Robotics VC Breaks Annual Records at Midyear: Why Physical AI Commands Software Multiples
Robotics startups have already exceeded every prior full-year global funding record — and the calendar still shows June. According to data from Dealroom, companies in the physical AI and robotics space have raised $55.8 billion so far in 2026, a figure nearly double what the sector raised in all of 2025 under the same broad measurement. Crunchbase’s narrower robotics-specific dataset captures $18.8 billion over the same period — itself already above every Crunchbase full-year record — and confirmed this week that robotics startup funding has entered record territory before July. The sector that Sand Hill Road once dismissed as a hardware graveyard is now being priced like AI infrastructure.
That repricing did not happen because robots got cheaper to build. It happened because of a specific architectural development that changed what a robotics company actually sells.
The Deals That Defined the Surge
The individual rounds tell a pointed story. Topping the list is Saronic, an Austin-based defense tech startup building autonomous naval vessels, which raised a $1.75 billion Series D in March led by Kleiner Perkins. The four-year-old company now carries roughly $2.6 billion in total funding and a $9.25 billion valuation — more than double where it stood after its previous round.
In January, Skild AI announced a $1.4 billion round led by SoftBank that tripled its valuation to above $14 billion. The Pittsburgh company is building what it calls an “omni-bodied” foundation model: a single AI system designed to operate any robot body for any task. That framing explains why SoftBank led the round and why the valuation reflects software-company multiples rather than robotics-hardware ones. In April, Skild acquired the robotics arm of Zebra Technologies, deepening the stack it controls.
Germany’s Neura Robotics closed a Series C in June worth up to $1.4 billion, backed by Tether, Nvidia, Amazon, Qualcomm, Bosch, Schaeffler, and the European Investment Bank — a syndicate that maps nearly exactly onto the supply chain a humanoid robot company needs. “The future of AI will not only live on screens,” Neura CEO David Reger said in a statement. “It will move, interact, learn and work beside us in the real world.” The full amount is tied to performance milestones; the investor list is not.
Apptronik extended its Series A to over $935 million in February with Google, Mercedes-Benz, John Deere, and AT&T Ventures participating. Mind Robotics, a Rivian spinout, raised two separate rounds — $500 million in March and $400 million in May — bringing its total past $900 million at a $3.4 billion valuation. Beijing-based Shihang Intelligent pulled in $1 billion in a Series A in June. And approximately $6 billion flowed into six or seven world-model companies in the first quarter alone, according to Bessemer Venture Partners — companies building simulation infrastructure that generates synthetic training data for robots at scale.
Why the Money Changed Its Mind
For most of the past decade, venture capital assessed robotics through a hardware lens: slow development cycles, expensive manufacturing, narrow use cases, thin margins. Even during the frothy years of 2021, global robotics funding peaked at roughly $14 billion under Crunchbase’s tracking. The money was patient, not exuberant.
The reframing began when large language model development demonstrated that AI could generalize across domains in ways previously considered impossible. That observation produced a specific question for robotics: if a single model could handle writing, coding, and legal analysis, could a single AI system learn to operate multiple robot bodies in multiple environments?
The answer came through what engineers now call Vision-Language-Action models. A VLA takes in a camera feed and a natural language instruction simultaneously and outputs motor commands — not through separate perception, planning, and control pipelines stitched together, but through a single unified neural network. The reasoning and the acting happen in the same system. A warehouse robot programmed for specific conditions fails when someone moves a bin two inches. A VLA-based system can reason about what it sees and adapt.
The critical implication is architectural. Because a VLA model can be pre-trained across multiple robot embodiments — different arm configurations, sensor suites, and form factors — and then fine-tuned for a specific hardware platform, the robot brain becomes a software product subject to platform network effects rather than a hardware product subject to manufacturing constraints. This is what Skild AI’s $14 billion valuation on essentially a software company reflects: it is being priced not as a robotics original equipment manufacturer but as AI infrastructure, the same category that commands the highest multiples in technology. The architecture also repurposes the same infrastructure that makes language models generalize across topics — a motor command trajectory is generated much the way text is, through a token prediction process applied to a learned representation of physical action.
Nvidia made this positioning explicit. At GTC Taipei on June 1, Jensen Huang announced the Isaac GR00T Reference Humanoid Robot — a fully open platform pairing a Unitree H2 Plus chassis with Sharpa Wave tactile five-finger hands and Nvidia’s Jetson AGX Thor compute module, capable of 2,070 FP4 teraflops of AI performance. The platform runs the Isaac GR00T software stack for simulation, training, and deployment, and Nvidia intends it to function like Android for robotics — the common software layer across which an ecosystem of hardware partners competes. The robot is expected to ship from Unitree in late 2026.
Three Bets, Not One
The funding surge is not a single thesis. A closer look at where capital is flowing reveals three distinct layers of the stack being bet on simultaneously.
The first is the AI foundation model for robotics — companies like Skild AI building general-purpose robot brains. These are valued like AI software companies and have attracted the largest individual rounds. The world-model category, which generates synthetic training data to solve the physical AI data scarcity problem, sits within this layer. Unlike large language models, which trained on the entire text of the internet, physical AI requires recordings of how physical matter actually behaves — data that does not exist online in usable form. World models offer a path to generating it synthetically at scale, which is why they have become the most capital-intensive sub-category.
The second layer is the integrated robot platform — companies like Neura Robotics, Apptronik, and China’s AgiBot building full-stack systems designed for specific deployment environments. These attract strategic investment from the industries they aim to serve, which explains why Neura’s syndicate included both Nvidia (for compute) and Bosch and Schaeffler (for industrial integration).
The third layer is specialized applications — defense and maritime (Saronic), surgical systems, agricultural automation — where use cases are clear, customers have budget, and competitive moats come from domain expertise rather than platform ambition.
