1. Introduction
For the past several years, the most important question in technology investing has been: who controls the intelligence layer? The answer has largely been a software story — large language models, inference infrastructure, AI applications. That story is not over. But a new chapter is beginning.
At CES 2026 in January, Nvidia CEO Jensen Huang declared that "the ChatGPT moment for robotics is here." The claim is ambitious, but the capital data supports it. Robotics and physical AI startups raised a record $27.6 billion across 1,009 deals in 2025, more than doubling the prior year, according to PitchBook. [1] In Q1 2026 alone, physical AI and robotics featured among the top categories in what Crunchbase recorded as the largest single quarter of global venture investment in history — $300 billion across 6,000 startups worldwide. [2]
Physical AI — the class of systems that combine artificial intelligence with the ability to perceive, navigate, and act in the physical world — is no longer a research story. It is a capital formation story. And for frontier allocators, the question is the same one that defined the early innings of the cloud, the smartphone, and the AI software wave: which layer of the stack captures the most durable value, and who gets there first? [3]
2. The Capital Story
The numbers that define this moment are striking in both scale and trajectory.
Global robotics and physical AI VC funding reached $27.6 billion in 2025, up from approximately $13.8 billion in 2024 and exceeding the prior peak of $13.1 billion set during the 2021 boom. [4] The composition of that capital has changed as much as its scale. Deal count has actually fallen — from 671 rounds in 2023 to 473 in 2024 — while deal value has surged. Fewer companies are being funded, but the ones that are raising are doing so at a pace and valuation that would have been extraordinary even two years ago. [5]
The most visible single data point is Figure AI's $39 billion valuation following its Series C round in September 2025, which exceeded $1 billion in committed capital — a 15x increase from its $2.6 billion valuation in February 2024. [6] More telling, perhaps, is Skild AI: a company founded in 2023 that raised $1.4 billion in January 2026 at a $14 billion valuation — more than tripling its value in seven months — with SoftBank and Nvidia co-leading the round and Sequoia, Lightspeed, Jeff Bezos, Samsung, LG, and Salesforce Ventures among the participants. [7] These are not niche bets. They are major institutional statements about where the next layer of AI infrastructure is being built.
The long-range market projections reinforce the investment logic. Morgan Stanley projects over one billion humanoid robots in operation by 2050, generating a $5 trillion market — with 90% of that demand coming from industrial and commercial applications. MarketsandMarkets estimates the humanoid robot market alone grows from $1.8 billion in 2023 to $13.8 billion by 2028, at a compound annual growth rate exceeding 50%. [8][9]
3. Two Roads: How the Physical AI Market Is Splitting
The physical AI investment landscape is splitting along a fundamental design question: should a robot look and move like a human — fitting into environments already built for people — or should it be purpose-built from scratch to do one thing extremely well?
These two paths have different investment profiles, different infrastructure requirements, and different time horizons to commercial return.
The first road is humanoid and general-purpose robotics. The bet here is that a human-shaped machine, operating in spaces already designed for humans, offers the broadest addressable market with the least environmental redesign required. Figure AI, Agility Robotics, 1X, and Tesla's Optimus programme are all pursuing versions of this thesis. The near-term commercial deployment is not the home — it is the factory floor. Figure's Figure 02 supported over 30,000 vehicles at BMW's Spartanburg facility, handling more than 90,000 parts across 1,250+ hours of real production. [10] Agility Robotics signed a Robots-as-a-Service agreement with Toyota Motor Manufacturing Canada in February 2026, with seven or more commercial Digit units now active supporting RAV4 material handling. [11] Boston Dynamics unveiled its production-ready electric Atlas in January 2026; all 2026 units are committed to Hyundai and Google DeepMind, with full factory deployment targeted for 2028. [12]
The consumer home use case is real but longer-duration. 1X has opened preorders for its NEO home robot at $20,000. Tesla targets sub-$30,000 Optimus pricing at scale, though as of Q1 2026 the programme remains in an R&D and learning phase with no robots performing productive tasks in Tesla factories. [13] Chinese manufacturer Unitree has shipped over 5,500 humanoids in 2025 and is targeting 10,000 to 20,000 units in 2026, with its R1 model priced at $5,900 — creating significant pricing pressure on Western competitors. [14]
The second road is purpose-built, specialized robotics. These systems are not trying to replicate human form. They are optimised for a specific task in a specific environment — and as a result they are generally closer to commercial deployment, more predictable to underwrite, and already at scale in several sectors. Industrial robotics incumbents ABB, FANUC, KUKA, and Universal Robots represent the established base: 542,000 industrial robots were installed worldwide in 2024, and the global operational stock has reached 4.664 million units. [15] The startup layer in this path includes Gecko Robotics (infrastructure inspection), Saronic (maritime autonomous vessels), and a wave of warehouse automation companies — though logistics and warehousing robotics deal value actually fell 28.5% in 2025 as defense robotics surged, suggesting capital is rotating toward higher-urgency applications. [1]
The infrastructure supporting each path differs significantly. Humanoid robotics bottlenecks are primarily about making the machine more capable: dexterity, spatial reasoning, battery density, and real-world AI training data. Purpose-built industrial robotics bottlenecks are about scale and integration: systems software, deployment expertise, supply chain buildout, and embedding into existing operations. Investors applying the same diligence framework to both will miss what makes each category distinctive.
