Affiliate disclosure: Some links in this article earn us a commission at no extra cost to you. We only recommend hardware we'd run ourselves. Read our full disclosure.
The AI bottleneck has officially migrated from chips to watts. While GPU supply lines stretch through 2027, the quieter story of 2026 is what happened behind the compute: hyperscalers collectively committed to roughly 10 GW of nuclear capacity, Nebius locked itself into a five-year, $17.4 billion (expandable to $19.4B) dedicated-compute agreement with Microsoft, and grid operators from Virginia to New Jersey started treating AI campuses as their largest single class of customer. Megawatts are the new GPU — and that repricing eventually lands on the same place every AI cost does: your electricity bill.
This is the fourth chapter in the shortage chain we've been tracking: the memory price crisis that tripled DRAM, the October 1 Nebius rate hike that pushed on-demand H100 hours to $4.50, and now the power market repricing to match. Every layer of the AI stack is bidding for the same scarce inputs — and when the grid becomes the constraint, energy price discovery reaches further than any GPU launch cycle.
The deals: who bought what
Trackers of the nuclear-datacenter wave estimate that US hyperscalers have signed for roughly 8.5–10 GW of nuclear-linked output — from operating plants, restarts and uprates — with a much larger shadow book of small modular reactor (SMR) milestones behind it. The anchor deals:
| Deal | What it buys | Timeline |
|---|---|---|
| Microsoft × Constellation (Three Mile Island Unit 1) | ~835 MW restart under a 20-year PPA, sold as “Crane Clean Energy Center” | Target 2027–2028 |
| Microsoft × Nebius | ~$17.4B (up to $19.4B) of dedicated GPU compute capacity from Vineland, NJ | Multi-year, from 2025–2026 ramp |
| Google × Kairos Power | SMR fleet deal, first ~500 MW | First units ~2030 |
| Amazon × X-energy | SMR development, 5+ GW target | 2030s |
| Meta × TerraPower / Constellation | Multi-GW nuclear expansion + Clinton (IL) uprate PPA | Mixed 2020s–2030s |
Read the last column carefully: most SMR output arrives in the 2030s. The near-term gigawatts come from restarts and uprates — one running tracker of every nuclear-datacenter deal counts only a few GW deliverable before 2030. That's why the next-best lever — multi-year dedicated-capacity contracts like the Nebius–Microsoft agreement — matters so much: compute sellers are committing output the same way nuclear sellers are, and both moves shrink the pool spot buyers compete for.
Why megawatts are the new GPU
Three structural reasons, and none of them resolve quickly. First, interconnection queues are measured in years. A finished data-center shell is worthless until the utility energizes it, and new large-load interconnects routinely wait longer than the chips inside them. Second, the electrical supply chain is its own shortage: high-voltage transformers, switchgear and gas turbines now run multi-year lead times — the grid equivalent of HBM allocation. Third, AI load growth is concentrated: a handful of metro regions absorb most of the demand, which strains local grids, wholesale prices and rate cases long before national capacity catches up.
The economics rhyme with the GPU market: when output is contracted years ahead, the un-contracted layer becomes scarce and reprices. That's the mechanism behind October's Nebius hike, now playing out in power — utilities in AI-heavy regions are filing higher industrial tariffs, and regulators are debating who pays for new transmission. Either way, the price signal points up.
What it means for your lab's power bill
Homelab operators feel this in two ways. Directly: an always-on rig is exposed to $/kWh drift. A dual-GPU inference box drawing ~800W runs about 576 kWh/month — roughly $100/month at $0.17/kWh. If your regional tariff rises 20% over the next two years (direction consistent with AI-driven demand), that's an extra ~$20/month, every month, on hardware you already own. Over a typical 4–5 year card life, tariff drift can approach the cost of the GPU itself — a quiet TCO line almost nobody models.
Indirectly: every GPU-sized purchase is now coupled to energy scarcity. The cloud side already repriced, and the hardware side is exposed through the memory crisis. The 2026 hardware tier guide increasingly reads as a watts-per-token document, not a dollars-per-VRAM one.
- Know what your lab actually draws: a metered power strip gives you the real number in a week. Check metered power strips & energy monitors on Amazon →
- Cut idle burn on unattended services: an efficient Proxmox mini-PC holds many labs' service layer at a fraction of an idling GPU rig's draw. Shop efficient mini PCs for Proxmox on Amazon →
- Protect the load you already run: AI-season grid stress means more sag events, not fewer. Check UPS pricing on Amazon →
Four plays to run this quarter
- Measure before you optimize. A $15 metered strip turns guesses into numbers. Most labs discover idle burn they didn't know about — a GPU rig idling at 150W is $19/month at $0.17/kWh doing nothing.
- Use time-of-use if you have it. Move batch indexing, backups, and model quantization to off-peak windows. On peak/off-peak differentials of 2–3×, scheduling is free money.
- Right-size the service layer. A modern NPU-equipped mini PC serves DNS, containers and lightweight AI (think the firewall boxes we tested, plus whisper-class models) at 15–50W — reserve the big rig for the workloads that justify it.
- Buy power protection before peak season. Utility stress events track heat waves and, increasingly, AI load ramp events. A decent line-interactive or double-conversion UPS is cheaper than one surge-damaged controller board.
Wired your lab to meter wattage yet? Power protection + measurement is the cheapest upgrade of 2026:
The bottom line: 2026 is the year AI stopped being a compute industry and started being an energy industry. Hyperscalers bid directly for reactors and decades of dedicated output; neoclouds committed billions of dollars of capacity; and grid operators plan transmission around model schedules. Consumers and homelabbers can't sign PPAs — but we can meter, schedule, right-size and protect. In an AI season where megawatts are the new GPU, wattage discipline is the last cheap optimization left.
FAQ
How much nuclear capacity have hyperscalers actually committed to? Trackers put the total at roughly 8.5–10 GW of committed output (restarts, uprates and PPAs like Three Mile Island's 835 MW for Microsoft), plus SMR options — Google–Kairos, Amazon–X-energy, Meta–TerraPower — that mostly deliver in the 2030s.
What was the Nebius–Microsoft deal about? Announced September 8, 2025: a five-year agreement worth about $17.4 billion (expandable to $19.4B) for dedicated GPU capacity from Nebius, anchored by its Vineland, NJ data center — the template for hyperscaler-side multi-year capacity contracts.
Will AI data centers raise my home electricity bill? Not line-itemed, but AI load concentration lifts industrial tariffs and puts regional rate pressure in buildout areas. For a 24/7 rig, a 20% tariff rise is ~$20/month on a typical 576 kWh/month dual-GPU lab — significant TCO drift over a card's life.
What can homelabbers do about it? Measure with a metered strip, shift batch jobs off-peak, right-size the service layer with efficient mini PCs, and buy UPS protection before AI-season grid events multiply.
Disclosure: This post contains affiliate links. We may earn a commission on purchases at no extra cost to you. See our affiliate disclosure for details.