China’s Accelerating Position
The funding surge is not a purely American phenomenon. China-based robotics companies raised $5.6 billion across 176 deals through mid-May alone, according to Crunchbase, a sum matching total Chinese robotics investment in all of 2021. China now accounts for more than 43% of global robotics venture investment, per Crunchbase data, and 2026 appears to be the year Chinese companies are pivoting from early funding rounds to mass production and public listings. Unitree Robotics, China’s most prominent humanoid maker, filed for a Shanghai STAR Market IPO in March targeting a $3 billion to $7 billion valuation and cleared its listing committee review in June.
Wedbush Securities analyst Dan Ives has called China “the clear leader” in humanoid robotics, with the United States “playing in catchup mode.” The country’s robots are manufactured at roughly half the cost of Western competitors, a structural advantage that is already shaping US defense and industrial policy responses. Senators introduced the Blocking CCP Spy Tech Act in May specifically naming Unitree, which would require a national security investigation that could restrict the company’s use in federally funded research settings. A documented backdoor in Unitree’s earlier Go1 robot dogs — a hidden service that transmitted data to servers in China — raised questions the company says it has addressed in newer hardware, though no independent audit of its humanoid models has been published.
On the M&A side, Big Tech is moving to consolidate positions. Symbotic acquired Fox Robotics in February, adding autonomous forklift capability to its AI-powered supply chain platform. Meta entered humanoid robotics directly in May by acquiring San Diego-based Assured Robot Intelligence and folding the team into its Superintelligence Labs unit.
What the Reliability Gap Means
The numbers conceal a hard problem. Bessemer Venture Partners, in an April 2026 assessment, placed the industry squarely in what it called “the GPT-2.5 moment for robotics”: capabilities are real, scaling laws are beginning to emerge, but the gap between lab performance and the 99.9% reliability threshold that production deployment demands remains wide.
The reliability math is unforgiving. At 95% per-step success — a figure achievable in controlled lab conditions — a 10-step manipulation task succeeds only about 60% of the time in practice. A warehouse robot failing that frequently requires constant human intervention that defeats the purpose of automation. Production environments typically demand far greater reliability, and the current generation of VLA-based systems is not there yet.
A second constraint is on-device inference. Unlike large language models, which can serve thousands of simultaneous users on shared infrastructure by batching requests, robotics models must generate an environment state every few milliseconds per robot — meaning each deployment effectively requires a dedicated GPU pipeline. The roughly 1,000-fold drop in LLM inference costs over three years has not yet reached robotics inference.
Hardware supply chains present a third constraint. The actuators, force sensors, and dexterous end-effectors that humanoid robots require are not manufactured at the volumes the investment thesis assumes. The EgoScale paper, published in February 2026, provided the first strong empirical evidence that robotics foundation model performance scales predictably with pretraining data size — a signal that the field has crossed into a new phase of capability. But the supply chain ramp-up required to translate that capability into factory-floor deployment involves physical manufacturing constraints that no software engineering can shortcut.
What the funding figures do tell you, unmistakably, is that the bet has been placed. Venture capital spent much of the last decade treating embodied AI as a science project. The market is now treating it as infrastructure — the physical layer of the intelligence stack, as indispensable in the long run as the cloud compute layer was in the prior cycle. Whether that thesis proves correct will play out over years on factory floors, in fulfillment centers, and eventually in homes.
Frequently Asked Questions
What is a Vision-Language-Action model, and why does it matter for robotics investment?
A Vision-Language-Action model, or VLA, is a neural network architecture that processes a camera feed and a natural language instruction simultaneously and outputs motor commands — in a single unified system rather than through separate perception, planning, and control pipelines. The architectural significance for investment is that a VLA can be pre-trained across multiple robot embodiments and then fine-tuned for a specific hardware platform, making the robot brain a software product subject to network effects rather than a hardware product. That shift is what allows companies like Skild AI to carry AI-infrastructure valuations despite not manufacturing robots themselves.
How much has been invested in robotics in 2026, and why do different sources show different numbers?
Dealroom’s broad definition of robotics and physical AI puts 2026 year-to-date global funding at $55.8 billion — nearly double what it captured under the same methodology in all of 2025. Crunchbase’s narrower robotics dataset, which excludes autonomous vehicles, defense platforms, and some AI-first companies, puts the same period at $18.8 billion — itself a record by Crunchbase’s own tracking. Both figures confirm the sector has already exceeded prior full-year records. The gap reflects measurement scope, not a conflict about the underlying trend.
What is the difference between the three investment layers in physical AI?
Investors are placing three distinct bets simultaneously. The first layer is the AI foundation model for robotics — software companies building general-purpose robot brains and simulation infrastructure, valued like AI platform companies. The second is the integrated robot platform — full-stack companies building complete systems for specific deployment environments. The third is specialized applications — defense, maritime, surgical, and agricultural robotics, where domain expertise and existing customer budgets provide clearer near-term revenue. The world-model companies generating synthetic training data sit within the first layer and attracted approximately $6 billion in Q1 2026 alone.
When will humanoid robots be reliable enough for mass commercial deployment?
Not yet. Bessemer Venture Partners characterized the current state in April 2026 as “the GPT-2.5 moment for robotics”: capabilities are real and scaling laws are emerging, but the gap between lab performance and the reliability threshold that production deployment demands remains wide. At 95% per-step accuracy — achievable in controlled settings — a 10-step task succeeds only about 60% of the time, which is insufficient for industrial automation. On-device inference costs and actuator supply chains present additional near-term constraints. The EgoScale paper published in February 2026 provided the first strong evidence that model performance scales predictably with data — suggesting the reliability gap is a solvable engineering problem rather than a fundamental ceiling, but not one that has been solved yet.
