📰 Key Takeaways

By the end of 2025, global AI data center power capacity reached 29.6 GW, equivalent to New York State’s peak usage. A Stanford University report notes that GPU computing costs have dropped over 99% since 2006, but efficiency gains haven’t translated into power savings—instead, they’ve been poured into training larger models, with grid pressure continuing to rise. Flagship training tasks (like Llama 4 Behemoth) can consume over 100 MW in a single run, equivalent to a small power plant; AI-dedicated compute has surged about 200x in three years, jumping from under 1 GW in 2022 to today, with growth expected to continue through 2030. The US has 5,427 data centers—over 10 times more than any other country; by the end of 2024, cumulative AI power consumption is estimated at 9.4 GW, close to Switzerland’s total electricity usage, about half of Bitcoin mining’s total power draw.

Bitcoin miners, facing a 34%+ price drop in 2026 and JPMorgan’s estimate of total costs at around $78,000 per coin (far above the market price of ~$53,400), with roughly 20% of miners now operating at a loss, are turning their existing power infrastructure—including power contracts, grid access points, and cooling systems—over to AI companies. While ASIC mining chips can’t be used for AI training, the surrounding facilities can be seamlessly repurposed, and mining sites are often located in low-electricity-cost US states like Texas, making the geography highly aligned with AI needs. Iren, for example, signed a five-year GPU cloud contract with Microsoft in November 2025 worth approximately $9.7 billion, powered by a 750 MW facility in Childress, Texas.

💬 JudyAI Lab Perspective

The scale of AI data center power consumption has grown too big to ignore, and the phenomenon of Bitcoin mines converting into compute hubs is reshaping the AI infrastructure layout.

Stanford’s report reveals a counterintuitive structure: every improvement in GPU computing efficiency doesn’t lead to reduced power usage—it gets reinvested into training even larger models, creating a paradoxical cycle where “saving power leads to consuming more.” By the end of 2025, global AI data center power capacity has reached 29.6 GW, equivalent to New York State’s peak usage, and this figure is expected to keep climbing through 2030. For AI builders, the trend isn’t just an environmental concern: the mine-to-AI compute transition (like Iren’s $9.7 billion five-year GPU cloud deal with Microsoft) shows that existing infrastructure with “power, land, and cooling systems” is harder to acquire than the chips themselves—power contracts are becoming the new competitive differentiator. We’re seeing the competitive focus of AI infrastructure quietly shifting from the technology layer to grid access rights.

Next time you’re evaluating AI cloud service providers, try asking one more layer: which grid is their compute connected to, and is their electricity rate stable—this question was redundant five years ago, but it’s now basic due diligence.

📅 Source Info

🔗 Further Reading