How Physical AI Is Quietly Becoming the Most Expensive Bet in Tech.
By Creatives Takeover · July 16, 2026
Why Physical AI could reshape the next trillion-dollar industry.
In November 2025, Jeff Bezos returned to an operational role in technology for the first time since stepping down as Amazon's CEO in 2021. The company he co-founded, Project Prometheus, launched quietly with $6.2 billion in initial funding, an amount that alone made it one of the best-capitalized early-stage AI companies on the planet. By April 2026, the company was finalizing an additional $10 billion round at a $38 billion valuation, backed by JPMorgan and BlackRock, pushing total committed capital above $16 billion.
Project Prometheus has not shipped a product. It has no public revenue, no announced customers, and no demonstrated commercial deployment. What it has is over 120 employees recruited from OpenAI, xAI, Meta, and DeepMind, a mission focused on what the industry calls physical AI, and a founding team that includes a former Google X scientist who worked on Wing drones and Waymo. Bezos has been explicit, in a May 2026 interview with Andrew Ross Sorkin, that Prometheus is building what he calls an "artificial general engineer," a dramatically more capable version of computer-aided design software meant to accelerate how physical objects like aircraft, cars, and factory equipment get designed and built.
This is not an isolated bet. It is the clearest signal yet of a structural shift happening across the entire AI industry, one that most people scrolling past chatbot headlines have not fully registered.
The Industry Just Split in Two
For the last several years, the AI investment story has had a single center of gravity: large language models trained on internet-scale text and image data. That category is not disappearing, but it is maturing in a very specific, very telling way. Per-token pricing for frontier language models has fallen dramatically as competition intensifies and inference costs collapse. The models are getting cheaper, more commoditized, and increasingly similar to one another in raw capability. That is what happens to any technology once the underlying training data source, in this case, the accumulated text of the internet, stops being a scarce or differentiating asset.
Physical AI is the industry's answer to that commoditization, and it is built on a completely different premise. Rather than training on text and images that already exist and that every competitor can access, physical AI systems learn by interacting with the real world directly: through robotics, sensors, manufacturing lines, and controlled experimentation. That real-world interaction data does not already exist in a scrapable form anywhere on the internet. It has to be generated, physically, one interaction at a time, which makes it genuinely scarce in a way that text no longer is. Scarcity, in this industry, is exactly what commands a premium.
Robotics and physical AI startups raised a record $27.6 billion in 2025 alone. The capital flowing into this category is not primarily coming from traditional venture funds either. Prometheus counts JPMorgan and BlackRock among its backers. That is Wall Street balance-sheet capital, the kind typically reserved for infrastructure and industrial assets, being redirected toward AI labs that have not yet generated a dollar of revenue.
The Money Is Not Just Chasing Robots
The instinct is to assume physical AI means humanoid robots, and a significant share of the capital is indeed flowing there. But the more interesting split in where the money is going reveals something sharper about how sophisticated investors are actually thinking about this category.
Figure AI, the most funded pure-play humanoid robotics company in the world, closed a Series C in September 2025 at a $39 billion valuation, a fifteenfold increase from its $2.6 billion valuation less than two years earlier. Its Figure 02 robots have already logged more than 1,250 hours of real production work at BMW's Spartanburg manufacturing facility, supporting over 30,000 vehicles and handling more than 90,000 parts. That is genuine commercial deployment, not a demo reel.
But some of the largest and fastest-growing checks in the category are going to companies that do not build a robot body at all. Physical Intelligence, founded in late 2023 by former Google DeepMind and UC Berkeley researchers, builds foundation models designed to control any robot, on any hardware, without requiring extensive retraining for each new machine. It raised a $600 million Series B at a $5.6 billion valuation in November 2025, and by March 2026 was reportedly in talks to close another billion dollars at a valuation above $11 billion, effectively doubling in four months. Skild AI, founded by former Carnegie Mellon professors and building what it calls an "omni-bodied" AI brain for robots, raised a $1.4 billion Series C in January 2026 backed by SoftBank, Nvidia's venture arm, and Jeff Bezos personally through Bezos Expeditions, alongside strategic investors including Samsung, LG, and Salesforce Ventures.
The strategic logic behind these bets is a picks-and-shovels thesis applied to robotics. If a single software layer can eventually run any robot regardless of who manufactured the hardware, the company that owns that layer collects a recurring toll on every deployment across the entire industry, the way an operating system monetizes hardware it never had to build itself. Even Boston Dynamics, one of the most advanced robotics hardware companies in the world, formalized a partnership with Google DeepMind in January 2026 to have Gemini Robotics models power its Atlas platform, an acknowledgment from one of the most capable hardware builders on earth that the intelligence layer, not the mechanical body, is now the critical dependency.
The Valuations Do Not Yet Match the Revenue
None of this should be mistaken for a sector with clean, provable economics. It is, almost by definition, the opposite.
Physical Intelligence remains pre-revenue, deploying its roughly $1.5 billion in funding almost entirely toward real-world data collection infrastructure, GPU compute, and diverse robotic hardware used purely for training purposes, while maintaining a deliberately lean team of approximately 50 people. Its valuation currently sits at only about five times its total funding raised, a ratio analysts point to as a direct reflection of how capital-intensive it is to train and deploy genuinely useful robot intelligence, compared to the far higher revenue multiples enjoyed by pure software companies. Goldman Sachs projects the entire humanoid robotics market will reach $38 billion by 2035. Figure AI's current valuation, on its own, already exceeds that entire projected market size nine years before the projection's end date. Morgan Stanley's more bullish long-range estimate puts the humanoid robotics market at $5 trillion by 2050, a number so large it functions less as a forecast and more as a statement of how open-ended investors currently believe this category's ceiling to be.
