Skip to content
‹ Back to IndustriesIndustrial IoT

Physical AI value may move from robot hardware to deployment, McKinsey finds

McKinsey now believes physical AI will generate over $1 trillion in value by 2035. In a June 2026 video series, it had estimated that milestone for 2040. Most robotics funding still goes to robot makers and model developers. McKinsey expects value to shift toward software, deployment and field operations, which puts the buyer's site and its field data at the center of physical AI deals.

This story was produced through MarketScale. See how Industrial IoT teams put it to work with AI Visibility (GEO).

By MarketScale Newsroom · Physical AiMckinseyGeneral-purpose RoboticsHumanoid Robots
Share
Physical AI value may move from robot hardware to deployment, McKinsey finds

Key takeaways

01

Robot makers and foundation model developers have taken about $33 billion of the $38 billion invested in general-purpose robotics, while deployment partners and key components split under $6 billion, the reverse of where value is expected to move.

02

If the most valuable training data is the record of a robot failing and a person stepping in, every pilot on a customer's floor doubles as a data collection run for the vendor.

03

Compare robot vendors on deployed performance and proof in similar environments—not just hardware specs or purchase price—and scrutinize how deployment data and intervention logs are handled.

Free workspace

Turn your Industrial IoT expertise into content.

Record interviews, organize footage, and write with AI on a free trial of the MarketScale platform for qualifying companies. No demo required, no credit card.

Try it Free

Add up McKinsey's funding figures and robot manufacturers and the labs building their AI models have taken about $33 billion of the $38 billion invested in general-purpose robotics. The companies that install those robots, supply training data, and build the actuators, dexterous hands and force sensors have shared less than $6 billion.

McKinsey's Industrials Practice expects value to move in the opposite direction. In an analysis published November 20 by Ani Kelkar, Bernd Heid, Christian Jansen and Mark Patel, the firm says value may increasingly shift from hardware, which draws most investment today, toward software, deployment and field operations. It also brought its headline number forward. By 2035, physical AI applications could generate over $1 trillion in economic value. In a June 2026 video series, "The age of thinking machines," McKinsey estimated that physical AI would generate over $1 trillion in value by 2040. It now believes it will hit this milestone by 2035.

For an operations leader deciding whether autonomous forklifts belong in a distribution center or a mobile manipulator belongs on a line, that forecast changes what is actually being bought. If deployment is where value builds up, then the site itself, the integrator's skill and the data a fleet generates on a real floor become the assets that count. A buyer controls more of those than of the robot.

Warehouses, mines and shipyards go first

McKinsey defines physical AI as systems that use sensors and learned models of their surroundings to decide and act in the physical world with minimal human intervention. That separates them from conventional automation, which runs fixed scripts under fixed conditions. The firm says these systems are well past the pilot stage in structured settings such as warehouses, factories, mines and ports, though adoption varies widely by sector.

Physical AI already running at scale

1 million+
Amazon mobile robots across more than 300 facilities, coordinated by one foundation model
10%
Fleet travel time Amazon expects its DeepFleet model to cut
11 billion metric tons
Material moved by Caterpillar's autonomous haul trucks
500,000
Waymo paid rides per week in March 2026, double the 250,000 of less than a year earlier

McKinsey & Company, citing Amazon, Caterpillar and Waymo

There's more depth behind those figures. Caterpillar's autonomous fleet numbered 690 trucks at the end of 2024, and the company aims to triple it by 2030. In China, EACON has put its autonomous driving technology on more than 2,000 haul trucks. Shared autonomous vehicles worldwide are delivering more than 700,000 fully autonomous rides a week.

The mechanism McKinsey describes is a loop. Each deployment produces data that improves performance, and better performance opens the door to wider deployment. The loop starts most easily where tasks repeat and the environment is controlled, which explains why mines and fulfillment centers got there before public roads. McKinsey compares it to industrial automation, where companies that built a large installed base early won on customer relationships, service networks and application know-how, sometimes against rivals with better products.

In practice, that turns a vendor's installed base in environments like the buyer's own into a real selection criterion, alongside the spec sheet. A system with thousands of hours in comparable aisles has had more chances to learn from conditions like that buyer's.

Where 2026 robotics money landed

The investment curve is steep. According to McKinsey's analysis of PitchBook data, general-purpose robotics funding was about $1 billion in 2020 and climbed to around $9 billion by 2025. In just the first eight months of 2026, it hit $21 billion. Simple arithmetic shows that this partial-year total is already more than double the figure for all of 2025. Autonomous vehicles are excluded from those numbers, and they still draw most physical AI investment.

