Eric Schmidt, former Google CEO, warned in April 2025 congressional testimony that AI dominance requires energy dominance. He projected the US needs 29GW more power by 2027 and 67GW by 2030 for AI data centres – equivalent to nearly 100 nuclear reactors.
His predictions are proving accurate. As AI technologies scale, the fundamental economics hinge on access to cheap, abundant electricity. Without it, the grand visions of AI-driven innovation could stall, not due to flaws in the algorithms or waning enthusiasm, but because the physical infrastructure simply can’t deliver the juice.
The demand for power from AI is exploding at an exponential rate, far outpacing what most analysts anticipated just a few years ago. Globally, electricity consumption for data centres is expected to double to around 945 terawatt-hours by 2030 in baseline scenarios, with AI workloads being the primary driver. In the US alone, AI-specific power capacity is projected to surge from about 5GW today to more than 50GW by 2030, equivalent to the total global AI compute demand from just a few years prior.
More aggressive forecasts paint an even starker picture: US data centre power demand could hit 130GW or 1,050 terawatt-hours by 2030, representing up to 12 per cent of the nation’s total electricity use. Goldman Sachs anticipates a 165 per cent increase in global data centre power demand by the end of the decade.
This isn’t hyperbole; the numbers reflect the inherent energy intensity of AI operations. A single query on a generative AI model like ChatGPT consumes nearly 10 times the electricity of a standard Google search. Training a large model such as GPT-4 devours more power annually than 5,000 average US households. AI server racks are now drawing 50 to 200 kilowatts each – significantly more than traditional servers – potentially increasing US AI power needs by 30-fold by 2035. Rack power densities in AI facilities are climbing from 40 kilowatts to 130 kilowatts, with projections reaching 250 kilowatts soon.
The commercial sector, including data centres, is seeing electricity use grow up to 5.0 per cent annually through 2026. Eight major hyperscalers are forecasting a 44 per cent year-over-year jump to $371 billion in spending on AI data centres and computing resources in 2025 alone. Generative AI queries are estimated to consume 15 terawatt-hours this year, ballooning to 347 terawatt-hours by 2030.
Yet, the US power supply is buckling under this pressure, with constraints tightening and costs spiralling. Utilities are grappling with unprecedented demand from data centres, which could account for 40 per cent of net new electricity additions through 2030, expanding at a 23 per cent compound annual growth rate. Grid interconnection queues are clogged, with delays stretching 5 to 10 years in prime data centre locations. For 84 per cent of developers, power availability has eclipsed all other site criteria, including fibre access and regulations, as the top priority.
Voltage fluctuations from AI’s rapid power draws – swinging tens to hundreds of megawatts in fractions of a second – pose risks to grid stability. Outdated infrastructure and fragmented oversight compound the problem, with some utilities projecting load growth exceeding 20 per cent by 2035. Data centres are expected to consume 12 per cent of US electricity by 2028, up from 4.0 per cent in 2023, and hyperscale campuses may soon demand over 1 gigawatt each, enough to power 800,000 homes.
Electricity prices are reflecting this strain, eroding the cost advantages that once fuelled tech growth. Residential rates have climbed more than 30 per cent since 2020, outpacing inflation by a wide margin. The national average reached 19 cents per kilowatt-hour in August 2025, with household bills averaging around $149 per month based on typical usage. In high-demand states, rates vary from 11.69 cents to over 39 cents per kilowatt-hour, but the trend is upward nationwide – up 27 per cent nominally from 2019 to 2024, with a 6.7 per cent year-over-year increase noted in mid-2025.
Renewables, while promising, can’t provide the reliable baseload AI requires; grid bottlenecks lead to up to 40 per cent waste in wind power generation. As a result, onsite power generation is set to triple by 2030, with 38 per cent of facilities incorporating it and 27 per cent relying fully on it for faster deployment and reliability.
This energy crunch casts a long shadow over AI stocks, where valuations are predicated on unrelenting growth. The ‘Magnificent Seven’ – Apple, Microsoft, Amazon, Alphabet, Meta, Nvidia and Tesla – now comprise over 34 per cent of the S&P 500’s market capitalisation as of August 2025, up from 28 per cent at the end of 2023 and around 33 per cent at the start of this year. However, they contribute only about 27 per cent of the index’s earnings, revealing a precarious valuation gap fuelled by AI euphoria.
Nvidia’s leadership has acknowledged electricity as the emerging bottleneck, pivoting from performance-per-dollar metrics to power-constrained designs. Hyperscalers view energy shortages as the critical hurdle, with individual AI facilities consuming as much power as mid-sized cities. The median AI data centre size is expected to swell from 175 megawatts today to 375 megawatts by 2035, amplifying the pressure.
The so-called ‘AI bubble’ may not deflate from overhyped models or investor burnout, but from these intractable energy realities. Analysts caution that if AI demand falters, utilities could be left with stranded assets, triggering a broader market correction. Energy inefficiency is hindering scalability, forcing companies to pursue unconventional solutions. Microsoft is reviving nuclear reactors and exploring fusion and geothermal energy; Amazon has acquired nuclear-powered campuses; Oracle is developing sites with small modular reactors; even Tesla is installing gas generators. Meta is constructing massive facilities backed by new gas plants, with the costs being passed on to consumers.
In China, unrestricted energy expansion – adding 400 gigawatts annually across hydro, nuclear, wind, solar and battery sources – provides a competitive edge, potentially allowing them to outpace the US in computing power despite inferior chip technology. This disparity raises national security concerns, as US regulatory hurdles slow progress.
Echoing Schmidt’s urgency, Bill Gates emphasises that access to affordable electricity is a huge limiting factor for virtually every sector of the economy today. Removing those limits, he argues, could be as transformative as the invention of the steam engine at the start of the Industrial Revolution. That’s the kind of breakthrough a future powered by fission and fusion could deliver, unlocking unprecedented growth not just for AI, but across industries. Gates’ vision aligns with his investments in advanced nuclear through TerraPower, which is pioneering reactors to provide clean, reliable baseload power tailored for data centres and beyond.
For investors, particularly those entrenched in momentum plays or passive index strategies, the exposure is massive. The Magnificent Seven’s outsized influence means any energy-induced slowdown could drag down major indices. Diversification beyond AI hype is imperative – consider pivoting to energy enablers like nuclear firms, uranium suppliers or utilities poised to benefit from the surge. On-site generation technologies, such as fuel cells and modular reactors, offer resilience against grid failures. The shift to direct-current architectures and advanced cooling for high-density servers highlights opportunities in power management and infrastructure.
Schmidt’s warnings aren’t apocalyptic but a roadmap. The AI revolution demands an energy revolution. Ignoring the grid’s limitations invites peril but addressing them head-on could unlock unprecedented growth. As power becomes the binding constraint – not chips or code – the winners will be those who secure it first. In this high-stakes game, the portfolio that powers through will prevail.
The writer is former head of Citigroup’s emerging markets investments and author of ‘The Gathering Storm’.