The landscape · Energy and data-centre power
Energy and data-centre power: what AI is really doing
Brief me in 2 minutes
AI in Energy and data-centre power.
AI data centers are now big enough that grids treat them like their own weather system: Texas's connection queue for new large loads sits at 474 GW, about 90% of it data centers, and a Virginia senator wants a three-year statewide pause on approving new ones. The latest nuclear deals split in two. The ones that move electrons first are 20-year power purchase deals on running plants, 890 MW for Google and 690 MW for Amazon, of which 190 MW is new uprate capacity due between 2030 and 2032, while the new-build ones are still 'exploring', such as a 960 MW Indiana plan with no timing and a Utah proposal for 456 reactors. Software from startups like Emerald AI and Karman squeezes 25% or more out of capacity that is already built.
The game: we show you 6 things happening. You call each one real now, early or mostly hype. Then the receipts.
Shift 1 of 6
Grid rules get rewritten mid-buildout, state by state
Real now
California signed seven laws and Texas rewrote its interconnection rulebook in the same September, that is enacted policy responding to demand already on the ground; Virginia's three-year pause is only a filed bill, so far.
How it works. Data center campuses now request hundreds of megawatts in a single interconnection application, more than most grids were built to add at once. Texas's queue for new large loads reached 474 GW, roughly 90% of it data centers, so the state paused new grid hookups to audit whether the number is even real.
Overstated: Executives call permitting delay bureaucratic drag on the AI boom. Texas's own numbers say otherwise: 90% of its 474 GW interconnection queue is data centers, which is why the state itself hit pause to check the figure.
Shift 2 of 6
Data centers learn to flex instead of demanding constant peak power
Early
Deployed commercially, per the company, but still counted in single-digit demonstration sites against a data center fleet in the hundreds, this is proof of concept moving fast, not yet standard practice.
How it works. Instead of building a data center to draw its maximum power nonstop, a control layer throttles GPU clusters down for minutes or hours whenever the local grid is stressed, then ramps back up. Emerald AI's Conductor software proved it could cut a 256-GPU cluster's power draw by 25% over three hours during an Arizona grid event without dropping compute quality, and NVIDIA, Google, Anthropic and utilities including AES, Constellation and Vistra now plan to build this flexibility into new AI factories from day one.
Overstated: Emerald AI's own release calls its software deployed commercially at multi-megawatt, full data center scale. That is true at five demonstration sites with friendly utility partners, not yet the default setting for the wider fleet of data centers still built to draw fixed, always-on power.
Shift 3 of 6
AI labs pay for reactor uprates now and 'explore' new reactors for later
Early
Uprate contracts on existing plants are signed, but the Indiana project has no stated timing, and Valar's 2028 first-reactor target sits in an application the Bureau of Land Management's Utah office says it is still reviewing for completeness.
How it works. Google and Amazon have signed 20-year contracts for specific megawatts of nuclear power, Google's from multiple Constellation-owned plants and Amazon's from Calvert Cliffs, but the contracts that deliver power soonest are on plants that already run. Google signed a 20-year agreement for 890 MW from Constellation nuclear plants that would be upgraded, and Amazon signed 690 MW from the Calvert Cliffs plant in Maryland, including a 190 MW uprate.
Overstated: The new-build headlines are mostly exploration and proposals: AEP's announcement gives no project timing, and Valar's 456-reactor plan is, per NPR, set out in a proposal to federal regulators, not a permit.
Shift 4 of 6
Gas turbines are the fast lane to power AI, until regulators start asking who pays
Real now
Vistra and Constellation are signing contracts for real megawatts, not preorders. The regulatory pushback in North Carolina shows the friction is now in permitting the next wave, not in whether gas already works.
How it works. While nuclear and geothermal projects take years to permit, gas plants can be built or repurposed much faster, so developers are locking down gas capacity directly next to data center sites. New Era signed a 20-year power deal to run its 250MW Texas data center off Vistra's existing gas plant next door, and Constellation is paying $715 million for a 609MW Rhode Island gas plant to expand its footprint.
Overstated: Gas gets sold as the pragmatic bridge fuel while nuclear and geothermal mature. North Carolina's regulator just said the actual bridge, a 255MW turbine, was not worth its staggering cost given uncertain demand, an early sign state regulators will not rubber stamp every gas request AI developers bring them.
Shift 5 of 6
Geothermal gets its first gigawatt-scale AI contract
Early
The PPA is signed and Fervo is a public, revenue-generating company, but the specific capacity Google is buying here will not come online until 2028 at the earliest.
