2025 Year-End AI Review: Four Acts in the Drama — From 'Frenzy' to 'Clarity'

By the end of Act I, you might naturally ask: Since everyone is pursuing smarter, more efficient AI, won’t compute demand decrease—and thus ease pressure on electricity supply? Quite the opposite. Act II delivers a stark, unyielding truth: In the era of AI inference, the electricity challenge has become even harder—and colder. Act II: Energy’s Life-or-Death Struggle — A Physical-World ‘Power Grab’ If Act I was a battle of algorithmic ingenuity, then by the second half of 2025, the war abruptly shifts to the most primal, heaviest, rust-scented domain of all: electricity. Isn’t AI just software on hard drives—strings of 0s and 1s? So how can it be held back by *electricity*? Here’s a simple analogy: You’re doing homework at home. No matter how sharp your thinking, a sudden blackout leaves you staring blankly. Worse still, large-model training isn’t like a desk lamp flickering off—it’s tens of thousands (even hundreds of thousands) of GPUs running at full power, simultaneously. Their electricity demand isn’t just *massive*—it’s *extremely stable*. A voltage fluctuation lasting less than one-millionth of a second—or an unexpected circuit breaker trip—could crash a training job that’s already consumed hundreds of millions of dollars and run for three months. It’s like assembling a 10,000-piece LEGO set, only to have someone flip the table just before snapping in the final piece—everything resets to zero. Thus, a surreal scene emerged in 2025: hoodie-clad software geniuses, sitting in climate-controlled offices, began thinking about energy the way 19th-century coal barons once did. The most dramatic example? The resurrection of the Three Mile Island nuclear plant. Three Mile Island was the site of the worst nuclear accident in U.S. history. For decades, nuclear power has been nearly taboo in Western societies; Germany, for instance, chose to endure energy shortages rather than keep its reactors online. But in 2025—driven by AI—the taboo broke. As early as late 2024, Microsoft signed a 20-year, massive power-purchase agreement to restart one reactor unit at Three Mile Island—exclusively powering Microsoft data centers. By 2025, this plan didn’t cool down; instead, it accelerated amid the AI compute explosion. By June 2025, reports even surfaced that the restart timeline could be moved up to 2027 to meet urgent demand. It’s not just Microsoft: Amazon began negotiating long-term power contracts with nuclear plants, and Meta started investing heavily to keep existing nuclear facilities operational. Why nuclear power specifically? Because while wind and solar are clean, they’re weather-dependent and inherently unstable. Nuclear power delivers constant, 24/7 output—regardless of time of day or weather—and is therefore AI’s true ‘soulmate.’ You’ll realize: GPUs are no longer the sole bottleneck. The real constraint is a *stable, reliable electricity infrastructure*. America’s aging grid system is like fitting a Ferrari with Qing-era rural dirt roads—it simply cannot handle these ‘electricity-guzzling behemoths.’ So tech giants are stepping directly into power generation and grid construction. By late 2025, U.S. energy regulators were even forced to draft new rules—how to balance ultra-large data centers’ electricity needs against ordinary residential consumption, preventing AI-driven power grabs from spiking prices and triggering social unrest. Shifting focus back to China, you’ll see a different competitive advantage: China possesses the world’s strongest ultra-high-voltage (UHV) and extra-high-voltage (EHV) power grid system—even as top-tier GPUs remain restricted.

AI’s Electricity Frenzy! Energy Solutions: Nuclear Power + New Energy + Long-Duration Storage — Advancing in Tandem

