For three years, "AI" meant something you talked to. A box on a screen that wrote your emails, summarized your meetings, and occasionally hallucinated a court case. That version of AI lived entirely behind glass, with no hands, no weight, no stake in the physical room it was supposedly helping you navigate. In 2026, that framing is starting to look dated, and not because the chatbots got worse.
In late 2025, Forrester's chief research officer, Sharyn Leaver, predicted that "the AI hype period ends" in 2026 as pressure for measurable results intensifies. Gartner's 2025 Hype Cycle for AI described a gradual pivot away from generative AI as the central focus toward enablers such as AI-ready data and agents, and coverage of the report places generative AI in the "Trough of Disillusionment."
The bills are showing up in earnings, too. In August, Canva, which Fortune calls a rare startup that grew fast and made money, cut its 2026 revenue growth forecast from 30% to 20% because the cost of delivering AI features came in higher than expected. The company serves more than 265 million monthly users, so even small per-request costs add up quickly. Figma's second-quarter results the same month told a related story: revenue grew 48% and the full-year outlook went up, yet the stock fell about 16% as AI-related costs and a slower third-quarter guide weighed on investors. Even companies with growing revenue are feeling the compute bill.
None of this means AI is going away. It means the industry is doing what every hyped technology eventually does: running out of road on its first story and looking for a second one. And the second story, this time, has a body.
The pivot nobody had to force
At CES 2026, Nvidia's Jensen Huang declared that "the ChatGPT moment for physical AI" had arrived. In June, SoftBank's Masayoshi Son told CNBC that physical AI and robotics are where he sees the next trillion-dollar company emerging. Boston Dynamics unveiled the production version of its Atlas humanoid at the show, aimed at tasks from material handling to order fulfillment, with its entire 2026 output already committed to Hyundai and Google DeepMind.
Carmakers are testing humanoids on real factory floors, not trade-show stages. BMW ran a ten-month pilot of Figure AI's Figure 02 at its Spartanburg plant, during which the robot helped produce more than 30,000 X3s. Audi trialed UBTech humanoids at its Chinese joint venture in 2025 and planned pilots at German sites this year.
What's actually shifted is the software. Vision-language-action (VLA) models, the physical-world cousins of large language models, take camera images and a language instruction and directly predict the actions a robot should perform, instead of chaining separate perception and planning systems. Simulation did much of the rest. Georgetown's CSET describes how training in simulated environments and digital twins, then transferring the result to real machines, addresses the cost and time of collecting real-world training data.
The field is also less settled than the keynotes suggest. That same February 2026 CSET brief notes that it lacks even a shared name ("embodied AI," "physical AI," and "generative physical AI" are all in use) and that the mature software, hardware, and data ecosystem a robot needs to navigate the three-dimensional world at scale does not yet exist. It adds that humanoids likely account for less than 1% of the robotics market by revenue, with total shipments in the hundreds. Nvidia sells the GPUs behind all of this, so its framing is convenient. That doesn't make it wrong, but it is context.
Robotics is the visible fraction of the problem
Say "embodied AI" and most people picture a humanoid robot. That's the part you can photograph. According to the people actually building these systems, it's also only part of the challenge. A white paper summarizing an SAE World Congress 2026 panel concluded that embodied AI should be treated as a systems challenge rather than only a software challenge. Safe deployment depends on sensors, hardware robustness, operating limits, human interaction, communications, and cybersecurity as much as on the model. A bad recommendation from software is an inconvenience. The same mistake in a vehicle or a robot can mean property damage or injury.
Underneath that systems view sits a quieter research problem: soft robotics. A 2021 review in Frontiers in Robotics and AI observes that the vast majority of research on soft components focuses on actuators, the parts that move, even though sensors, mechanisms, power supply, and flexible electronics matter just as much. It singles out power as a major open issue, more serious still for soft systems. The review also highlights "multifunctional" materials. Dielectric elastomers, for example, are best known as actuators, but because they are structurally equivalent to capacitors they can also act as stretch or pressure sensors, store charge, or harvest energy. That is a design idea worth watching, partly because it will never make a viral demo video.
Then there's the compute inside the robot itself. CSET's primer describes a split: heavy training and simulation happen in data centers, while energy-efficient edge processors handle real-time inference within the robot's power and thermal limits. The SAE paper makes the same point from the vehicle side: unlike data-center AI, ground vehicles have strict power limits and still need low latency. It is the same energy problem the rest of this piece is about, shrunk down to a single machine.
The bill nobody wants to look at directly
Here's the number that complicates all of the above. Gartner expects worldwide data center power demand to rise 27% in 2026 alone, to 132 gigawatts from 104 gigawatts in 2025, and to reach 290 gigawatts by 2030. The IEA puts global data center electricity consumption at around 415 terawatt-hours in 2024 and projects it will roughly double to 945 terawatt-hours by 2030 in its base case. Servers, meaning CPUs plus accelerators such as GPUs, account for around 60% of a modern data center's electricity use on average. Forecasts differ: Gartner's own estimate is that consumption will exceed 1,200 terawatt-hours by 2030.
In the US, a June 2026 analysis from Lawrence Berkeley National Laboratory, as reported by E&E News, suggests data centers could go from roughly 4.5% of US electricity today to between 9.5% and 15% by the end of the decade.