Picture 2. Two Roads in Physical AI | Humanoid and general-purpose robots are targeting broad commercial deployment by 2026–2030; purpose-built industrial systems are already deployed at scale today. The two paths have different bottlenecks, different investor profiles, and different time horizons. Sources: [6][10][11][12][13][14][15]
4. The Third Layer: Foundation Models for Robotics
Sitting above both hardware paths is an investment thesis that may ultimately capture more durable value than either: the software intelligence layer.
The analogy to the broader AI market is instructive. In software AI, the most valuable positions have accrued not just to the companies building the applications, but to those controlling the underlying models. The same dynamic is emerging in physical AI, and the companies competing for the foundation model layer in robotics are attracting capital at a pace that suggests investors see it the same way.
Skild AI ($14B valuation, $2B+ raised) is building what it calls the "omni-bodied brain" — a single AI foundation model designed to control any robot for any task, without bespoke retraining for each new hardware platform or environment. The round's investor base — SoftBank, Nvidia, Bezos, Sequoia, Samsung, LG, Salesforce — reflects the breadth of strategic interest in whoever solves this problem. [7] SoftBank simultaneously acquired ABB's robotics hardware division for $5.375 billion (at roughly 2.3x revenue); its investment in Skild at a far higher software multiple signals exactly where it believes recurring value will concentrate. [16]
Physical Intelligence ($5.6B valuation) is building generalised AI for robot control, competing in the same foundation model space. [17] Jeff Bezos' Project Prometheus, launched in November 2025 with $6.2 billion in initial funding, focuses explicitly on physical AI for engineering-intensive industries including chip manufacturing, aerospace, and automotive — with a reported additional $10 billion raise in progress. [18]
The strategic logic is compelling for investors: if one software layer can run any robot, the companies owning that layer collect recurring licensing revenue across every deployment, regardless of who manufactured the hardware. It is a picks-and-shovels thesis for a market where hardware is commoditising faster than most expected. Boston Dynamics formalised this dynamic in January 2026, announcing a strategic partnership with Google DeepMind — whose Gemini Robotics models will underpin Atlas's AI capabilities — signalling that even the most advanced hardware companies now see the intelligence layer as the critical dependency. [12]
Picture 3. Key Physical AI Startups | The five most capitalised venture-backed physical AI companies, spanning humanoid hardware, purpose-built industrial systems, and the foundation model layer — with valuations ranging from ~$10B to $39B. Sources: [4][6][7][12][13][17][18]
5. Where AI Meets the Physical World: The Convergence Opportunity
The most important investment insight in physical AI is not about any single company or hardware path. It is about what happens when intelligence becomes infrastructure.
Physical AI is not replacing the digital AI story — it is extending it into new substrates. The same AI capabilities that transformed software are now being applied to manufacturing, logistics, defense, agriculture, construction, healthcare, and the built environment. McKinsey has named embodied intelligence a defining technology trend of 2026. PitchBook analysts identify the convergence of vision-language-action models, humanoid robotics, and world models in automotive manufacturing as one of the highest-priority investment themes of the year. [3][19]
The convergence creates a new kind of infrastructure company — one that is harder to describe in a single sentence but also harder to displace once embedded in real operations. Several patterns are worth noting for allocators:
Picture 5. Physical AI Convergence Map | Intelligence is converging simultaneously with manufacturing, defense, logistics, healthcare, agriculture, construction, and space — creating a new class of infrastructure companies that are harder to displace once embedded in real operations. Sources: [1][3][19][20]
AI and manufacturing are converging faster than most traditional industrial companies anticipated. Western original equipment manufacturers now face a Chinese cost edge and shorter development cycles in robotics. The companies that close this gap — through AI-native manufacturing software, vision-language-action models, or proprietary training data — will disproportionately capture procurement budgets from automotive, aerospace, and semiconductor manufacturers.
