2026-07-08T18:30:00.000Z
The Next Robotics Boom Is Being Built on Existing Machines
The 39.03% CAGR reflects a land grab for the AI layer sitting on top of two million already-installed robots, not adoption of an entirely new robot category.
Boston Dynamics built its entire 2026 production run of the electric Atlas humanoid for two customers only: parent company Hyundai Motor Group and research partner Google DeepMind. Every other manufacturer waits until 2027. That decision, made months before a single commercial Atlas unit shipped, says more about this market than any growth chart could.
The hardware is no longer the constraint. NVIDIA's Isaac platform already runs inside FANUC, ABB Robotics, KUKA, and Yaskawa controllers, four companies that together operate more than two million industrial robots worldwide. What remains scarce is the reasoning layer that lets those robots adapt to new tasks without months of reprogramming.
Kaiso Research's primary data puts the global robotics intelligence market at USD 6.24 billion in 2025, climbing to USD 168.37 billion by 2035 at a 39.03% CAGR. This is not a market growing because more robots get built. It's a market growing because the intelligence running on top of robots already on factory floors is being rebuilt from scratch, and the rebuild has already split into incompatible camps.
Software, Not Hardware, Is the Scarce Input Behind a 39.03% CAGR
Software, not hardware, is the scarce input behind this market's growth rate. The segment accounts for the larger and faster-growing share of total revenue, with foundation model-based platforms posting the steepest growth curve inside it, per Kaiso Research's primary data. The market this report tracks excludes robot hardware entirely, no actuators, no sensors, no mechanical structures. It covers the reasoning layer instead: perception software, autonomous decision-making engines, reinforcement learning platforms, and multi-agent coordination systems that decide what a robot does once it is switched on.
Manufacturing and logistics together generate the largest combined share of application revenue, a pairing that makes sense once hardware and intelligence are treated as separate purchases. A FANUC arm bolted to a factory floor and an autonomous forklift moving pallets through a KION-operated warehouse need the same underlying capability: perceive an unstructured environment and decide what comes next without an engineer rewriting code for every variation.
In reviewing the demand curves across North America and Asia-Pacific in this report series, Kaiso Research's primary data confirms a familiar split. North America holds approximately 38% of global revenue, the largest single regional share, while Asia-Pacific is growing fastest. Jensen Huang's framing from CES 2025, that physical AI had reached its "ChatGPT moment," reads less like marketing now and more like the procurement signal that manufacturers, logistics providers, and healthcare systems are all responding to at once.
NVIDIA's Platform Already Runs Inside Two Million Installed Robots
NVIDIA's Isaac and Cosmos platforms already run inside the controllers that FANUC, ABB Robotics, Yaskawa, and KUKA use to operate a combined installed base exceeding two million industrial robots worldwide. GTC 2026 confirmed how far that reach now extends beyond those four incumbents: Figure and AGIBOT for humanoids, Skild AI and World Labs for general-purpose robot brains, and Hexagon Robotics, CMR Surgical, and Medtronic for precision and surgical applications. Skild AI, for its part, partners directly with ABB Robotics and Universal Robots to deploy generalized robot intelligence across automotive and electronics manufacturing, while also working with contract manufacturer Foxconn on the high-precision assembly lines that build the chips NVIDIA's own platform depends on.
The technical core of that adoption is Jetson modules embedded directly into robot controllers, giving each arm real-time AI inference at the edge rather than relying on a remote server. KION Group is running the same playbook for autonomous warehouse forklifts, building physics-accurate digital twins with NVIDIA and Accenture before a single forklift moves a real pallet for its anchor customer GXO. PTC is doing the equivalent for engineering workflows, connecting Onshape CAD files directly into Isaac Sim so FANUC America and Fauna Robotics can validate robot designs before they are built.
Boston Dynamics chose a different reasoning layer entirely. Its new production Atlas humanoid runs on Gemini Robotics foundation models developed by Google DeepMind, not NVIDIA's platform, and the CES 2026 partnership commits the entirety of Boston Dynamics' 2026 Atlas output to two buyers: Hyundai Motor Group's new Robot Metaplant Application Center and Google DeepMind itself. Hyundai owns Boston Dynamics outright and is backing the bet with a $26 billion U.S. manufacturing investment that includes a dedicated humanoid factory targeting 30,000 units a year by 2028.
