Can Computing and Energy Evolve Without Each Other?
From the Future of Computing and Energy collection
Computing and energy used to belong to different conversations. One appeared in the technology section, the other on the business or policy page. That separation no longer holds.
Large AI training runs now use as much electricity as many thousands of households consume in a year. A December 2024 Lawrence Berkeley National Laboratory report estimated that U.S. data centres used about 4.4% of the country's electricity in 2023 and projected roughly 6.7–12% by 2028. The largest technology companies have begun signing agreements directly with nuclear plants and reactor developers to secure power for infrastructure that barely existed a decade ago.
The common framing treats these as parallel transitions — one in technology, one in energy — that happen to overlap in time. A more accurate framing is that they are structurally coupled: each is the other's most important input and most important customer. Understanding either transition in isolation produces a distorted picture.
Intelligence has a watt cost
The coupling begins with a concrete physical fact: intelligence has a watt cost. Every answer an AI model produces runs on chips that draw power and shed heat, and a small cost per answer multiplied by billions of answers becomes a load large enough to appear in national electricity forecasts. Training the largest models draws tens of megawatts for weeks or months.
This is a new category of industrial demand, arriving faster than grids were designed to accommodate. The energy cost of computation has become a binding constraint on what can be built, where it can be built, and who can afford to build it.
That constraint reshapes geography. Data centres migrate toward cheap, reliable power, and their operators contract directly with nuclear operators and solar developers. Iceland, powered almost entirely by hydropower and geothermal energy, has attracted data centres for the same reason it once attracted aluminium smelters.
The map of computational advantage is being redrawn along energy lines. Regions that can offer power, grid connections, and permits together accumulate a structural lead.
Computing accelerates the energy transition
The coupling also runs in the opposite direction. Computing is accelerating the energy transition itself. Google DeepMind's GNoME project predicted about 2.2 million new crystal structures in 2023, of which roughly 380,000 were judged most stable. Researchers still debate how many are genuinely novel or useful, but the method lets candidate battery, solar, and catalyst materials be screened by computer before anyone mixes chemicals in a laboratory.
Grid management presents a parallel case. A continental grid that integrates millions of solar panels, wind turbines, batteries, and flexible loads is an optimisation problem of enormous size. Machine-learning models forecast renewable output, schedule storage, and help balance supply and demand minute by minute.
These tools make grids with high shares of renewable power easier to run reliably. That operational capability depends directly on computing.
A feedback spiral with governors
Together these create a feedback loop with a specific structure: better computing enables better energy technology, and better energy enables more computing. A new battery chemistry improves grid storage. Cheaper storage makes renewable power more reliable, lower electricity costs make larger computing workloads affordable, and larger workloads speed up the next round of discovery.
Each turn of the spiral can compound the previous one. Like every positive feedback loop, it can also stall.
The spiral has governors, and they are overwhelmingly material, infrastructural, and institutional. Much of the U.S. electricity grid was designed in the mid-twentieth century for large central power plants sending electricity in one direction. Integrating distributed renewables, two-way power flows, and large new data-centre loads requires rebuilding transmission on a vast scale.
A new high-voltage line can take a decade or more to approve and build. New reactors face similar timelines: Georgia's Vogtle Units 3 and 4 entered service in 2023–2024, years late and far over budget. The pace of both transitions is set by permits, wires, and construction crews as much as by any laboratory result.
Concentrated supply chains and scarce skills
Supply chains add a second set of governors. TSMC in Taiwan manufactures the large majority of the world's most advanced chips. China dominates the refining of lithium and rare earths, and Russia holds around 40% of global uranium enrichment capacity.
Batteries need lithium, nickel, cobalt, and graphite, and large wind turbines use magnets containing rare-earth elements. The Democratic Republic of Congo mines roughly 70% of the world's cobalt. These concentrations create geopolitical dependencies that echo the oil dependencies the energy transition was meant to ease, and diversifying them is a decade-scale building project.
The workforce dimension compounds the infrastructure gap. Electricians, line workers, nuclear engineers, and quantum computing specialists are all scarce relative to the plans that depend on them. Training an electrician takes several years, and training an engineer takes longer.
Every technology transition is ultimately a workforce transition. The training pipeline has the longest lead time of any bottleneck and receives the least attention.
Why efficiency raises demand
The efficiency paradox adds a structural wrinkle. William Stanley Jevons observed in 1865 that more efficient steam engines increased total coal consumption, because cheaper work invited more work. Computing shows the same pattern: each generation of more efficient chips makes far larger workloads affordable, and total energy use tends to rise.
Efficiency gains remain valuable, because they make new applications possible. Their effect on total demand, though, is usually growth. Efficiency multiplies capability; supply still has to be built.
The stakes of abundance
What emerges from these interlocking dynamics is a picture more nuanced than either optimism or pessimism can capture. The system contains powerful accelerating mechanisms, as computing compresses discovery timelines in energy technology. It also contains powerful decelerating mechanisms: grid infrastructure, permitting regimes, concentrated supply chains, and workforce gaps that algorithmic progress cannot shortcut.
The pace of both transitions is set by whichever factor moves slowest. The slowest factors are rarely the technologies themselves.
That points to the largest stake of all. If advanced nuclear, renewables, and storage deliver even part of their potential, the world could reach energy abundance at a scale it has never experienced. Cheap, clean, plentiful energy would change the economics of water, food, manufacturing, transport, and computation at once.
Much of the underlying physics is already demonstrated. The open question is whether institutions, supply chains, and political systems can reorganise fast enough to let the technology deploy at the pace the feedback loop allows.
Computing and energy now form a single coupled system whose transitions shape each other's trajectory. Understanding either one requires understanding both, along with the constraints, feedback loops, and bottlenecks that connect them.