The Capital Has Arrived: Where Physical AI Value Will Actually Land

By Kriv Naicker | Published: 18-11-2026 | Physical AI, AIoT, Strategy & Investment

Physical AI has had its capital moment. Venture investors put US$47.4 billion into the field in the first half of 2026, more than they invested across 2022, 2023 and 2024 combined.[1] What it hasn't had yet is its business-case moment. This piece is about the gap between the two, and where I think the value will land when the gap closes.


I'm writing it for four kinds of reader: the boards and executives deciding where to spend, the investors deciding what to back, the operators in New Zealand and Australia who will end up buying most of this technology, and the people shaping policy around it. The argument is the same for all of them, so I'll make it once and then spell out what it means depending on where you sit.

1. The Capital Arrived First

The headline number needs unpacking. Crunchbase counts robotics, autonomous vehicles, aerospace, drones, industrial automation and sensors as physical AI, and the largest cheques went to autonomy and defence, not humanoids. Waymo's US$16 billion round, at a US$126 billion valuation, made up nearly a third of the half-year total. Anduril raised US$5 billion at a US$61 billion valuation, and defence-tech companies Shield AI and Saronic each raised US$1.5 billion or more.[1] Funding in the second half of 2025 was US$12 billion, so the jump in six months was close to fourfold.[1]


My read is that capital is going first where a buyer with a budget already exists: robotaxi fleets and defence procurement. Humanoids are where expectations are highest and sales are still small. Worldwide humanoid sales were about US$2 billion in 2025, according to analyses cited by JP Morgan, and the same analyses project US$300 billion by 2035.[2]


Unitree's listing in Shanghai on 19 August shows how far expectation has run. It priced at a valuation of about US$9 billion, then closed its first day of trading 460 percent higher, which takes that valuation to roughly US$50 billion.[2] On day one, a single company was worth about 25 times the entire world's humanoid sales for 2025. I don't read that as proof of a bubble. Early capital is how new infrastructure gets built, and some of these companies will deserve every dollar. But it is a price, and the useful question is what has to be true for the price to be right.


A price is a claim about the future. The business case is the evidence for it.

2. The Business Case Is Still Catching Up

Executives are not short of enthusiasm. The Capgemini Research Institute surveyed 1,678 executives at organisations with more than US$1 billion in revenue in early 2026. Two-thirds ranked physical AI a high priority for the next three to five years, and 79 percent said they were already engaging with it. Only 27 percent were deploying or scaling anything, and just 4 percent said they were operating at scale. Nearly eight in ten called scaling a challenge, mostly because of technology and operating readiness.[3]


The humanoid numbers are sharper. Executives named technical immaturity (72 percent), cost (63 percent) and training (58 percent) as barriers. More than six in ten said they were unclear on the return from humanoids, and a similar share expected public resistance to be a critical obstacle.[3] Unitree's own prospectus points the same way: research and education customers made up 73.6 percent of its humanoid revenue in the first nine months of 2025, and industrial applications just 9 percent. Its adjusted profit then fell by more than half in the first quarter of 2026 as research and development spending rose, even with revenue up 68 percent.[4]


The machines executives expect to scale first are the unglamorous ones. More than half named autonomous mobile robots, industrial arms and cobots as the fastest-growing form factors in their organisations, well ahead of humanoids.[3] Executives aren't rejecting the technology. They're being realistic about what is ready. Capital doesn't change that on its own: several advanced robotics startups have already closed, including Cartwheel Robotics this year, K-Scale Labs in 2025 and Alphabet's Everyday Robots in 2023, which Mergermarket attributes to high development costs and failed commercialisation.[5]


Humanoids get the headlines. Mobile robots, arms and cobots are what executives expect to scale first.

3. Where Value Will Land

Think of the market as three layers. Bodies are the machines that act. Brains are the models that reason and plan. Between them sits the layer that connects a machine to a real site and a real outcome: the sensing, the data, the integration and the accountability when something goes wrong. Most of the capital is going to the first two. I think most of the value will be decided in the third.