How it works. Enhanced geothermal drilling, borrowed from oil and gas fracking techniques, can deliver round-the-clock carbon-free power without waiting on a reactor's decade-long permitting process. Fervo Energy signed a 396MW power deal with Google to feed a planned Utah data center, with an option to expand to nearly 1 GW by 2030, building on a smaller 115MW deal the two signed in 2024.
Overstated: Geothermal gets pitched as an already-proven alternative to nuclear waiting games. Fervo's own contract is not due online until 2028, and the option to nearly triple it to 1 GW is not due until 2030, so this remains a very well-funded construction project, not power on the grid yet.
Shift 6 of 6
AI weather models quietly become the forecast billions of people already see
Real now
This already ships by default inside Search, Maps and Gemini worldwide, not a research demo waiting for a partner to adopt it.
How it works. Google DeepMind's WeatherNext models have replaced older, physics-only supercomputer simulations as the base layer behind Search, Maps and Gemini weather. The newest version, WeatherNext 3, ingests live satellite data every hour instead of every six, producing forecasts roughly five times sharper than its predecessor, and independent evaluator Brightband confirmed it is now the most accurate model on the market for 2-meter temperature on 26 of the last 30 days tested.
Overstated: Google's launch post reads like consumer forecasting has replaced the need for official weather services. The same page tells users to check their local meteorological agency or national weather service for actual severe weather warnings, because a sharper phone forecast is not a certified warning system.
The scary part
What happens to jobs in Energy and data-centre power
What AI is already taking:
- Deciding, minute to minute, how much power an AI data center cluster is allowed to draw during a grid stress event NVIDIA ↗
- Drafting the base weather forecast that used to come from a meteorologist running a physics simulation Google ↗
- Real-time rack-level power monitoring and control inside a data center that used to run on fixed engineering margins Karman ↗
What stays human:
Deciding whether a reactor, a rate case or a gas turbine is actually safe and fair to ratepayers still runs through human regulators, courts and licensing boards, as North Carolina's regulators just proved by rejecting a gas turbine on cost grounds. No AI model signs a 20-year power purchase agreement or accepts liability for a grid failure.
Get better · 15 minutes
The ability to read and negotiate energy contracts and interconnection filings, the paperwork that actually decides who gets power and when, stays scarce and human no matter how good the forecasting or dispatch software gets.
Try this this week:
Pull up PJM's or ERCOT's public interconnection queue dashboard and find a data center project near you, then ask an AI assistant like Gemini or ChatGPT to explain in plain English what stage it is at and what happens next. That is the same queue the deals in this briefing are fighting over.
Your move
The open lane is software that makes existing power capacity flexible or visible, in the model of Emerald AI and Karman, cheaper to build than a reactor and it ships in months, not years. Reactor and fusion hardware itself is already crowded: Kairos, X-energy, Aalo Atomics, Commonwealth Fusion and Helion are all chasing the same handful of hyperscaler customers.
The signal to watch is whether a deal has an actual in-service date attached; most new-reactor headlines are dated 2027 to 2034 or carry no date at all, like the 960 MW Indiana plan. The red flag to probe in a pitch is a gigawatts-under-exclusivity number with no interconnection filing behind it: Elementl has only put 600MW of its claimed 1.5GW plant into PJM's actual queue so far, and Valar's 456-reactor Utah plan is a proposal, not a permit.
If you work at a utility or a large power user, test flexible load curtailment software this quarter the way Salt River Project and Portland General Electric already have with Emerald AI. A demonstrated 25% cut over three hours is a cheaper first move than requesting a new gas interconnection.
The valuable skill is reading a power purchase agreement and an interconnection queue filing, not just a model card. Utilities and AI labs are hiring people who can tell a signed PPA with a near-term delivery date from one that is really a press release for the 2030s. Start by pulling one recent PPA or interconnection filing from a public company like Constellation, Vistra or Fervo.
What's actually changing
- 1
Grid rules get rewritten mid-buildout, state by state
How it works. Data center campuses now request hundreds of megawatts in a single interconnection application, more than most grids were built to add at once. Texas's queue for new large loads reached 474 GW, roughly 90% of it data centers, so the state paused new grid hookups to audit whether the number is even real. California went the other way and signed seven data center bills in September that, per Utility Dive, require data centers to "pay their fair share of grid update costs" and disclose water use, and in Virginia a state senator filed a bill that would pause approvals for new data centers for three years from 1 July.