• Natural Gas: Serving as the ‘emergency tourniquet,’ natural gas offers mature technology and short construction timelines, delivering stable 24-hour power ideal for rapid deployment of AI data centers. However, it remains constrained by fuel price volatility and carbon-emission regulations—making it unsuitable for matching the 30-year operational lifespan of modern data centers. • New Energy (Photovoltaics / Wind): Paired with short-duration storage (e.g., lithium-ion batteries), it leverages peak–valley electricity price differentials to reduce costs—ideal for small-to-medium AI data centers near urban areas (e.g., the ‘Digital Xichong • Urban Superbrain’ project, which uses a 200 kW / 430 kWh storage system to ensure stable operation). • Shortcoming: New energy sources are highly intermittent; short-duration storage (2–4 hours) can only smooth out momentary fluctuations—not resolve cross-day supply deficits. Under extreme weather conditions, natural gas backup remains indispensable. 2. Medium-Term Efficiency Gains (3–8 Years): New Energy + Long-Duration Storage + Grid Modernization (Core Enabler) • Dominance of New Energy: Solar PV and wind power become the backbone of green electricity supply—aligning with policy mandates such as the requirement that *new data centers source no less than 80% of their electricity from renewable sources* (per the joint notice ‘On Promoting High-Quality Development of the Renewable Energy Certificate Market,’ issued by China’s National Development and Reform Commission, National Energy Administration, Ministry of Industry and Information Technology, and other departments). The EU has introduced an ‘AI Compute Carbon-Footprint Traceability System,’ while U.S. tech giants like Google have pledged 100% renewable electricity for AI workloads. • Breakthroughs in Storage Technology Are Critical: • Vanadium Redox Flow Batteries (VRFBs) Emerge as the Near-Term Optimal Solution: Technologically mature, water-based electrolytes eliminate fire/explosion risks, cycle life exceeds 15,000 cycles, calendar life reaches 25 years—perfectly matching AI data centers’ >10-year operational horizon. Crucially, power and capacity can be scaled independently, enabling >4-hour long-duration discharge. The Jimsar 1-GWh vanadium flow battery project in Xinjiang has already validated this at scale. • Hybrid Storage Complements Deployment: Urban AI data centers increasingly adopt ‘vanadium + lithium-ion’ configurations—lithium-ion handles millisecond-level responses to high-frequency fluctuations, while vanadium batteries manage long-duration peak shaving, valley filling, and emergency backup. • Diversified Long-Duration Technologies Advance in Parallel: In Europe and the U.S., iron–air battery demonstration projects are already underway. With low cost, multi-day duration, and low technical risk, iron–air storage holds clear strategic value for long-duration applications. • Grid Modernization Support: Concurrent upgrades to ultra-high-voltage transmission and smart grids—alongside streamlined approval processes for new-energy grid integration—address the longstanding bottleneck of ‘plenty of renewable generation, but no way to deliver it.’ 3. Long-Term End State (>8 Years): Nuclear Power as Primary, New Energy + Storage as Auxiliary (Ultimate Solution) While new energy plus storage reduces carbon emissions, geographic and weather constraints prevent them from guaranteeing year-round, stable power for gigawatt-scale AI compute hubs. Nuclear power remains the long-term cornerstone—with two competing technological pathways intensifying: (1) SMRs (Small Modular Reactors): A ‘LEGO-style’ Revolution of Traditional Nuclear Power • Core Logic: Transform nuclear plants from bespoke civil-engineering projects into factory-manufactured, modular units assembled on-site like LEGO bricks—enabling phased installation and lowering capital thresholds. • Advantages: Proven technology, regulatory acceptance, and inherent safety via physics-based passive cooling; multiple modules can be combined flexibly to match dynamic load profiles. • Drawbacks: Ancillary infrastructure costs remain stubbornly high; grid interconnection is still required—and permitting bottlenecks persist at ~7 years. (2) Oklo Model: Building Nuclear Reactors in AI’s ‘Backyard’ • Key Innovation: Reactor size shrinks to that of a dining table—capable of direct integration into data center campuses, operating entirely off-grid, with 15 MW output tailored precisely for large AI data centers. • Business Model Innovation: Instead of selling nuclear hardware, Oklo offers a ‘power subscription service’—handling everything from installation and operations to regulatory compliance—dramatically lowering the barrier to entry for AI companies. • Risks: High regulatory uncertainty; zero revenue to date; enormous capital requirements; and extremely high cash-flow failure risk—making large-scale rollout financially precarious.

Germany’s Idle Nuclear Plants Eye New Life as AI Data Centers

19 hours ago

PreussenElektra partners with Nvidia to explore turning idle German nuclear plants into data centers, capitalizing on existing grid infrastructure while utilities monetize decommissioned sites.

bloomberg

Germany’s Idle Nuclear Plants Eye New Life as AI Data Centers

FAQ