arrived, nfinite tilted toward generative imagery, and robotics labs, world-model labs and vision teams began asking de Vigan's team for something different: not prettier pictures, but exact data on the situations their models failed in, in thousands of copies. That request became Physicl.

The founding year carries conflicting public accounts. Multiple company databases (SVTR, Clera) label Physicl "Founded 2026" and describe it as roughly six months old as a standalone entity as of mid-2026, with a first public appearance at NVIDIA GTC in March 2026. But Physicl is best understood as a spin-out of nfinite, inheriting the high-fidelity 3D rendering pipeline de Vigan built over a decade. This report records that Physicl launched as an independent brand or entity in 2026 (GTC 2026 debut; seed round and Studio beta announced 6 November 2026), with technology derived from nfinite, and notes that the precise legal entity (likely a French SAS or a US Delaware corporation) and its registry details are not yet public and should be confirmed against company disclosure and national registries.

On headquarters and team, Physicl is explicitly Paris-based with teams in the San Francisco Bay Area and New York (some sources list only San Francisco, reflecting the US expansion focus); the team is roughly 30 people, heavily weighted toward engineering and research, a small and engineering-heavy structure. The company has not disclosed valuation, pricing, revenue or recurring revenue, and remains very early stage.

Product and Technology

The product core is Physicl Studio, a synthetic-data-generation application for physical-AI teams, currently in private beta and free during beta, opened in waves to research labs, universities and companies. Studio's differentiation comes in three parts.

First, ground truth is constructed, not estimated. Depth (metres, EXR float), surface normals (camera space, EXR or RGB-encoded PNG), material index, instance and 2D/3D boxes, and camera pose are all emitted by the renderer from the scene geometry, not estimated from the picture afterwards. That keeps labels exact and stable in the dark corners, occlusions and clutter where models fail, the central advantage of synthetic data over hand-labeled captures.

Second, change one condition at a time. On the same scene a user can vary, alone or in combination, the camera (orbit, pitch, egocentric head-mounted, crane, lateral track), layout (rotated desks, aisle obstacles, restocked shelving, moved clutter) and light (midday, late afternoon, golden hour, blue hour, dusk, night, overcast), leaving everything else fixed so two runs compare directly and one rare edge case is systematically covered. A 480-frame aisle run produces roughly 4.0 hours and 14.2 GB of output.

Third, a scaled real-3D library to start from. Studio ships 10,000+ environments built from real geometry with known materials and 100,000+ placeable assets, so users vary real 3D from day one rather than modeling from scratch. A single run delivers per-frame beauty, depth, normals, materials, instances, boxes and uv layers, plus per-frame JSON of camera, 2D and 3D boxes, pose and visibility.

Underneath is an agentic workflow that transforms unstructured, heterogeneous real-world information (massive image collections, hours of LiDAR capture) into high-fidelity 3D representations with correct geometry, materials and physics. In short, Physicl ports a decade of e-commerce 3D rendering into generating, for physical AI, training worlds that are physically consistent, exactly labeled and infinitely variable. Its value anchor is narrowing the sim-to-real gap so that what a model learns in simulation transfers to the real world.

Applications and Target Customers

Physicl addresses four clear user groups: robot-learning teams (training robot policies that must face restocking, lighting change and clutter before real deployment), perception and vision teams (needing exact depth, normals and instance masks without an annotation queue), world-model researchers (needing physically consistent, camera-known variant video), and ML evaluation teams (needing scenes a model has never seen and a repeatable benchmark rerun on every checkpoint).

Among disclosed customers and partners, the most concrete is SpAItial, a world-model company founded by TUM professor Matthias Niessner that turns real rooms into explorable 3D. According to Physicl, SpAItial more than halved its 3D reconstruction error on real rooms after training on Physicl data, in an evaluation led by Niessner's team; this is a company-reported result, not yet an independent third-party benchmark. The company announcement also names Meta, World Labs and Epic Games as collaborators; secondary sources (e.g. Clera) add Meta Reality Labs, Adobe (including a $1.5M order), DeepMind and NVIDIA and call the company "the Scale AI of physical AI," but these finer customer and order claims mostly come from recruiting or third-party aggregators and need verification in diligence (see DISCREPANCIES).