Why it matters. Whoever writes the interconnection queue rules decides which AI company gets power first and which household absorbs the bill for new transmission, so a permitting vote in Austin or Sacramento now moves faster than any GPU order.
What's overstatedExecutives call permitting delay bureaucratic drag on the AI boom. Texas's own numbers say otherwise: 90% of its 474 GW interconnection queue is data centers, which is why the state itself hit pause to check the figure.
ERCOTCalifornia Public Utilities CommissionPJM InterconnectionVirginia General Assembly
3 receipts
- Texas's pending data center interconnection requests total 474 GW, about 90% of the state's entire queue, prompting a state audit.
the state audits the Electric Reliability Council of Texas' 474 GW interconnection queue, approximately 90% of which is data centers
Utility Dive, 21 Sept 2026 ↗ - California Gov. Gavin Newsom signed seven data center bills that, per Utility Dive, require data centers to "pay their fair share of grid update costs" and disclose water use to local governments.
The new laws require data centers “pay their fair share of grid update costs,” disclose water use to local governments and give up eligibility for blanket environmental exemptions
Utility Dive, 22 Sept 2026 ↗ - A Virginia state senator filed a bill for the upcoming General Assembly session that would pause approvals of new data centers in the Commonwealth for three years.
A state senator representing Virginia's data center alley in Loudoun County has filed a bill for the upcoming General Assembly session that would temporarily pause any new data centers from being approved in the Commonwealth.
WTVR, 15 Oct 2026 ↗
- Texas's pending data center interconnection requests total 474 GW, about 90% of the state's entire queue, prompting a state audit.
- 2
Data centers learn to flex instead of demanding constant peak power
How it works. Instead of building a data center to draw its maximum power nonstop, a control layer throttles GPU clusters down for minutes or hours whenever the local grid is stressed, then ramps back up. Emerald AI's Conductor software proved it could cut a 256-GPU cluster's power draw by 25% over three hours during an Arizona grid event without dropping compute quality, and NVIDIA, Google, Anthropic and utilities including AES, Constellation and Vistra now plan to build this flexibility into new AI factories from day one.
Why it matters. This is the fastest way to add AI compute without waiting a decade for new transmission, and one widely cited estimate says flexing load for less than 200 hours a year could unlock 100 gigawatts of grid capacity that already exists, capacity worth more than $2 trillion in avoided data center investment.
What's overstatedEmerald AI's own release calls its software deployed commercially at multi-megawatt, full data center scale. That is true at five demonstration sites with friendly utility partners, not yet the default setting for the wider fleet of data centers still built to draw fixed, always-on power.
Emerald AINVIDIAGoogleOracleKarman
3 receipts
- Emerald AI's software cut a 256-GPU cluster's power draw by 25% for three hours during an Arizona grid stress event while preserving compute quality.
reduce the power consumption of AI workloads running on a cluster of 256 NVIDIA GPUs by 25% over three hours during a grid stress event
NVIDIA, 1 July 2025 ↗ - A Duke University study estimates flexing AI data center load for under 200 hours a year could unlock 100 gigawatts of new grid capacity.
if new AI data centers could flex their electricity consumption by just 25% for two hours at a time, less than 200 hours a year, they could unlock 100 gigawatts of new capacity
NVIDIA, 1 July 2025 ↗ - Emerald AI raised $150 million in a Series A at a $1.05 billion valuation, saying its tech is deployed commercially at multi-megawatt scale.
With five global demonstrations complete, Emerald AI's technology is now deployed commercially, dynamically flexing power consumption at multi-megawatt, full data center scale.
Emerald AI, 25 Aug 2026 ↗
- Emerald AI's software cut a 256-GPU cluster's power draw by 25% for three hours during an Arizona grid stress event while preserving compute quality.
- 3
AI labs pay for reactor uprates now and 'explore' new reactors for later
How it works. Google and Amazon have signed 20-year contracts for specific megawatts of nuclear power, Google's from multiple Constellation-owned plants and Amazon's from Calvert Cliffs, but the contracts that deliver power soonest are on plants that already run. Google signed a 20-year agreement for 890 MW from Constellation nuclear plants that would be upgraded, and Amazon signed 690 MW from the Calvert Cliffs plant in Maryland, including a 190 MW uprate. New reactors are at the 'exploring' stage: Amazon, AEP and X-energy are looking at a 960 MW Indiana project with 12 small reactors and no stated timing, and Valar Atomics has proposed 456 reactors for Utah.
