The Global "Reliable Power" Shockwave. Article 1 of 2: Big Tech's Nuclear Race

datacenterprimerja
Feb 27
4 min read

James Soh. First published on 14th of January, 2026.

The rise of the Speed-to-Power currency.

In 2019, Three Mile Island Unit 1 powered down—a nuclear cautionary tale from the 1970s. January 2026: it's the crown jewel of AI infrastructure. Microsoft's pioneering 20-year PPA with Constellation to resurrect it as Crane Clean Energy Center (2027 target, 835MW firm power) sparked a domino effect. This "Sputnik Moment" declares: Big Tech demands Firm Power—24/7 reliable power—more than just carbon-free.

Big Tech's Nuclear Arms Race: Racing Ahead of the Power Wall

Microsoft blazed the trail, but 2026 reveals a full hyperscaler stampede—Meta, Amazon, Google, Oracle—all betting billions on nuclear to sidestep grid chokes and secure sovereign capacity:

Big Tech's nuclear binge solves two killers: 7-10 year U.S. grid queues + AI demand doubling data center power by 2028. They restart proven plants now (fastest firm MW) while funding SMRs for 2030s scale—creating private "power islands" that skip utilities entirely. Meta grabs 6.6GW (nation-scale). Amazon sticks servers right at reactors. No more waiting or grid begging—hyperscalers generate their own power.

The Personal Frontline: Lessons from a Data Center Power Hunt

I witnessed this power paradigm shift firsthand at a major data center operator. There, I collaborated with an exceptional electrical design engineer at a Chinese state-owned electrical company, and he was part of a team that designed and implemented a nuclear plant in Guangdong Province, China. He joined as a knowledge powerhouse and team player, and his ability to handle the scale (2023) of multi-building 200MW and upwards data center facilities is very much appreciated.

The Dawn of "Speed-to-Power" as the Ultimate Currency

A $10B AI Data Center's location no longer pivots on tax breaks. Enter 2026: interconnection timelines rule. Can a grid deliver 500MW+ firm baseload today? No? Disqualified. U.S. grids stagger under AI surges doubling data center consumption by 2028.

Hyperscalers pioneer "sovereign power"—private generation, dedicated lines, off-grid islands. Virginia's queues stretch 7-10 years; Texas/Pennsylvania lure with fast plugs into oil/gas/nuclear. My colleague's mantra rang true in design reviews: "A delayed megawatt is a lost billion in compute." Digital ghost towns await solar-heavy laggards.

Nvidia amplifies this: CEO Jensen Huang calls power "the biggest constraint," pushing AI factories overseas—UAE 5GW hubs, Taiwan fabs, SEA pilots—while U.S. stalls. Rivals AMD/Intel lag sovereign hunts.

The SMR Silver Bullet? A 2026 Reality Check

The Small Module Reactors (SMR) Silver Bullet? Promised Quick Fix, Delayed Reality

Small Modular Reactors (SMRs) were billed as the perfect solution—small nuclear plants built in factories like cars, promising faster construction and lower costs than traditional giant reactors. The reality in 2026 is much tougher.

Three major roadblocks:

Regulatory delays: U.S. Nuclear Regulatory Commission (NRC) safety approvals take 5-7 years for new designs
Cost overruns: NuScale's flagship Utah project saw costs triple from $3B to $9B before cancellation in 2023
Fuel shortages: SMRs need special High-Assay Low-Enriched Uranium (HALEU)—U.S. production won't scale until 2030+

Timeline: Estonia targets first SMR grid connection in 2030. Widespread commercial fleets? Post-2035. Meanwhile AI data center demand hits crisis levels by 2028.

Bottom line: Big Tech can't wait 10+ years—they're reviving proven 1970s reactors like Crane instead. SMRs solve tomorrow's problem when today's crisis hits now.

Compare Crane Clean Energy Center (Three Mile Island restart) to delayed SMRs:

Proven technology: Pressurized water reactor design used safely for 50+ years
Already approved: Full safety licensing complete—no regulatory redo needed
Fuel ready: Uses standard nuclear fuel already available globally
Government backed: $1 billion U.S. loan guarantee removes financial risk
Fast timeline: Targets full operation by late 2027—delivering 835 megawatts of reliable 24/7 power when AI data centers hit peak demand

Simple verdict: Restarting Three Mile Island delivers power in 2 years. SMRs take 10+ years. Hyperscalers facing 2028 power cliffs choose proven plants now over futuristic promises.

Southeast Asia's Power Safety Margins Meet AI's Explosive Surge

SEA boasts cushions, but AI erodes fast:

~30% avg fuels 7-10% GDP booms. Data centers quadruple to 10.7GW by 2035 (3-4% peak, 7-10% growth), racks at 106MW density—$10B costs. Nvidia scouts margins pre-firming lags.

If Johor state in Malaysia needs 5.8GW as announced in 4Q 2025 as data center projects underway or in planning statge, it may already erode the peak load safety margin by the next couple of years. as the 5.8 GW "reserve" for data centers in Johor alone would represent nearly 25% of the Peninsular's current average peak load. It is also the same situation for Batam island.

Southeast Asia Can't Afford America's Wake-Up Call

America's hyperscalers reviving retired nuclear plants signals a brutal truth: "speed-to-power" now trumps all other site selection factors. Regions offering only daytime solar—even with generous tax breaks—face exclusion from the AI gold rush.

Southeast Asia faces identical math but lacks nuclear shortcuts:

30% average reserves look comfortable today
Thailand 36-40%, Malaysia 28-38% seem safe through 2030
AI data centers quadruple to 10.7GW by 2035—erasing margins fast

The region's solar success creates a dangerous low reserve safety margin: daytime surplus eroded at evening AI peaks. If the U.S. needs 1970s reactors for firm capacity, how will ASEAN power its digital future urgently?