By scenario, Physicl's sweet spot is the long-tail edge case: occluded objects, moving shadows, cluttered counters, unexpectedly placed items, rare in real datasets and expensive to capture, exactly where robots fail. The intended usage is on-the-fly generation: as a model detects a capability gap in training, it triggers automatic generation of the corresponding synthetic scenes, closing a data-training loop.

Business Model and Ecosystem

Physicl's current commercial form is "SaaS plus data infrastructure": Studio is free in beta and opened in waves to labs, universities and companies, with post-beta monetization expected through usage- or seat or project-based billing; neither pricing, license model nor recurring revenue has been disclosed. Its strategic narrative is explicitly a data layer, not a model layer: it does not train its own robot models and does not compete with downstream customers, instead supplying environments, assets and labels for training and post-training on hard cases.

A deeper ambition is a continuous, agentic, on-the-fly synthetic-data workflow: customers feed massive image or LiDAR data into the platform, convert it to high-fidelity 3D, and generate billions of variants on demand; in de Vigan's vision this is "you are running Claude, it checks it needs more data, then auto-generates it," making Physicl the data kitchen embedded in the training loop. This narrative maps onto Scale AI's labeling-infrastructure role in the LLM era and is the main source of the company's valuation imagination.

On ecosystem, Physicl inherits de Vigan's decade of 3D and retail-visualization asset library, customer network and investor relationships (Inovia and US Venture Partners also backed nfinite), and an advisory board of Thomas Wolf (Hugging Face), Christoph Lassner (World Labs), Matthias Niessner (TUM), Yaser Sheikh and Sebastien Deguy, figures at the core of the physical-AI and world-model networks whose endorsement and reach materially help customer acquisition.

Competitive Landscape

Physicl enters a young and fast-crowding field: many start-ups now supply data, tooling or simulation environments to robot companies and foundation-model builders, and on this site Physical Intelligence, Skild AI, World Labs, Veeda AI, Sudo Tech, Qianjue, XIRRA and Qingyan all approach physical AI and world models from different angles. de Vigan himself concedes the space is "very young and sometimes blurry."

Physicl's relative strengths are threefold: the "ground truth by construction" reliability, where labels come from geometry and stay exact in occlusion and dark corners, distinct from post-hoc estimation; the asset scale (10,000+ scenes, 100,000+ assets) and the engineering ability to turn unstructured real information into physically consistent 3D, both honed over a decade; and a strong advisory network and disclosed headline customers (SpAItial, World Labs, Meta) that build early trust. Its weaknesses are equally clear: the company is extremely young (roughly six months to under a year as a standalone entity), about 30 people, with no disclosed revenue or pricing, and a free beta means willingness to pay is unproven, while the sim-to-real benefit rests on a single company-reported SpAItial case. The real moat, data assets, industry know-how and depth of embedding in the training loop, still needs time to compound.

Financing and Capital Structure

Physicl announced a $10M seed round on 6 November 2026, led by Inovia Capital with US Venture Partners, funding team expansion and continued build-out of its synthetic-training-data platform. This is a confirmed round backed by primary sources such as TechCrunch. Notably, Inovia and USVP were also nfinite investors, so the new company inherits a capital base that already knows the team and technology.

The total funding carries a conflicting account. Some third-party aggregators (e.g. Clera) list Physicl's total raised at roughly $20M, but that figure likely conflates nfinite's historical financing (over $120M) or folds in undisclosed early or pre-seed contributions, and cannot be taken as the confirmed financing of Physicl as a standalone entity. This report takes the disclosed $10M seed as the figure of record and flags the $20M account as unverified. The company has not disclosed a valuation; secondary sources (Clera) mention a planned $100M-plus Series A in early-to-mid 2027, which is forward-looking speculation, not company-confirmed.

By comparison, Physicl is substantially behind physical-AI peers that have disclosed hundreds of millions or billions in financing (Physical Intelligence, Skild AI, Sudo Tech), but as a roughly one-year-old seed-stage company a $10M launch with a reputable lead investor is a solid start. The size, valuation and timing of the next round, and the conversion from free beta to paid, will determine its platform and scale.

DISCREPANCIES

1. Founding year and entity nature. Account A: multiple databases label Physicl "Founded 2026," roughly six months old as a standalone entity as of mid-2026, with a GTC 2026 debut. Account B: Physicl is described as a spin-out of nfinite, with technology and team derived from de Vigan's decade at nfinite since 2017. Ruling: record Physicl as launched as an independent brand or entity in 2026, technology derived from nfinite; the precise legal entity (French SAS or US Inc) and registry are not public, marked unconfirmed. Source grade: DATABASE (SVTR/Clera) versus MEDIA.

