Where the money goes in physical AI

The private mega-rounds (2024-2026)

Between 2024 and 2026, a handful of Western robotics and physical AI startups raised private rounds of a size unprecedented for the sector, some rivalling the largest venture rounds in history across any industry. The table below gathers the most significant deals, with the amount raised, the resulting valuation, the investor who led the round, the date and the source.

For comparison, the 2025 sector total for humanoids alone is estimated between 2.3 B$ (CB Insights, via Sifted, humanoids only) and 14 B$ (PitchBook, robotics broadly defined, all segments), depending on scope: the methodologies are not comparable, but even at the widest end of the range, a handful of rounds is enough to match an entire year of the sector.

Who is funding: investor logic

Three families of investors fund physical AI, each with a distinct logic that shows up directly in the rounds from the table above. Traditional venture capital (Sequoia, Lux Capital, Thrive Capital, CapitalG, Founders Fund) is betting on the models/brain layer, the most technically uncertain but also the most valuable one if it succeeds: that is what explains soaring valuations at Figure AI, Physical Intelligence and Skild AI. Strategic corporates (NVIDIA, Amazon/Bezos Expeditions, Samsung, LG, Google, Qualcomm, Microsoft) are buying an upstream option on compute and components: NVIDIA took part in dozens of robotics VC deals in 2025 (67 per TechCrunch and PitchBook, 48 per aifundingtracker, an unresolved discrepancy) while locking in the onboard inference layer. Sovereign funds and mega-funds (SoftBank, PIF/Alat, QIA, Mubadala, Temasek) deploy patient capital and are increasingly looking to buy already-installed distribution capacity directly, rather than a simple equity ticket in an early-stage startup.

The clearest signal of this shift is SoftBank's acquisition of ABB's industrial robotics division for 5.375 B$ in October 2025: rather than betting on a new humanoid champion still in the research phase, SoftBank bought an installed base of industrial robots already deployed at thousands of customers, explicitly under the "Physical AI" strategy named by Masayoshi Son. The same fund also bought out full control of Boston Dynamics from Hyundai (325 M$ for the remaining 9.65%, finalisation expected June 2026) and chipmaker Ampere Computing (6.5 B$, March 2025): three deals that trace a logic of vertical control rather than a simple venture-capital bet.

Source: Reuters, 8 October 2025 (ABB acquisition); ByteBridge (SoftBank Vision Fund portfolio).

Public plans, bloc by bloc

Every major bloc waves its own headline figure for physical AI, with scopes, horizons and calculation methods that are not comparable to one another. Here are the plans most commonly cited, with their announced amount and horizon.

The key asymmetry: China outside VC, the West inside it

The most structuring contrast in this dossier fits in one sentence: in the West, the funding layer that builds a champion (Figure AI, Physical Intelligence, Skild AI, Neura Robotics, Apptronik, Wayve) is private venture capital and strategic corporates, with individual rounds from 300 M$ to over 1 B$ and valuations that soar (Figure AI 39 B$, Project Prometheus 41 B$, Skild AI more than tripling in a year). In China, that role is played by public vehicles: national and local funds that reach over 100 B yuan at the scale of a single city (Hangzhou), ministerial plans quantified in physical targets (100,000 units deployed targeted by 2027), and stock market exits at far more conservative valuations (AgiBot 5.1 to 6.4 B$, Unitree 5.9 to 7 B$) despite a higher production volume than Western peers.

Private venture capital does exist in China (HongShan, formerly Sequoia China; Hillhouse; Tencent), but per-company tickets remain an order of magnitude below Western ones: AgiBot had raised only around 84 M$ in private funding through mid-2025, Fourier Intelligence around 83 to 100 M$. The gap is not offset by more venture capital, but by the scale of state support upstream and downstream: direct subsidies, guarantee funds, public deployment procurement, five-year plans quantified in production targets.

A direct consequence for reading the numbers: comparing a Chinese funding round to an American one systematically underestimates the real Chinese effort, which mostly sits in public budget lines outside venture capital databases like Crunchbase or PitchBook, and symmetrically overestimates the risk that Chinese champions are underfunded, since their solidity does not depend on a fresh round of private fundraising but on a near-unlimited public backstop.

The hourglass: extreme capital concentration

The physical AI funding landscape looks like an hourglass: very wide at the top, nearly empty one step down. A small number of companies concentrate most of the world's venture capital in the sector. Project Prometheus's closing round alone, 12 B$ in June 2026, exceeds CB Insights's entire estimate for global humanoid funding across all of 2025 (2.3 B$ projected); add the rounds from Figure AI, Skild AI, Neura Robotics and Wayve, and the top five deals of 2024-2026 combined comfortably exceed 25 B$, well above Dealroom's estimate of the global humanoid market in 2025 (3.2 B$, itself described as bigger than the previous six years combined).

Below that peak, the fabric thins out fast: most robotics startups, including Chinese players that actually lead on production volume such as AgiBot (around 84 M$ cumulative) or Fourier Intelligence (around 83 to 100 M$), raise amounts an order of magnitude below the Western mega-rounds. For an industrial SME or an integrator looking to fund a pilot rather than become a global champion, the consequence is direct: the relevant funding mechanisms are not the ones making headlines for Figure AI or Project Prometheus.

See our robotics value chain pillar to see where defensible value sits outside the mega-rounds, and our funding your project guide for the schemes actually accessible to a European SME or industrial company.

The Atlas verdict

First observation: the money is rational about its destination, flowing to the layers where value and defensibility look strongest, robotic foundation models for Western venture capital, installed industrial deployment capacity for sovereign funds, manufacturing scale-up for the Chinese state. Second observation: the money is optimistic about timing, valuations of 39 or 41 B$ assume a trajectory toward a market worth hundreds of billions of dollars that even the most bullish analysts (Morgan Stanley, Goldman Sachs) do not project before 2035, and more often not before 2050. Third observation, the most awkward one: a large share of this Western capital sits on the least defensible layer in the medium term, the generalist models, already locked in by three or four American players backed by OpenAI, NVIDIA and SoftBank, while the most accessible residual value likely sits in sector integration, real-world deployment data and regulatory compliance. Fourth observation: this reading stays fragile, because the dossier feeding it contains its own unresolved discrepancies, on NVIDIA's number of VC deals in 2025 (67 per TechCrunch and PitchBook versus 48 per aifundingtracker), on Saudi PIF's assets under management (from 900 B$ to 1.15 T$ depending on the source and year) or on the exact scope of the US CHIPS Act (52.7 B$ in direct grants versus 280 B$ for the total package, nested rather than contradictory scopes). Fifth observation: the window stays open for a European player, not by trying to outspend head-on, but by using the two years bought by the AI Act's Annex I deadline being pushed to 2 August 2028 to build a compliance and integration offer that neither China, deployment first and standardisation later, nor the early American entrants, an unregulated market, have prioritised so far.