“Whoever wins the energy race likely will win the AI race.” – William Thompson – Barclays analyst
The escalating electricity requirements of artificial intelligence infrastructure have created a critical bottleneck, transforming power generation capacity into the decisive factor determining which nation secures dominance in AI development. Data centres in the United States consumed 176 TWh in 2023, equivalent to 4,4% of total national electricity use, with projections indicating growth rates accelerating to between 13% and 27% annually through 2028 under various scenarios4. This surge stems directly from the proliferation of GPU-accelerated servers optimised for AI workloads, which have more than doubled energy demand since 2017. In parallel, China’s data centre power needs are forecasted to expand at 17% per year, reaching 479 TWh by 2030-a volume matching France’s entire national consumption-concentrated in provinces like Hebei, Shanghai and Zhejiang2. These figures underscore a fundamental constraint: AI training and inference processes demand immense computational power, rendering energy availability the limiting resource in the global race1.
Historical trends amplify this tension. US electricity demand stagnated for two decades, leaving utilities unprepared for the abrupt AI-driven spike, while China sustained 8% annual growth through relentless investment in generation and grid infrastructure2. The asymmetry is stark: American renewables’ levelised cost of electricity has risen slightly due to reshoring, tariffs and policy shifts, contrasting with China’s solar PV costs halving and onshore wind falling by two-thirds since 20202. Such disparities position energy infrastructure as the fulcrum of strategic competition, where delays in permitting, grid upgrades or capacity additions could cede ground to rivals. Utilities now identify data centres-particularly AI-enabled ones-as the primary driver of future load growth, with firms like Southern Company anticipating 6% yearly sales increases from 2025 to 2028, up from prior 1-2% expectations4.
Mechanisms Driving the Power Crunch
At the core lies the thermodynamics of computation. Modern large language models require clusters of thousands of high-end GPUs, each consuming hundreds of watts continuously during training runs that can last weeks. A single frontier model iteration might draw power equivalent to a mid-sized city, with inference queries adding persistent load. This has inverted traditional efficiency gains: despite per-flop energy improvements, total demand explodes as models scale under laws like those posited by OpenAI’s scaling hypothesis, where performance improves predictably with compute investment. The result is a feedback loop: greater AI capabilities spur more ambitious projects, necessitating ever-larger data centres.
Geographically, concentrations exacerbate strains. In the US, facilities cluster in regions like Virginia’s ‘Data Center Alley’ and Texas, overwhelming local grids and prompting blackouts or moratoriums. China mirrors this in coastal hubs, pushing expansion inland for cheaper renewables, yet faces similar interconnection delays2. Globally, the International Energy Agency projects data centre electricity use quadrupling by 2030, potentially rivalising Japan’s total consumption by 20255. Such projections challenge net-zero pathways, as fossil fuel backfilling risks entrenching emissions while clean capacity lags.
US Vulnerabilities and Policy Responses
America’s grid, fragmented across three major interconnections, grapples with underinvestment. A recent Resource Adequacy Report warns that supply expansions fail to match AI-driven demand, imperilling reliability2. Nine of the ten largest utilities cite data centres as the chief growth vector, with AI explicitly named by seven of eight in forecasts4. President Trump’s ‘America’s AI Action Plan’ underscores national security imperatives, yet execution hinges on reforming permitting-often taking years-and innovating large-load tariffs. Proposals include nuclear restarts, small modular reactors and gas peakers, but timelines clash with hyperscalers’ 2026-2030 buildouts.
Financial commitments magnify risks. Tech giants have pledged trillions in capex, with AI accounting for 90% of recent S&P 500 gains, 80% of profits and 75% of investments3. Nvidia derives 15% of projected sales from ‘circular deals’ among ecosystem players, while OpenAI’s $1,4 trillion commitments-contingent on future revenues-evoke perpetual motion concerns3. Debt-financed data centre construction invites overbuild, akin to past infrastructure bubbles, should adoption falter or efficiency leapfrog projections.
China’s Structural Advantages
Beijing’s command economy enables swift mobilisation. State-directed investments have sustained grid expansions, with renewables now competitive on cost. Data centre demand, though nascent, benefits from this foundation: by 2030, key provinces will absorb loads rivalising US markets like PJM2. Policies prioritise AI as a ‘new quality productive force’, channelling subsidies into domestic chips, though US export controls on advanced semiconductors force workarounds like Huawei’s Ascend series. Energy self-sufficiency via coal-backed renewables provides resilience, contrasting US reliance on imported components.
Yet challenges persist. Coastal concentrations strain urban grids, while water cooling for data centres competes with agriculture in water-scarce north. Export ambitions-hyperscalers like Alibaba eye global expansion-require overseas power deals, exposing vulnerabilities.
Debates and Objections
Sceptics question the linkage’s inevitability. Howard Marks argues AI’s transformative potential invites bubble behaviour, with unpredictable demand growth defying forecasts: ‘a year from now, AI may do 10x or 100x what it can today’3. Parallels to the internet era cite overcapacity risks, though proponents counter with existing revenues-Anthropic’s 100x two-year growth-and productivity gains3. Critics like those at NewClimate Institute warn of a ‘climate strategy crisis’ for net-zero pledges, as AI erodes margins for clean transitions5.
Technological counters include specialised ASICs promising 10-100x efficiency over GPUs, potentially easing loads, but incumbents like Nvidia dominate transitions3. Fusion or quantum breakthroughs remain speculative. Economists debate pass-through: will AI savings boost profits or fuel price wars, diluting incentives3? Job displacement-43% task automation per Vanguard-complicates political buy-in3.
Strategic Tensions and Geopolitical Stakes
The US-China rivalry frames energy as proxy warfare. Washington’s CHIPS Act and export bans aim to starve China’s compute, but Beijing retaliates via rare earths and supply chains. Trump’s plan signals escalation, yet power shortages could neutralise semiconductor edges. Allies like Taiwan (TSMC) and South Korea face similar crunch, tilting balances.
Europe lags, with regulatory hurdles stifling builds; Malaysia and Singapore impose caps5. This fragments the race, favouring agile powers.
Why Supremacy Hinges on Power
Victory demands not just generation but resilient, low-cost supply. Winners will train larger models faster, iterate superior architectures and deploy at scale, compounding advantages. Laggards risk technological lockout, as first-mover data moats solidify. Economically, AI could add trillions to GDP, but uneven energy access concentrates gains. Environmentally, fossil dependence undermines legitimacy; clean dominance bolsters soft power.
Investors face asymmetry: equity rewards winners, debt punishes losers in winner-take-most arenas3. Policymakers must balance innovation with reliability, lest blackouts erode public support. Ultimately, the energy race dictates AI’s trajectory, forging a new geopolitical order where electrons, not silicon, crown champions.
References
1. Tech: Who will win the US-China AI race? – The Edge Malaysia – 2025-12-04 – https://theedgemalaysia.com/node/782629
2. Powering China’s data centres – Wood Mackenzie – 2025-07-25 – https://www.woodmac.com/blogs/the-edge/powering-chinas-data-centres/
3. Is It a Bubble? – Oaktree Capital Management – 2025-12-09 – https://www.oaktreecapital.com/insights/memo/is-it-a-bubble
4. [PDF] 2024 United States Data Center Energy Usage Report – 2024-12-17 – https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
5. The AI race could make net zero impossible. Here’s why – 2025-11-11 – https://www.context.news/net-zero/the-ai-race-could-make-net-zero-impossible-heres-why