Brains Models that reason and plan The loop Sensing, data, integration, trust Bodies Machines that act Capital racing in Capital racing in Value is likely decided here

Bodies are capital-heavy and increasingly contested. Mergermarket reports investors treating hardware as a moat, while noting these businesses are capital-intensive and harder to scale, though also harder to displace.[5] They carry geopolitical exposure too. The US announced a ban on imports of Chinese-manufactured humanoid robots in July, citing national security, and JP Morgan puts China's share of the global humanoid and autonomous vehicle market at about three-quarters.[2] Even Unitree's challenge, in the words of Humanoid Analytics founder Rinat Mirzaitov, is "turning hardware and manufacturing leadership into application and workflow leadership."[2]


Brains are getting easier to get. NVIDIA, for one, launched Cosmos 3 in June as an open model it says can serve as a world model or as the foundation for robot action models.[6] When capable models are open, access stops being the scarce thing. What you feed the model, and what you let it do, becomes the question.


That leaves the layer in between, which is familiar territory for anyone who has worked in IoT: sensors, connectivity, data quality, and integration with the systems a business already runs. Mergermarket notes that unlike digital AI, which has consolidated around a handful of platforms, physical AI is highly fragmented, and no single company controls the full stack. Some investors see room for specialists in the bottlenecks, including data collection, edge computing, autonomy software and the human-machine interface.[5] Crunchbase's investors describe the same shift from another angle: companies bundling hardware into recurring or usage-based models, with the hardware acting as the distribution channel for a data flywheel.[1]


Not everyone agrees the middle is where the money ends up. One investor in the Crunchbase piece expects the strongest moats to belong to companies that vertically integrate and own several layers.[1] I think both views can be right. The winners will be the ones that own the loop from sensing to outcome, whether they build all of it or contract carefully for it.


A Kiwi Example

Halter is the clearest New Zealand example I know. It's a sensing business that most people wouldn't call robotics: solar-powered GPS collars that use audio cues and gentle vibrations to move cattle around virtual fences drawn on a smartphone. In March it raised US$220 million at a US$2 billion valuation, led by Founders Fund, double the roughly US$1 billion it was valued at in June 2025.[7] It says it has sold one million collars to more than 2,000 farmers and ranchers across New Zealand, Australia and the US, and American ranchers have built 60,000 miles of virtual fencing since it launched there in 2024.[7] What interests me is where it is heading next: animal health monitoring and pasture management, an operating layer for livestock operations, with field teams working alongside customers rather than selling from a distance.[7] The collar is the body. The value is in the loop around it.


Capital is chasing bodies and brains. Value will be decided by sensing, integration and trust.

4. A Leader's Playbook

If you're deciding where to start, this is how I'd approach it.


Start with the job you can't staff, not the one that demos best. Labour shortages, more than labour costs, are the top driver of physical AI investment in the Capgemini survey, which tells you where the return is easiest to defend.[3] Pick a site you control reasonably well, a clear task, and a clear number attached to it.


Buy the body, own the loop. Machines will change and suppliers will come and go. The operational data, the integration with your systems and the accountability for outcomes are what compound, so contract for them explicitly. If a vendor's model improves from your site's data, know that, and decide whether it's a fair trade.


Prove the return at one site before you scale. Only 4 percent of executives report operating at scale, and nearly eight in ten find scaling hard.[3] Set the number before the pilot starts, count readiness (sensors, connectivity, data quality, process change) as part of the cost, and scale only when the number holds.


Plan for the people working alongside the machine. More than half of the executives surveyed point to improved safety and reduced physical strain as benefits, and more than six in ten expect public resistance to be a critical obstacle for humanoids.[3] Make the case to the people who will work next to it before the decision is made, not after.


Know where your machine comes from. Supply is concentrated, and policy is moving. Ask what happens to support, parts and software updates if import rules change, and whether a second source exists.

5. What It Means for Work

The near-term story is about filling gaps more than removing people. Labour shortages are the top reason executives give for investing in physical AI, led by agriculture, retail, high tech, warehousing and logistics, and automotive.[3] Mergermarket reports construction and mining firms anticipating workforce gaps in the tens of thousands over the next decade.[5]


People also seem to want help with the right jobs. In a survey built by Columbia's Yunzhu Li with Fei-Fei Li's group, about a third of the roughly thousand tasks people said they wanted robots to do were cleaning.[8] That fits the pattern: the work people are happiest to hand over is the work they least want to do.