Tesla represents a third position on the same map: full vertical integration, with no third-party robotics intelligence platform involved at all. Optimus runs on Tesla's own compute infrastructure and shares its underlying AI stack with the company's self-driving software, a choice that keeps Tesla independent of both NVIDIA's platform and Google DeepMind's models. It also means Tesla owns every failure mode by itself.
Deployment Cost Fell 40 to 60 Percent. Funding Arrived Before Proof
Deployment cost is the mechanism, not novelty. Multi-step task planning grounded in real-world physics, the core capability inside Google DeepMind's Gemini Robotics model, cuts total robot deployment cost by an estimated 40 to 60%, since a manufacturer no longer needs a specialist to hand-code a new task sequence every time a product line changes. That single cost reduction explains why foundation model-based learning is now the fastest-growing learning model inside this market, and it matters most for exactly the kind of multi-week reprogramming cycles that have made robot redeployment too expensive to justify for shorter manufacturing runs, the same constraint NVIDIA's Jetson-based edge inference is built to remove.
The academic evidence behind that shift is stark. Submissions covering vision-language-action, or VLA, architectures at the International Conference on Learning Representations grew from a single paper in 2024 to 164 in 2026, per Kaiso Research's primary data, evidence that the field is reorganizing itself around one architecture in under two years. Patent filings covering physical AI robotics methods confirm the same pattern from a different angle, growing 187% across 2024 and 2025 compared with the prior two years, concentrated heavily around perception and manipulation claims rather than locomotion.
Capital followed the same signal well before the underlying technology had shipped commercially. Robot-related startups raised USD 6.4 billion in just the first 11 months of 2024 alone, a figure that confirms investor confidence in robotics intelligence crossed a structural threshold before Boston Dynamics or Tesla had a commercial humanoid on a single factory floor. That timing only makes sense if investors believed the foundation-model approach to robotics represented the same kind of platform shift large language models delivered for text, arriving on a compressed timeline because the underlying compute infrastructure already existed.
The capital arrived first.
Atlas Ships Exclusively to Hyundai and DeepMind Through 2027
Boston Dynamics committed its entire 2026 Atlas production run to exactly two buyers, Hyundai Motor Group's Robot Metaplant Application Center and Google DeepMind, locking out every other customer until 2027 at the earliest. The partnership, announced at CES 2026 in Las Vegas, integrates Gemini Robotics foundation models directly into the new electric Atlas, with joint research already underway at both companies' facilities. Hyundai's robotics timeline names specific milestones rather than vague ambitions: high-precision sequencing tasks by 2028, complex assembly tasks by 2030.
Figure AI took a different route toward the same kind of commercial proof. The company closed a Series C round exceeding USD 1 billion at a USD 39 billion post-money valuation in 2025, funding its BotQ manufacturing line and proprietary Helix reasoning system, and it has already delivered Figure 02 units to paying automotive customers, a commercial milestone neither Boston Dynamics nor Tesla had matched as of this report's publication.
Tesla's position is the outlier worth naming directly. Optimus Gen 3 entered hand production at Tesla's Fremont facility in January 2026, and those Gen 3 hands moved into 24/7 autonomous factory shift testing by the second quarter, but Tesla has announced no external Optimus customers, and Musk himself confirmed as recently as June 2026 that the robot is not in usage in Tesla's own factories in any material way. A robot that is shipping and a robot that is working are not the same announcement. Conflating the two is the single most common mistake in current humanoid robotics coverage.
World Models, Not Bigger Robots, Decide Who Wins
World models, not larger robots or faster actuators, are what separate a machine that generalizes to an unfamiliar task from one that needs to be reprogrammed for it. The report segments robotics intelligence into seven distinct capability types: perception, navigation, decision, manipulation, collaborative, cognitive, and autonomous intelligence, each addressing a different point in the perceive-reason-act loop that defines an intelligent robot. Foundation model-based and world model-based learning are pulling ahead of older approaches like supervised and imitation learning because they generalize, while a reinforcement learning policy trained for one warehouse layout typically fails inside a second warehouse with different aisle widths.
Physical Intelligence's pi0.5 foundation model, released in 2025, illustrates the property directly. It is built to perform tasks in new environments without prior site-specific training, the opposite of how industrial robot software has always worked. Older robot software was deterministic. Program the exact sequence, and the robot executes it precisely and nothing else.