AI and defense have already converged, as covered in our previous issue. Defense robotics funding surged 139% in 2025. The same autonomous systems, sensing platforms, and AI control layers that serve commercial customers serve sovereign ones — and sovereign customers offer larger contracts, more durable demand, and a strategic moat that commercial customers rarely provide. [20]
AI and logistics is the most established application, but capital is rotating. Logistics and warehousing robotics deal value fell 28.5% in 2025 while defense robotics surged. This does not mean logistics is uninteresting — it means the easy first wave has been funded, and the more differentiated second wave is still forming. [1]
The common thread across all of these is that the most valuable physical AI companies will not be pure hardware or pure software businesses. They will be companies that own a closed loop: proprietary training data from real-world deployments, AI models that improve with each deployment, and hardware or software platforms embedded deeply enough in customer operations that switching becomes structurally difficult.
Picture 4. The Physical AI Architecture | Three layers are forming: hardware at the base, foundation models and intelligence in the middle, and sector applications at the top. The middle layer — where Skild Brain, Physical Intelligence, and Gemini Robotics compete — is where the most durable value is likely to concentrate. Sources: [7][12][16][17][18]
6. What Remains Accessible for Frontier Allocators
The headline valuations have moved. Figure AI at $39 billion, Skild AI at $14 billion, and Boston Dynamics at an implied $21–28 billion are not early-stage entry points for most VC and family office capital. But the physical AI investment landscape is large enough that meaningful early-stage opportunities remain across several dimensions.
The foundation model layer is still forming. While Skild and Physical Intelligence have attracted significant capital, the question of which foundation model will emerge as the dominant platform for robotics is genuinely unsettled. Smaller, earlier-stage foundation model companies — particularly those focused on specific verticals such as surgical robotics, agricultural automation, or construction — represent earlier entry points into the same thesis.
Purpose-built industrial robotics remains undervalued relative to humanoids. The humanoid narrative has attracted a disproportionate share of investor attention and capital, while purpose-built systems with clearer near-term commercial traction — autonomous underwater vehicles, inspection robotics, agricultural automation — remain less crowded from an investor standpoint.
The picks-and-shovels layer is the most accessible. The companies building the enabling infrastructure — actuators and motion systems, training data platforms, simulation environments, edge compute for real-time robot inference, and systems integration software — benefit from the sector's growth regardless of which hardware platform wins. Manufacturing costs for humanoid robots are currently dominated by actuators and motion systems, which account for 40–50% of total unit cost. Companies that solve the actuator cost curve have a large and durable market regardless of which robot form factor prevails. [9]
Europe is an underserved market. Unlike defense tech, where European venture capital has begun to develop real depth, the physical AI investor base in Europe remains thin relative to the market opportunity. The continent's industrial base — automotive, precision manufacturing, logistics — represents some of the highest-density commercial deployment opportunities for physical AI, but the startup and investor ecosystem has not yet scaled to match.
7. Conclusion
The ChatGPT moment that Jensen Huang described at CES is not simply a marketing claim. The capital data, the deployment data, and the corporate strategic positioning all point in the same direction: intelligence is moving from servers into the physical world, and the infrastructure layer for that transition is being built now.
The two roads — humanoid general-purpose robots and purpose-built industrial systems — will both generate returns, but on different timelines and through different mechanisms. The foundation model layer sitting above both may ultimately capture the most durable value, following the same pattern that played out in software AI.
For frontier allocators, the question is not whether physical AI is real. That debate is over. The question is which layer of a rapidly forming stack remains early enough, differentiated enough, and underinvested enough to generate the returns that this audience requires. The picks-and-shovels layer, the European market, and the vertical-specific foundation model opportunity are where that answer is most likely to be found in 2026.
The machines are leaving the lab. The capital is following. The investors with the clearest edge will be those who understand the full architecture — not just the robots that are making headlines, but the intelligence, infrastructure, and enabling technology that make those robots possible.