That gap between valuation and demonstrated revenue is the honest center of this story. Prometheus has no announced product. Physical Intelligence has no announced revenue. Figure AI, despite genuine commercial deployment at BMW, has not disclosed the kind of revenue figures that would typically justify a $39 billion valuation under conventional software or hardware multiples. Investors are not pricing physical AI companies on what they currently generate. They are pricing them on the size of the market they believe will exist once the underlying technical bottleneck, reliable general-purpose robot intelligence, gets solved by whoever gets there first.
Why Even the Cautious Money Is Getting In
What makes this bet different from a typical speculative technology cycle is who is actually writing the checks and why.
Apptronik's February 2026 Series A extension brought in AT&T Ventures, John Deere, and the Qatar Investment Authority, alongside repeat investors Google and Mercedes-Benz. NEURA Robotics closed a $1.2 billion round in June 2026 backed by Amazon, Nvidia, Qualcomm, Bosch, Schaeffler, and the European Investment Bank, a combination that mixes AI infrastructure capital, automotive manufacturing expertise, and sovereign industrial policy backing in a single cap table. Nvidia, notably, is positioned to profit from this category regardless of which specific robot body or foundation model eventually wins, since its Isaac GR00T, Cosmos, and Jetson Thor platforms provide the simulation and training infrastructure underlying nearly every major player in the space simultaneously.
That pattern, corporate strategics with deep industrial expertise sitting alongside sovereign wealth funds and traditional venture capital, is not typical of an early-stage speculative bubble. It looks more like the early formation of a new category of essential infrastructure, priced the way investors once priced early cloud computing or early mobile networks: not on current revenue, but on the conviction that whoever controls the layer other companies eventually have to build on top of will capture enormous and durable value later.
The Part of the Story Worth Sitting With
There is a real risk sitting underneath all of this enthusiasm, and treating it as settled would be dishonest. The path from an impressive demonstration to reliable, production-grade commercial deployment has historically defeated far more robotics companies than it has rewarded. Figure AI's own CEO, Brett Adcock, told an interviewer in November 2025 that he would not currently let his own robot operate unsupervised around his young children for hours at a time, a striking admission from the leader of a company marketing its product as home-ready. China's Unitree is already shipping humanoid units at a fraction of the cost basis of its Western, AI-first competitors, raising a genuine question about whether reliability and general intelligence, rather than hardware, will actually be the deciding factor, or whether cost and manufacturing scale will simply win regardless.
The honest read is that physical AI has crossed the threshold from speculative narrative into a genuine, well-capitalized industrial category, evidenced by billions of dollars in committed capital from institutions that do not typically make purely speculative bets. It has not yet crossed the threshold into proven, reliable, at-scale commercial value creation. Both of those things are true simultaneously, and the size of the bet is precisely a reflection of how large investors believe the payoff will be if the remaining technical uncertainty resolves in their favor.
What This Means If You Are Not Raising Billions for a Robotics Lab
Most founders are never going to raise a ten-figure round to build humanoid robots or train foundation models on real-world physics. But the underlying logic driving this entire category is worth understanding regardless of what you are building, because it describes something durable about where genuine competitive advantage is heading across the wider technology industry.
Internet-scale text and image data, the resource that built the last several years of AI value creation, is no longer a differentiating asset. Every major lab has access to a broadly similar version of it, and the resulting models are converging in capability while their price per token collapses. Physical AI's core insight is that the next genuinely defensible moat comes from data nobody else can easily access: information generated through direct, proprietary interaction with a specific real-world environment. That principle scales down considerably from Bezos-scale industrial ambition. A startup with genuine, proprietary operational data about a specific workflow, industry, or customer behavior, data that a general-purpose AI model trained purely on public internet text simply does not and cannot have, is applying the exact same structural logic that is currently commanding sixteen-billion-dollar valuations at the frontier.
Five Things Worth Taking From This
Commoditization at the frontier creates opportunity at the edges. As general-purpose language models converge in capability and collapse in price, the value shifts toward whoever controls data, workflows, or environments that remain genuinely scarce. Ask what your business generates that a foundation model trained on public text simply cannot replicate.
Capital follows scarcity, not effort. The billions flowing into physical AI are not rewarding companies for working hard. They are rewarding companies for sitting on top of a data source, real-world physical interaction, that is structurally difficult for competitors to replicate at any price.
The infrastructure layer often outlasts the product layer. Nvidia profits from physical AI regardless of which specific robot or foundation model wins. Consider whether your own business is better positioned as a product competing for a single outcome, or as infrastructure that benefits from the category's growth regardless of who wins within it.
Valuation ahead of revenue is not automatically irrational, but it is not automatically justified either. Physical Intelligence trades at roughly five times its funding precisely because training real-world intelligence is extraordinarily capital-intensive. Understand whether a valuation reflects genuine scarcity and a real technical bottleneck, or simply momentum and narrative.
Reliability, not capability, is usually the real bottleneck. Figure AI's own CEO will not yet trust his product unsupervised around his children. The gap between an impressive demonstration and dependable, everyday commercial deployment is where most ambitious technology categories actually die. Building toward reliability, however unglamorous, is usually the more durable strategy than chasing the most impressive demo.
Physical AI is not a hype cycle in the way that term usually implies. It is backed by more institutional, industrially sophisticated capital than most AI trends before it, and it is solving a real structural problem: the value of purely text-based intelligence is running out precisely because it became too easy to replicate. Whether it becomes the most important bet of this decade or simply the most expensive one still being figured out will depend entirely on whether reliability catches up to ambition before the money runs out of patience.