The foundation-model approach inverts that relationship entirely, asking the robot to reason about an unfamiliar task the way Gemini Robotics or NVIDIA's Isaac GR00T models do, using simulation-trained world models to predict the physical consequences of an action before taking it. NVIDIA's Cosmos world models exist specifically to generate the synthetic training data this approach requires, converting what Jensen Huang has called robotics' data problem into a compute problem instead, one that scales with GPU capacity rather than with how many real robots a company can physically deploy to collect data.
Three Strategies Compete: Platform, Vertical Integration, and Pure Intelligence
Three distinct competitive strategies are visible inside this market, and they carry different risk profiles for the executives evaluating them. NVIDIA's platform strategy, anchored by Cosmos, Isaac, and GR00T, captures the broadest base of industrial incumbents, FANUC, ABB Robotics, KUKA, Yaskawa, and Universal Robots among them, by selling simulation and inference infrastructure rather than competing with any single robot maker directly. That breadth is the strategy's strength and its exposure: NVIDIA does not build robots, but its margin depends entirely on the robotics industry's willingness to standardize on its compute and simulation stack.
Google DeepMind and Boston Dynamics chose vertical depth over platform breadth, restricting Gemini Robotics to a single hardware partner and committing scarce production capacity to two buyers through 2027. That's a slower path to revenue but a tighter feedback loop between model and hardware, the same argument Tesla makes for keeping Optimus entirely in-house.
Pure-play intelligence vendors, Skild AI, Physical Intelligence, and World Labs among them, are betting that the reasoning layer becomes valuable enough to license across whichever hardware platform ultimately wins, the same position software companies held during the personal computer hardware wars of the 1980s. None of the three strategies has proven definitively correct. NVIDIA's own equity position in Figure AI, detailed next, suggests even the platform leader is hedging its bet rather than betting the company on one outcome.
Figure's $39 Billion Valuation Prices In a Bet NVIDIA Also Made
Figure AI's Series C round, led by Parkway Venture Capital with participation from Brookfield Asset Management, NVIDIA, Intel Capital, and Qualcomm Ventures, closed in 2025 at a USD 39 billion post-money valuation on more than USD 1 billion in committed capital. NVIDIA's presence in that round is the detail worth isolating. The same company building the platform that FANUC, ABB Robotics, and KUKA rely on also took a direct equity position in a humanoid robot maker, a hedge that only makes sense if NVIDIA expects multiple winners rather than one.
The broader funding pattern tells the same story at a larger scale. Robot-related startups raised USD 6.4 billion in the first 11 months of 2024 alone, well before most of the foundation-model architecture behind these companies had reached commercial deployment. That timing matters for anyone evaluating valuation multiples in this segment today, since capital arrived ahead of revenue, and the next 18 to 24 months of commercial deployment data, not further funding announcements, will determine which of these bets actually pays off.
Europe's Dual Compliance Burden Hits Humanoid Vendors First
The EU AI Act classifies autonomous robots as high-risk AI systems under Annex I, with full compliance obligations applying from August 2027, while the EU Machinery Regulation 2023/1230 layers a second, parallel requirement from January 2027. A robotics intelligence vendor selling into Europe has to satisfy AI Act documentation for the software and Machinery Regulation conformity assessment for the physical robot at the same time, a dual burden that does not exist in the U.S. market yet.
Safety standards have not caught up with humanoid hardware either. ISO 10218 governs stationary industrial arms, but dynamically stable, walking robots like Atlas and Optimus fall outside that standard's original scope. CEN/TC 310, working alongside ISO, is revising ISO 10218 directly, while a separate working group, with experts from Agility Robotics and Boston Dynamics, develops ISO 25785-1 specifically for robots that require active balance control. Final publication is not expected before 2026 or 2027, which leaves manufacturers applying draft principles as best-practice guidance rather than binding compliance today.
Manufacturing CTOs Need a Platform Bet. Investors Need a Different One.
Manufacturing and logistics executives evaluating a robotics intelligence platform now face an architecture decision with real switching costs attached, not a vendor comparison they can revisit casually next budget cycle. Choosing NVIDIA's Isaac and Cosmos stack means joining the same camp as FANUC, ABB Robotics, and KUKA, with the broadest current installed base but no guarantee that breadth survives a future where Gemini Robotics or a pure-play intelligence vendor proves technically superior for a specific task class. Mid-market manufacturers running legacy automation systems will absorb whichever platform choice they make across two to three procurement cycles before a migration path to an alternative becomes economically realistic.
