Hyperscalers are signing twenty-year nuclear deals and outbidding utilities for transmission capacity. The next bottleneck for AI is not chips — it is megawatts.
For most of the last decade, the rate-limiting step for AI was access to high-end GPUs. In 2026, that has quietly changed. The new constraint is electricity, and the scramble to secure it is reshaping power markets across three continents.
Microsoft, Google, Amazon, and Meta have collectively announced more than 40 GW of new data center capacity for 2026 and 2027 — roughly the entire generating capacity of Poland. Yet in Northern Virginia, Dublin, and Singapore, regulators have already paused new connections. Builders are now relocating to wherever they can find a free substation, even if the fiber routes are inconvenient.
The nuclear pivot
The most striking shift is the embrace of nuclear power. Microsoft's twenty-year offtake from Three Mile Island Unit 1 set the template; since then, Amazon, Google, and Oracle have all signed similar deals for small modular reactors that won't come online until 2029 or 2030. Hyperscalers are now effectively financing the next generation of US nuclear plants in exchange for guaranteed clean-energy supply.
What it means for builders
If you operate at small or medium scale, expect cloud inference prices to stop falling. The compounding cost reductions of the 2020–2025 era were largely a chip-and-software story; from here, electricity is the new floor. Teams building latency-sensitive products should also plan for regionally fragmented capacity — your favourite region may simply run out of headroom by 2027.