General-purpose robotics funding by region, January to August 2026 ($ billion)
McKinsey & Company analysis of PitchBook data · © MarketScaleDownload chart

Through 2025 the United States led on cumulative funding, while China took the lead in 2026. China's cumulative total is about $17 billion, which is 46% of all general-purpose robotics investment. In hardware its share climbs further. It holds about three-quarters of global funding in force and precision sensing, and more than half in robot makers and integrated actuators. With about $5 billion cumulative, Europe ranks third. McKinsey says the region's strength lies in its component suppliers, its mechatronics expertise and a manufacturing base suited to deployment, not in private capital.

For procurement teams qualifying suppliers of actuators, tactile sensors and other components, that concentration is a sourcing fact worth mapping before volumes grow. It describes where the capital has gone, which tends to be where capacity gets built.

The customer's floor is the training set

Simulation is cheap and scales well for geometry and navigation. McKinsey points out that it falls short where small differences in force, friction and slip decide whether a gripper holds a crushed carton or drops it. Those details are learned most reliably in live operations, which is why the firm calls field deployment the richest source of insight.

L.E.K. Consulting, in an October 14 analysis of humanoid robotics by Harsha Madannavar, Philip Meier and Shaaina Dayal, takes that a step further. Once a robot can do a task reliably, routine data from that task adds less, and the scarce material is failure and intervention data: what happened when the policy broke and how a person recovered. L.E.K. cites University of Michigan research published in Nature Communications in 2026, in which simulation training built on near misses and failures improved autonomous vehicle safety performance by 90%. The firm notes that result comes from a different application and may not carry over directly to manipulation.

If the most valuable training data is the record of a robot failing and a person stepping in, every pilot on a customer's floor doubles as a data collection run for the vendor. That suggests the operator running a pilot is contributing something of value, and contract terms about how site data, intervention logs and resulting model improvements are used deserve the same attention as uptime guarantees.

Pricing is moving the same way. Venture firm Decasonic argues physical AI business models are shifting away from selling hardware and toward robots-as-a-service, priced per task, per order or per system, much as AI customer-service software has moved to pricing by outcome. It suggests that as more robot model APIs become publicly available, deployers can increasingly choose between models for each task. Decasonic's conclusion fits McKinsey's: when intelligence gets abundant, value moves to the layers that route, evaluate and deploy it.

Ask each vendor for cost per useful autonomous hour, and how many of the quoted hours were teleoperated. L.E.K. notes teleoperated hours are expensive and are not autonomous hours, so a low sticker price can hide a lot of human supervision.

Why integrated hardware still wins today

The hardware-heavy funding pattern has a reason behind it. L.E.K. observes that most leading Western humanoid developers, including Tesla, Figure, 1X, Apptronik and Agility Robotics, are vertically integrated, and calls that a rational response to current conditions. The most useful data combines robot state, forces, tactile feedback and operator interventions, which requires control of both the sensors and the policy. Actuators, tactile sensors and dexterous hands are also hard to source at the needed performance and cost.

The category is also still small. By L.E.K.'s count, roughly 19,100 humanoids shipped worldwide during the first half of 2026. That is a 272% increase over the prior year, though from a low base, and full-year revenue is estimated at $1.6 billion. At individual Western customer sites, fleets range from a handful of robots to a few dozen. McKinsey's own forecast has general-purpose robotics and autonomous vehicles still leading applications in 2035, with intelligent factories gaining ground. As L.E.K. notes, a buyer weighing a single work cell built around one fixed, repetitive task may find that an application-specific robot still costs less and runs more reliably than any humanoid.

Even so, the interfaces are starting to separate. At its South Carolina assembly plant, BMW runs Figure robots, and it is also piloting Hexagon Robotics' AEON humanoid at a German plant. In June 2026, NVIDIA released an open reference design for humanoids that explicitly separates body from brain. It combines a Unitree chassis and tactile hands from Sharpa with Jetson Thor compute and Isaac GR00T models. L.E.K. cautions that this shift may not happen: a cross-hardware layer may never take shape if custom drivers and integration work remain expensive.

Milestones that will test the forecast

Shipbuilding offers an early look at less structured work. HII, the largest US naval shipbuilder, intends to partner with two firms: Path Robotics for adaptive welding, and GrayMatter Robotics for autonomous surface preparation, coating and inspection. The schedule calls for proof-of-concept demonstrations in 2026 and a full pilot in 2027. Work on a hull varies far more than a tote traveling along a conveyor, so a pilot that succeeds would prove the deployment loop can work beyond the orderly settings where it started.

The other signal to watch is in the funding data itself. If McKinsey's read is right, the under-$6 billion going to deployment partners, training data and key components should start growing faster than the robot makers' share, and more contracts should be written per task or per outcome. Caterpillar's goal of tripling its autonomous fleet by 2030 is a public marker of what scale in one structured domain looks like.

Featured companies

Your experts belong here

Every story in MarketScale Industrial IoT starts with a company putting its controls engineers, plant-floor specialists, and integration partners on the record. Buyers are already reading this topic. The only question is whose experts they find.

Plant and controls buyers research deep before contact, and your engineers get to shape that research.