Investment committees evaluating exposure to this market should treat the NVIDIA, Google DeepMind, and Tesla positions as genuinely different risk classes rather than interchangeable robotics plays. NVIDIA's platform model captures value regardless of which specific humanoid or industrial robot maker wins, a structurally safer position than betting on Figure, Boston Dynamics, or Tesla individually. Direct exposure to a single hardware platform, the way NVIDIA itself took a position in Figure's Series C, concentrates risk on one company's ability to convert funding into reliable commercial deployment, the exact gap Tesla has not yet closed with Optimus. European exposure compounds that calculus, since any platform choice now also has to clear the EU AI Act and Machinery Regulation compliance burden described above, a cost that lands differently depending on which stack a vendor has already standardized on.
The Data Problem NVIDIA Wants to Convert Into a Compute Problem
Jensen Huang has framed the central constraint on physical AI bluntly: robotics has a data problem that the industry needs to convert into a compute problem, since robot learning requires physical interaction data, sensor streams, and manipulation video that simply does not exist at internet scale the way text did for large language models. NVIDIA's Physical AI Data Factory Blueprint and Cosmos world models are a direct response, generating synthetic training data to substitute for real-world collection, but synthetic data carries its own fidelity risk that has not been resolved at scale.
Regulatory fragmentation compounds the data risk. A humanoid certified safe under draft ISO 25785-1 principles in one jurisdiction may not satisfy EU AI Act Annex I documentation in another, and no finalized global standard exists for walking robots yet. The third risk is reputational rather than technical: Tesla's own executives have acknowledged that early Optimus Gen 3 units are not performing useful factory work, a credibility gap that competitors with actual commercial deployments, Figure's BMW pilot among them, will keep using as a selling point.
By 2028, This Market Has Far Fewer Plausible Winners
By 2028, when Hyundai's humanoid factory is meant to reach 30,000 units a year and Tesla's stated million-unit ambition either materializes or does not, this market will have far fewer plausible winners than the dozen or so companies currently competing for attention. The next 18 months will separate platform rhetoric from deployed reality, since Boston Dynamics' exclusivity arrangement with Hyundai and Google DeepMind expires for new customers in 2027, the same year EU Machinery Regulation obligations take full effect and ISO 25785-1 is expected to reach final publication. NVIDIA's installed base inside FANUC, ABB Robotics, KUKA, and Yaskawa gives it the broadest current distribution, but distribution is not the same as proof that Cosmos and Isaac-trained models outperform Gemini Robotics on the tasks a specific buyer actually cares about.
The Stack You Pick Is a Capital Commitment, Not a Software Choice
The USD 168.37 billion question by 2035 is not whether robots get smarter. Kaiso Research's primary data confirms that trajectory already, at a 39.03% CAGR that few software categories outside generative AI itself are matching. The real question is which of three incompatible bets, NVIDIA's open platform, Google DeepMind and Boston Dynamics' closed integration, or Tesla's vertical isolation, captures the margin once deployment moves from pilot programs to production at scale.
History offers a mixed precedent. Open platforms won the personal computer era. Closed integration won the smartphone era.
Robotics intelligence does not have to follow either script, and the fact that NVIDIA is hedging its own platform bet with a direct equity stake in Figure suggests even the platform leader isn't certain which model wins this time. Manufacturers locking in a stack this year are not making a software decision. They are making a multi-year capital commitment to whichever camp turns out to be right, and the data needed to know which camp that is does not exist yet.
---
About Kaiso Research and Consulting
Kaiso Research and Consulting is a global market intelligence firm publishing 5,000+ research reports across 11+ industry verticals.
[email protected] | +1 872 219 0417
Isha Paliwal, Lead Industry Analyst, Kaiso Research and Consulting | Covering robotics and physical AI markets across North America, Europe, and Asia-Pacific
Published: 2026-06-29 | Report Code: LSDB14
Market Study: Access the full index or request a complimentary sample directly via the Robotics Intelligence Market Size, Trend and Opportunity Analysis Report page
