Electrons are the new GPUs
AI stopped being limited by chips and started being limited by power. This decade, speed to electricity is the product, and most of it will not come from new reactors.
Neil Gaikwad
The size of the prize
What the consensus misses
The consensus trade is nuclear: small modular reactors, fusion, restarts. Restarts are real, but most new reactors arrive after 2030. NuScale's Utah project was cancelled when its power price rose to $89 a megawatt-hour, and the most-watched advanced reactor company still holds only a materials license. If your thesis needs new reactors before 2030, you are early by a cycle.
What I think the market underprices is time. Turbines are sold out for years, the grid queue runs past five years, and a gigawatt of compute sitting dark costs more than the power itself. So the product people will pay for is speed to electricity: generation on the site, batteries, power electronics that skip transformer lead times, and software that finds spare capacity in the grid already built.
The other underrated asset is flexibility. Texas now requires new loads over 75 MW to accept curtailment in emergencies, and FERC has told every grid operator to justify how it connects large and flexible loads. A data center that can turn itself down gets connected sooner, which makes flexibility worth real money for the first time.
Why now
The buyers have committed. Microsoft is restarting Three Mile Island Unit 1 for 2028, Amazon contracted 1,920 MW from Talen through 2042, Google signed for up to 3 GW of geothermal from Fervo, and Meta lined up as much as 6.6 GW of nuclear.
The rules are being rewritten right now. FERC ordered PJM to create co-location services in December 2025 and, in June 2026, ordered all six regional grid operators to justify or reform their large-load rules.
Prices are rising in public. Residential power is forecast at 18.2 cents a kWh in 2026, up from 17.3 cents, and ratepayer politics will push data centers to bring their own power.
Where the value pools
| Layer | My call | Why |
|---|---|---|
| Firm generation | Incumbents keep it | Gas turbines, nuclear fleets and restarts sit with GE Vernova, Constellation and Vistra, and their backlogs show it. |
| Speed to power at the site | Most value | Behind-the-meter generation, storage and power electronics that get a campus energized in months, not years. |
| Flexibility and grid software | Underpriced | New rules turn curtailment and spare grid capacity into a faster path to connection. |
| Advanced nuclear and fusion | Long-dated | Real, but mostly after 2030; venture returns depend on licensing timelines no one controls. |
| Cooling and the rack | Durable value | Every new GPU generation runs hotter; strategics are already leading rounds here. |
Who is winning so far
| Company | What it does | Stage and last round |
|---|---|---|
| GE Vernova | Gas turbines, grid and electrification equipment | Public; $176B total backlog Q2 2026 · sec.gov |
| Fervo Energy | Enhanced geothermal; Google framework for up to 3 GW | IPO raising about $2.2B May 2026 · telcomagazine.com |
| Crusoe | AI data centers with on-site power; the Abilene Stargate campus | Over $3B at about $30B, reported Sep 2026 · news.bgov.com |
| Commonwealth Fusion | Tokamak fusion; SPARC demo and a first plant in Virginia | $863M Series B2; close to $3B raised Aug 2025 · cfs.energy |
| Valar Atomics | Helium-cooled microreactor; reached criticality in a DOE pilot | $1B Series B led by Sequoia Aug 2026 · canarymedia.com |
| Base Power | Home batteries aggregated as grid capacity in Texas | $1B Series C led by Addition Oct 2025 · news.bgov.com |
| Heron Power | Solid-state transformers for data centers, solar and storage | $140M Series B Feb 2026 · canarymedia.com |
| Emerald AI | Software that flexes data center load to grid needs | $150M Series A at a reported $1.05B Aug 2026 · thenextweb.com |
Where I would write a seed check
Interconnection studies, siting and spare-capacity discovery for a queue that holds 2,061 GW and new rules in every region.
Early examples: GridCARETransformers and switchgear are bottlenecks, and GPUs, batteries and solar all run on DC. Semiconductor-based conversion skips the wait.
Early examples: Heron PowerTurbines are booked for years, so teams that package permits, fuel, storage and modular generation for sub-100 MW sites win the gap years.
Capacity prices are at their caps. Aggregated batteries earn that price; Base Power proved the model in Texas and other regions are open.
Early examples: Base PowerStrategic buyers are leading rounds in direct-to-chip cooling, which tells you where the exits are.
Early examples: AccelsiusHow it plays out
- 2027
Fervo's first commercial geothermal power and the first co-location services under FERC's orders. Both test whether firm clean power can arrive fast.
- 2028
Three Mile Island Unit 1 returns as the Crane Clean Energy Center, and US data center use lands somewhere in Berkeley Lab's 325 to 580 TWh range.
- 2031
The first advanced reactors deliver commercial power for Google and Meta, if licensing holds. My expectation: on-site power is standard for new AI campuses.
- 2036
My expectation: data centers are one of the largest buyers of new generation in the US, and the companies that sold speed have been bought by utilities and equipment makers.
As of Oct 2026
Risks
- Efficiency shocks: on the DeepSeek news Vistra fell nearly 30% and Constellation and GE Vernova more than 20% in a day. nbcsandiego.com
- Small reactors have cost and execution risk: NuScale's Utah project was cancelled at $89/MWh. power-eng.com
- Ratepayer politics: capacity prices at their caps and rising household bills invite rules that slow data centers. publicpower.org
- Regulators can pull the plug: Texas lets the grid operator disconnect 75 MW+ loads in emergencies. utilitydive.com
- Demand rests on a handful of buyers whose guidance moves quarter to quarter. fortune.com
Sources
- IEA, Energy and AI
- Berkeley Lab, Data center electricity demand (Jan 2025)
- APPA on EPRI, Data centers 9 to 17% by 2030 (Feb 2026)
- APPA, Interconnection backlog (Jul 2026)
- APPA, PJM capacity auction (Jul 2026)
- GE Vernova, Q2 2026 results
- Constellation, Crane Clean Energy Center
- TechInformed, Meta nuclear deals (Jan 2026)
- McGuireWoods, FERC large-load orders (Jun 2026)
- EIA, Short-term energy outlook
- Fortune, Amazon capex (Jul 2026)
- Power Engineering, NuScale cancels Utah project
Nothing here is investment advice: these are my own notes on public information, written to sharpen judgment, not to recommend buying or selling anything.