2. Headquarters. Account A: the company and most coverage place the base in Paris with teams in the San Francisco Bay Area and New York. Account B: some recruiting or database sources list only San Francisco (e.g. Clera says "based in San Francisco"). Ruling: record headquarters as Paris with San Francisco and New York teams, and flag the single-source San Francisco account as inconsistent with other sources. Source grade: OFFICIAL (announcement) versus DATABASE.

3. Total funding amount. Account A: confirmed $10M seed on 6 November 2026 (Inovia lead, USVP participate; primary sources TechCrunch et al.). Account B: third-party aggregators list total raised at roughly $20M. Ruling: take the confirmed $10M seed as the figure of record and flag the $20M account as likely conflating nfinite history and unverified. Source grade: OFFICIAL / MEDIA versus DATABASE.

4. Customer and order details. Account A: the company announcement discloses collaboration with SpAItial, World Labs, Meta and Epic Games, with SpAItial's error cut by more than half (Niessner team evaluation). Account B: secondary sources add Meta Reality Labs, Adobe (including a $1.5M order), DeepMind and NVIDIA. Ruling: cite the four announced collaborations and the SpAItial result, and flag the Adobe order and DeepMind or NVIDIA details as third-party-aggregator claims not confirmed by the company. Source grade: OFFICIAL versus DATABASE.

5. Founder name spelling. Account A: company materials and most coverage use "Alex de Vigan." Account B: a founder profile (YesPress) uses "Alexandre de Vigan." Ruling: record "Alex de Vigan (also Alexandre de Vigan)" and note the two are the same person with a full-name spelling difference. Source grade: OFFICIAL versus MEDIA.

6. Team size. Account A: the announcement and StartupBooted give roughly 30 people with a high engineering share. Account B: a database source (Clera) gives a range of 11 to 50. Ruling: record roughly 30 (the intersection of most sources) and flag the single-source wide range as imprecise. Source grade: MEDIA versus DATABASE.

7. Nature of the SpAItial result. Account A: Physicl says SpAItial cut its real-room reconstruction error by more than half after using its data, evaluated by Niessner's TUM team. Account B: that result is company-reported, not an independent third-party benchmark, and SpAItial itself has product progress such as Echo-3. Ruling: record the result as an early company-disclosed signal and flag it as non-independent and needing more customer validation. Source grade: MEDIA, unverified.

Risks and Diligence Gaps

The primary gap is independent validation of the sim-to-real benefit. Physicl's strongest evidence is SpAItial's "error cut by more than half," but it is company-reported and evaluated by a collaborator (Niessner's team), with no independent third-party benchmark against a baseline on a standardized task, and only one customer, so generalization to other physical-AI tasks such as robot manipulation or autonomous driving remains unproven.

The second gap is commercialization and willingness to pay. Studio is in a free beta, with no disclosed pricing, license model or recurring revenue; the conversion from a free research tool to a paid data infrastructure is the most common failure point for early SaaS, and there is no public evidence of willingness to pay.

The third gap is capital thickness and certainty. Although the $10M seed provides launch funding and a planned $100M-plus Series A is only speculation, as a roughly 30-person, no-revenue early company it depends heavily on the next round's timing and cash flow; if the Series A misses or the funding environment tightens, platform build-out will be constrained.

The fourth gap is the competitive moat. The physical-AI data and simulation field is young and crowded, and "physically consistent 3D synthesis" at the method level is not a durable moat; what defends is data-asset scale, industry know-how and training-loop embedding. Physicl has a starting asset library and strong advisors, but embedding depth and customer lock-in still need time.

The fifth gap is information transparency. As a private company under a year old, Physicl has disclosed no valuation, revenue, pricing or specific legal entity, and most details (customer orders, total funding, headquarters) come from a mix of announcements, media and third-party aggregators with several conflicting accounts (see DISCREPANCIES), and should be confirmed against formal company disclosure and national registries.