A fair reading needs some perspective. The IEA notes that data centers account for less than 10% of global electricity demand growth through 2030. The strain comes from concentration. Nearly half of US data center capacity sits in five regional clusters, and in Virginia data centers consumed about 26% of the state's electricity in 2023. The same Pew summary of IEA findings says a typical AI-focused hyperscale data center uses as much electricity as 100,000 households, and the largest ones now under construction are expected to use about 20 times that.
Now add bodies to the picture. Today the load is overwhelmingly in the data center. CSET describes physical AI as needing one computer to train the AI, one to deploy it, and another to run the digital twin where robots practice, while its hardware table lists battery packs of 2 to 4 kilowatt-hours for a typical humanoid. Individual robots are modest consumers. But every step toward real fleets adds simulation compute, onboard inference, and charging on top of demand that already strains local grids. The industry spent three years solving how to make AI think. Making it act adds to the power bill instead of shrinking it.
Where the electrons are supposed to come from
The headline answer is nuclear. A tracker of hyperscaler nuclear deals counted 13 announced projects committing about 9.8 gigawatts for Google, Amazon, Meta, and Microsoft as of May 2026, a mix of existing plants and advanced reactors. Meta alone has agreements for up to 6.6 gigawatts. Oklo says its agreement with Meta supports up to 1.2 gigawatts in southern Ohio, with the first phase targeted as early as 2030 and full capacity by 2034. Amazon's expanded contract with Talen Energy allows supply of up to 1.92 gigawatts from Pennsylvania's Susquehanna plant through 2042.
The newer entrants are small modular reactors, which the IAEA defines as advanced reactors producing up to 300 megawatts of electricity per module and designed to be built in factories and shipped to the site. The pitch is speed and flexibility compared with conventional plants.
Renewables will carry a large share too. In the IEA's base case, renewables and natural gas lead in meeting data center demand, with half of the growth met by renewables supported by storage and the broader grid. But AI facilities want round-the-clock power, and intermittent generation can't supply that on its own, which is where storage comes in. One example is NANEXBAT, a five-year, €8 million Dutch consortium led by the University of Twente that is working to improve sodium-ion batteries for heavy-duty transport and stationary grid storage, partly to reduce dependence on scarce materials.
The skepticism deserves a hearing. The Bulletin of the Atomic Scientists argues that the announcements are largely PR. It points out that the largest nuclear investment announced so far, $650 million, would not fund a single Natrium reactor, while Amazon and Google said in February that they expect to spend up to $200 billion and $185 billion, respectively, on data centers this year. Whether the reactors show up on schedule is a separate question from whether the press release was accurate.
What if computation itself needed less power?
Everything above assumes the fix is on the supply side: build more generation, store more power, feed a hungrier system. There's a smaller, odder line of research asking a different question. What if the computation didn't need to be this power-hungry in the first place?
In September 2026, a team at Maynooth University's Hamilton Institute in Ireland, led by Professor Damien Woods, published a paper in Nature describing a Scaffolded DNA Computer. The design borrows the scaffold principle from DNA origami: a long DNA scaffold plus a pool of short "compute" strands, with the program encoded in the DNA sequences themselves. According to the university's announcement, the team mixed water, salt, and DNA in a test tube, heated it, and let it cool. The molecules interact and settle into a structure, and as Woods puts it, "that structure is the answer." In the paper's terms, the system computes by relaxing toward its thermodynamically favoured state instead of spending continuous energy to enforce correctness.
The researchers ran ten programs, including adding two 25-bit numbers, which the paper counts as a 100-bit computation. Small computations finished in under a minute, while the largest additions took up to around 14 hours. The system could be reused up to 25 times in a row. Woods points out that computing and data storage account for 23% of Ireland's electricity, and he argues that computing has been fixated on silicon even though other forms of computation exist, the brain among them.
To be precise about what this is: the DNA computer does not generate energy, and it will not run a robot's actuators or charge a battery. It is also not a free lunch. The paper lists heating and annealing among the system's costs, the experiments ran on lab instruments, and the authors do not offer an energy comparison with silicon. Co-author Abeer Eshra says it is slower than silicon and not meant to compete on speed, and the paper itself says scaling up required slower anneals and further optimization. What the work offers is a different lever from the nuclear and battery stories above. Those are about finding more power. This is about questioning how much computation has to depend on continuous power at all. It is early, exploratory research, but it is a reminder that the premise "AI needs more electricity" rests on one particular way of building computers.
Humanity, still trying to make itself
Step back from the gigawatts and the DNA strands for a second, and the whole story starts to look less like a technology roadmap and more like a mirror. For three years, we built something that could talk like us. Now we are building it hands, eyes, joints, and a body to stand in a room the way we do, and discovering, the way parents eventually do, that a body is expensive, fragile, and needs feeding constantly.
There is an old, unresolved question sitting underneath all of this. When humans build a mind and then insist on giving it a body, what are we actually trying to make? Every civilization that reached for artificial life, from golems and automatons to Frankenstein's creature, and now humanoid robots on a factory floor in South Carolina, was in some sense trying to answer a question about itself by building an answer outside itself. The post-hype era of AI is less a retreat from ambition than ambition getting more literal. We are no longer just simulating a mind. We are constructing the whole apparatus a mind needs to exist in physical space, one gigawatt and one DNA strand at a time.
Whether that apparatus ends up running on small modular reactors, sodium-ion batteries, molecular computers, or some combination nobody has proposed yet, the underlying project has not changed since long before anyone had a GPU. We are not just trying to build something intelligent. We are still trying to build something like ourselves, and finding out, again, that making a body is the hard part.
