Energy Worker Gap of 500,000 by 2030 Tests Humanoid Robot Hype

A labor problem caused by AI's appetite for power

Most talk about AI and jobs is about software replacing office workers. A new Goldman Sachs analysis looks at a different problem: the people needed to build and run the physical systems that AI depends on. Goldman estimates that the U.S. power and grid value chain will need about 500,000 more workers by 203034. Electricity demand is rising, and AI data centers are a big part of that increase4.

The hard part is training time, not the size of the number. According to the report, many of these jobs take three to four years to learn1. Energy apprenticeship programs took in only about 45,000 new entrants in 2024, while closing the gap would take roughly 65,000 a year14. One outlet ties that 65,000 target specifically to transmission and distribution work, which suggests the shortfall across the whole sector could be bigger5. In the report's own framing, labor is turning into a structural constraint alongside power, because training cannot keep up with the speed of capital spending on new infrastructure12.

Retirements add to the strain. Coverage of the report says about 41% of the energy workforce is close to retirement age7. The sector employed about 8.5 million people in 2024, according to U.S. Department of Energy figures cited in the coverage2.

How big the demand from data centers could get

Lawrence Berkeley National Laboratory estimates that data centers could use about 11.8% of all U.S. electricity by 2030, and as much as 15.3% in some scenarios5. Projects in development show how that demand turns into local grid pressure. In Chicago, a former options-exchange building is being converted into a 33-megawatt data center, about as much power as 28,000 homes use on average23. The developer wants another 17 megawatts if the local utility can supply it23. The city's central business district is expected to use more than 95% of its current substation capacity by 203423.

The political pushback is growing too. New York put a statewide moratorium in place in July on building permits for data centers drawing 50 megawatts or more23. In August, Chicago's mayor joined other officials calling for a construction pause23. Pauses like these might slow demand, but they do nothing to train the line workers and electricians who will be needed either way.

Where robots come in

Goldman's argument for robots follows from the labor shortage. The firm says rising interest in "physical AI," including humanoid robots, comes mainly from the need for workers17. It projects that the humanoid robot market will grow from about 20,000 units in 2025 to 1.4 million in 203517. Barclays makes a similarly optimistic call: its head of thematic FICC research expects the humanoid market to grow from about $3 billion today to $200 billion by 20352.

This is where the numbers start to disagree, and the disagreement matters. The International Federation of Robotics estimates that only about 7,000 humanoid robots were sold worldwide in 2025, and that many went to research and AI development rather than productive work9. Another estimate, from Smart Analytics Global, puts 2026 shipments at about 60,000 units, three times the 2025 level, rising to about 500,000 by 20308. These figures roughly agree that the market is growing fast. They disagree on how big it was to begin with, and 2025 totals range from about 7,000 to about 20,000 depending on the source. A market that small cannot yet do much for a shortfall of half a million workers in one sector.

The coverage also agrees on one thing: big humanoid fleets are still years off. Building robots and the factories to make them takes years and a lot of money. Lenders have little history to judge these projects by, and companies are still working out how to use humanoids at scale1. Goldman expects broad commercial deployment of Chinese humanoids from companies like Unitree and UBTECH between 2027 and 20291. Even Tesla's much-promoted Optimus is not yet being made in volume. As of its July 2026 shareholder update, Tesla was still installing first-generation Optimus production lines at Fremont, and the first units are going to internal training-data collection rather than customers27.

The robots doing grid work today are mostly not humanoids

Read together, the coverage shows that the automation actually running in the energy sector today is mostly specialized machines, not general-purpose humanoids. National Grid uses Boston Dynamics' four-legged Spot robot at its Sandy Pond high-voltage converter station in Massachusetts. There, the robot goes into areas that could be dangerous for people while equipment is live, and uses optical and infrared cameras to look for leaks and overheating7. Data center operators are buying the same kind of robot for security and inspection. Boston Dynamics says interest from that market jumped sharply over the past year, at prices between $175,000 and $300,000 per unit29.

Construction is getting specialized robots as well. A drilling robot built by DeWalt and August Robotics drilled more than 90,000 holes across 10 phases of a hyperscaler's data center builds with 99.97% accuracy. The companies say it can cut construction timelines by as much as 80 weeks25. Other heavy industries were already using autonomous equipment before the humanoid boom, including Deere's farm systems, Caterpillar's mining platforms, and Amazon's Proteus warehouse robot2.

China is furthest ahead in deploying robots on the grid. State Grid Corporation of China announced a roughly $1 billion program this year to acquire about 8,500 AI-powered robots for grid inspection and maintenance24. Robots are also already being used for power-line repair and plant inspection there1. My reading is that the main competition right now is over grid maintenance robots in general, not humanoids specifically.

Robot foundation models need to read a gauge before they can replace a lineworker

Humanoids are getting so much attention because of a change in how robots learn. Robot foundation models, usually built as vision-language-action (VLA) systems, take in camera images and spoken or written instructions and turn them directly into motor commands. That replaces the older approach of hand-coding a separate pipeline for each task17. Google DeepMind, NVIDIA, Physical Intelligence, Figure and others now ship competing models, and they take different strategic paths. Companies that build their own robots tend to keep their models closed. NVIDIA offers GR00T and its Cosmos world models as open components for anyone to build on13.

One recent release is especially relevant to energy work. Gemini Robotics-ER 1.6 came out in April 2026 and added a feature, built with Boston Dynamics, for reading complex gauges and sight glasses11. Inspecting substations and plants means reading exactly those kinds of instruments. Google DeepMind and Boston Dynamics also announced at CES 2026 that they will run Gemini Robotics models on the electric Atlas humanoid and on Spot22. In my view, this is the most direct route from embodied AI research to the inspection work the energy sector needs done.

NVIDIA's GR00T N1.7 points to another possible advance. NVIDIA says it found the first scaling law for robot dexterity: increasing the human first-person video used in training from 1,000 to 20,000 hours more than doubled average task completion11. If that holds up, the shortage of robot training data, which every lab struggles with, could be eased by video of people doing the work13. Skilled trades are a natural source of that footage.

Some caution is warranted. Getting skills learned in simulation to work in the real world is still hard for fine manual tasks12. Labs also measure their models on different tasks, which makes direct comparisons difficult18. Many humanoids still run only a few hours per charge. Atlas, for example, is rated at about four hours with swappable batteries22. The hardware is getting more demanding as well. Counterpoint Research expects the DRAM in each humanoid to roughly double by 2030 as robots run larger foundation models on board8.

Retraining comes first

The Goldman projection works better as a deadline than as a prediction about robots. The 500,000-worker gap is expected by 2030, and most credible humanoid timelines put large-scale use on a similar or later schedule. That leaves human training as the only solution available right away. Pay is a strong draw. Power-line installers earned a median of $92,560 in 2024, according to federal data, and those working for utilities earned more5. Most of these jobs require apprenticeships rather than college degrees5.

For the rest of this decade, the most likely outcome is a stretched workforce helped by specialized robots: quadrupeds, drones and single-task machines, increasingly guided by foundation models that can read instruments and move through hazardous sites7. Humanoids are more likely to start out handling inspection and helping crews than replacing them. Whether they eventually take on more depends on progress in dexterity, battery life and real-world reliability, and the public data does not yet show that progress clearly.