I'd be cautious about leaning on the shortage story forever. It is the easiest case to defend today, and it may be harder to defend in five years if the machines become far more capable. What I'd say to any organisation deploying now is that how you treat the people working alongside the first machines will set the tone for every deployment that follows.

6. Where You Sit

The argument is the same for everyone, but the question it raises depends on where you sit.


If you're on a board or executive team, ask for the number, not the demo. What does this remove from cost, downtime or risk at one site in the first year, and who owns the data the machine generates?


If you invest, read the price as a claim about the future and ask what the revenue is made of. Revenue that recurs because a machine is working on a customer's site is a different thing from revenue that comes from research and education buyers. Crunchbase also notes that exits so far have been concentrated in aerospace, defence and drones more than in robotics.[1]


If you run an operation in New Zealand or Australia, your advantage is the environments: primary industries, mining, logistics and infrastructure, where labour is hard to find and a lot is already sensed. Halter shows a company from here can capture global value from a sensing-led product. You will probably buy most bodies from global suppliers, so differentiate on the loop and on your own data.


If you shape policy, watch three things: public acceptance, supply-chain dependence, and whether to back sovereign capability in sensing, data and integration. That last one is where a small country can compete. Halter shows the opportunity and the question together. It was founded in New Zealand, it is now led from Boulder, Colorado, and it plans to hire more than 200 people at its Auckland headquarters.[7] How much of the value stays here depends on choices made long before a round like that.

Closing Thought

The machines will keep getting better, and the money will keep coming while they do. What decides who benefits is something less visible: who owns the loop between what a machine senses and what it does. I've spent a long time around IoT, and the sensing and integration work is rarely the part that gets the headline. It is often the part that decides the outcome. That is where I'd put my attention.


The machines will change every year. The loop between what they sense and what they do is what you keep.


References

  1. Azevedo, M. A. Crunchbase News. (2026, August 18). VCs Pour Billions Into Physical AI As The Next Wave Of AI Investing Takes Shape. https://news.crunchbase.com/venture/physical-ai-funding-startups-robotics-aerospace-h1-2026/
  2. Power, J. Al Jazeera. (2026, August 19). China's Unitree soars in market debut as investors bet on humanoid robots. https://www.aljazeera.com/economy/2026/8/19/chinas-unitree-soars-in-market-debut-as-investors-bet-on-humanoid-robots
  3. Capgemini Research Institute. (2026, April 16). Two-thirds of organizations rate physical AI as a high priority for the next three to five years. Press release for the report Physical AI: Taking human-robot collaboration to the next level. https://www.capgemini.com/nz-en/news/press-releases/two-thirds-of-organizations-rate-physical-ai-as-a-high-priority-for-the-next-three-to-five-years/
  4. CGTN (with input from Reuters). (2026, August 10). Unitree Robotics opens IPO subscription, set to become China's first humanoid robot stock. https://news.cgtn.com/news/2026-08-10/Unitree-Robotics-opens-IPO-subscription-a-milestone-for-industry-1PuxbxhJhks/p.html
  5. Hooper, T. & Ansari, I. Mergermarket, ION Analytics. (2026, April 14). Physical AI draws record capital as investors look beyond software (Dealspeak North America). https://ionanalytics.com/insights/mergermarket/physical-ai-draws-record-capital-as-investors-look-beyond-software-dealspeak-north-america/
  6. Walter, N. Interesting Engineering. (2026, June 1). NVIDIA unveils full-stack platform for humanoid robots, robotaxis and smart factories. https://interestingengineering.com/ai-robotics/nvidia-physical-ai-cosmos3-humanoid-robots-tsmc
  7. Stan, A. M. The Next Web. (2026, March 24). Halter raises $220M at $2 billion valuation as virtual fencing for cattle goes global. https://thenextweb.com/news/halter-series-e-virtual-fencing-cattle
  8. Casado, M. (host). (2026, July). Fei-Fei Li on Spatial Intelligence and Robotics, featuring Fei-Fei Li and Yunzhu Li. The a16z Show. https://a16z.com/podcast/fei-fei-li-on-spatial-intelligence-and-robotics/

← Back to Articles