The sixth gap is talent and expansion risk. The company is small and engineering-heavy, and its core 3D-rendering and physics-pipeline capability depends heavily on de Vigan's nfinite accumulation and key engineering talent; as a roughly one-year-old company, retention of core talent and equity incentives, plus cross-time-zone (Paris and San Francisco) collaboration efficiency, are potential constraints.

Overall Assessment

  • techScore 8.0. Physicl's "ground truth by construction" base is reliable, its starting asset library is substantial (10,000+ scenes, 100,000+ assets), and it has a visible customer result in SpAItial, giving clear differentiation in physically-accurate synthetic data. Deducted points: the sim-to-real benefit rests on a single company-reported case, no independent benchmark, cross-task generalization unproven, company extremely young.
  • bizScore 5.5. A free beta and a wave-based opening to labs, universities and companies is a clear strategy, and the data-infrastructure-layer positioning is sensible; but no pricing, license model or revenue is disclosed and willingness to pay is entirely unproven. Deducted points: no revenue, model undisclosed, free-to-paid conversion risk.
  • fundScore 6.0. The $10M seed led by Inovia with USVP, inheriting nfinite's investor relationships, is a solid start; but the amount is modest among physical-AI peers, the $20M total account is unverified, no valuation is disclosed, and the next Series A is only planned. Deducted points: early-stage capital base, conflicting total-funding account, limited certainty.
  • productScore 6.5. Studio is shipped and in beta, with a complete "describe scene, vary on multiple axes, per-frame ground truth" loop and a real-3D starting library, ahead of a pure concept stage. Deducted points: still beta, free, paid path unverified, limited cross-industry generalization evidence.
  • marketScore 7.0. Physical-AI data infrastructure is a high-heat, high-need direction, the company has a strong advisory network and headline-customer endorsement, and the "Scale AI of physical AI" narrative carries imagination; but the field is crowded, most competitors (PI, Skild, World Labs, Veeda) are more mature or better funded, and the company is extremely young. Deducted points: under one year old, intense competition, limited scale and proven metrics.

The composite view is 6.6. Physicl is a roughly one-year-old physical-AI synthetic-data infrastructure company spun from a mature 3D-rendering asset base by a serial entrepreneur, with a clear technical differentiation of "physically accurate, infinitely variable" data and visible early customer signals, while carrying the standard early-stage deficits of unindependently-validated benefit, unclear willingness to pay, early-stage capital and low information transparency. The most informative next disclosures will be independent benchmarks from customers beyond SpAItial, the first paid contract after beta, and the actual landing of the 2027 Series A.

References

  1. Physicl official site physicl.ai and Studio pages (2026): positioning, Studio capabilities, output layers and per-frame ground truth.
  2. Physicl official post "Introducing Physicl" (2026-11-06): founding motivation, $10M seed (Inovia lead, USVP), Studio launch, advisory board.
  3. TechCrunch / StartupBooted / KeeVurds (2026-11): seed round and Studio beta coverage, team about 30, Paris and San Francisco and New York.
  4. YesPress founder profile (2026): Alex de Vigan bio (M&A lawyer, Matchimmo, nfinite 2017, Physicl 2026) and nfinite history.
  5. EdgeChat / nfinite materials (2026): nfinite founded 2017, $100M Series B led by Insight Partners in 2022, over $120M total raised.
  6. SVTR company profile (2026-11-07): Physicl founded 2026, GTC 2026 debut, $10M seed, about 30 staff, Paris.
  7. Clera company page (2026): headquarters and team and funding accounts (including $20M total raised, Adobe $1.5M order), partly unverified.
  8. SpAItial and Physicl collaboration disclosure (2026): reconstruction error on real rooms cut by more than half after training on Physicl data (TUM Niessner team evaluation).
  9. Physicl advisory-board public information (2026): Thomas Wolf (Hugging Face), Christoph Lassner (World Labs), Matthias Niessner (TUM), Yaser Sheikh, Sebastien Deguy.
  10. Industry background: the sim-to-real gap in physical AI and the role of synthetic data in robotics and world-model training (2026).
This report is compiled from public sources and does not constitute investment advice. Data is current as of 8 November 2026. Physicl is a private, unlisted company with no audited financials; the seed amount, customer collaborations, the SpAItial error result, team size, total funding and headquarters carry several conflicting public accounts (see DISCREPANCIES), and uncertainties are flagged throughout the text and in DISCREPANCIES; please rely on the company's announcements, national registries and primary sources.

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.