Book DemoSee how it works15 minutes, straight to a calendar.

About the author

MarketScale Newsroom
MarketScale NewsroomEditorial Team, MarketScale

The MarketScale Newsroom reports on the companies, technologies, and trends shaping 16 B2B industries. It turns primary sources and expert commentary into clear, useful coverage for the people doing the work.

B2B Weekly

The week in Industrial IoT, and sixteen other industries, every Monday.

Ten stories, one-line takes, five minutes. Free.

Industrial IoT: are you visible to AI?

Before they reach out, Industrial IoT buyers ask AI engines which vendors to trust. Explore how your experts, customers, and partners can become useful content for buyers and AI search.

Free Trial

You just read one Industrial IoT expert. Your company is full of them.

This article was produced through MarketScale. The same platform turns your controls engineers, plant-floor specialists, and integration partners into the articles, video, and social content Industrial IoT buyers are searching for. Start a free trial and see it with your own people. For qualifying companies, no credit card, no demo required.

NPS +73 · 1,000+ creators · 38+ countries

What your free trial includes

Hands-on access to the MarketScale platform
Media requests to your crowd, remote recording, AI writing tools
No demo required. No credit card.
For qualifying companies. Company confirmation required.

More Industrial IoT Insights

Factory AI that tunes machines works best with limits on what it can change

Factory AI that tunes machines works best with limits on what it can change

GridRaster co-founder Dijam Panigrahi expects closed-loop quality control, spatial AI and physical AI (adaptive robotics) to shape manufacturing over the next 12 to 18 months, with each enabling software to affect what happens on the plant floor. Parsec’s 2026 report says 72% of manufacturers use AI in some form, but only 10% have deployed it at scale. The piece argues that moving from pilot to production starts with one measurable line, clean time-stamped machine data, clear limits on what systems may change (with operator override and logging), and accountable ownership before go-live.

  • 01Seventy-two percent of manufacturers use AI somewhere and 10% use it at scale; the article argues that closing that gap requires deciding before go-live what each system is allowed to change on its own and who owns the outcome.
  • 02Bounded parameter ranges, operator override and a log of every automatic adjustment are the guardrails that let closed-loop quality control keep the traceability regulated plants need, and they can be written into a purchase specification today.

Nov 20, 2026

Atlas Sayaç opens international R&D project call for water metering and industrial IoT

Atlas Sayaç opens international R&D project call for water metering and industrial IoT

Atlas Sayaç, a Turkish engineering and manufacturing company, opened an international R&D call on November 19, 2026 for water metering and industrial IoT projects. The company proposes linking payments to project success and revenue sharing, inviting water utilities, equipment manufacturers, systems integrators and researchers to collaborate on engineering problems or technologies ready for commercialization.

  • 01Atlas Sayaç is proposing a success-based R&D arrangement for mutually accepted projects, deferring payment until the project meets agreed success criteria and the partner begins earning revenue.
  • 02The company combines electronics, embedded software, mechanical design, manufacturing and software-platform integration across M-Bus, LoRaWAN, NB-IoT/GSM and RS-485 technologies.
  • 03Prospective partners submit a nonconfidential outline of their engineering problem or technology, application, target market and commercial goals to https://atlassayac.com/en/rd-partnerships/ for feasibility discussion.

Nov 19, 2026

Rockwell gives AI agents only the simulation tasks a validated digital twin can support

Rockwell gives AI agents only the simulation tasks a validated digital twin can support

At Automation Fair 2026, Rockwell Automation laid out rules: AI agents may chain digital twin simulation tools only for tasks a validated model supports. Model validation comes first. Green, yellow and red zones set where virtual tests replace physical ones, starting with one or two engineering questions.

  • 01An AI agent's scope in Rockwell's method is set by how well the digital twin has been validated: green-zone tasks can replace physical tests, yellow-zone tasks can trim some, red-zone tasks are off limits.
  • 02People who use simulation results outnumber the engineers who run simulations by about 10 to 1, so agents are aimed at the queue of handoffs between specialists, not at engineering judgment.
  • 03A model can match conveyor dimensions and machine motion and still assume perfect part supply, so any green label should be tied to the specific question it was validated for.

Nov 19, 2026

Explore More Industrial IoT Insights

Read more expert perspectives from across Industrial IoT.

Browse Industrial IoT Hub

About the Expert

MarketScale Newsroom
MarketScale Newsroom

Editorial Team

MarketScale

The MarketScale Newsroom reports on the companies, technologies, and trends shaping 16 B2B industries. It turns primary sources and expert commentary into clear, useful coverage for the people doing the work.

For B2B teams

Your experts could be publishing here

Stories like this one run on content MarketScale captures from real practitioners. See how your team's expertise becomes coverage in Industrial IoT and beyond.

Book a Demo

Or call us. No forms required. We pick up. 214-945-2512