Why it matters. Uprates and extensions add capacity on existing sites with existing licences, so they are the nuclear money that can show up on a grid soonest (Amazon's 190 MW uprate is due between 2030 and 2032), while the new-build projects have to survive permitting, financing and a first-of-a-kind build.
What's overstatedThe new-build headlines are mostly exploration and proposals: AEP's announcement gives no project timing, and Valar's 456-reactor plan is, per NPR, set out in a proposal to federal regulators, not a permit.
GoogleConstellation EnergyAmazonAmerican Electric PowerX-energyValar AtomicsKairos Power
5 receipts
- Google signed a 20-year agreement for 890 MW from multiple Constellation nuclear plants that would be upgraded.
In the latest arrangement, Google entered into a 20-year power purchase agreement for 890 megawatts of electricity from multiple Constellation-owned nuclear power plants, which would be upgraded to produce more electricity from existing reactors.
The Star (Bloomberg), 6 Oct 2026 ↗ - Amazon will receive 690 MW from Calvert Cliffs under a 20-year deal that includes a 190 MW uprate due between 2030 and 2032.
The deal will also enable around 190 MW of new generating capacity, expected to come online between 2030 and 2032. Through the agreement, Amazon will receive 690 MW of power from the nuclear plant, including the 190 MW uprate.
ESG Dive, 30 Sept 2026 ↗ - Amazon, AEP and X-energy are exploring a 960 MW Indiana nuclear project and gave no timing.
The facility would utilize 12 of X-energy’s Xe-100 small modular reactors, each capable of producing 80 MW. The companies did not give details around the project’s timing.
Utility Dive, 19 Oct 2026 ↗ - NPR reports Valar's Project Beehive timeline calls for the first reactors to come online sometime in 2028.
The project's timeline calls for nonnuclear construction to begin as soon as the end of this year, with the first reactors coming online sometime in 2028.
NPR, 30 Sept 2026 ↗ - The Bureau of Land Management's Utah office told NPR it is still reviewing Valar's application for completeness.
We are currently reviewing the application for completeness,
NPR, 30 Sept 2026 ↗
- Google signed a 20-year agreement for 890 MW from multiple Constellation nuclear plants that would be upgraded.
- 4
Gas turbines are the fast lane to power AI, until regulators start asking who pays
How it works. While nuclear and geothermal projects take years to permit, gas plants can be built or repurposed much faster, so developers are locking down gas capacity directly next to data center sites. New Era signed a 20-year power deal to run its 250MW Texas data center off Vistra's existing gas plant next door, and Constellation is paying $715 million for a 609MW Rhode Island gas plant to expand its footprint. But not every gas request survives contact with a regulator: North Carolina just rejected Duke Energy's proposed 255MW gas turbine as too costly with uncertain need.
Why it matters. Gas is currently the power source most able to match AI's construction timeline, but every megawatt locked into a 20-year contract is a megawatt regulators cannot easily reroute if cheaper or cleaner capacity shows up later, and ratepayers often end up guaranteeing the plant's return.
What's overstatedGas gets sold as the pragmatic bridge fuel while nuclear and geothermal mature. North Carolina's regulator just said the actual bridge, a 255MW turbine, was not worth its staggering cost given uncertain demand, an early sign state regulators will not rubber stamp every gas request AI developers bring them.
VistraConstellation EnergyDuke EnergyNew EraShell
3 receipts
- New Era signed a 20-year gas power deal to run its 250MW Ector County, Texas data center directly off an adjacent Vistra gas plant.
New Era signs 20-year gas PPA with Vistra subsidiary to power 250MW data center in Ector County, Texas
Data Center Dynamics, 22 Sept 2026 ↗ - Constellation Energy is paying $715 million to acquire a 609MW Rhode Island gas plant from Shell.
Constellation Energy will acquire the 609-MW RISEC from Shell Energy North America for $715 million, the companies announced Sept. 10, 2026.
Utility Dive, 14 Sept 2026 ↗ - North Carolina regulators denied Duke Energy's proposed 255MW gas turbine, citing staggering cost and uncertain need.
denied Duke Energy's proposal to construct and operate a 255-MW natural gas-fueled simple cycle combustion turbine at its Smith Energy Complex in Richmond County, North Carolina, citing uncertainty around load growth and the cost of the proposal.
Utility Dive, 22 Sept 2026 ↗
- New Era signed a 20-year gas power deal to run its 250MW Ector County, Texas data center directly off an adjacent Vistra gas plant.
- 5
Geothermal gets its first gigawatt-scale AI contract
How it works. Enhanced geothermal drilling, borrowed from oil and gas fracking techniques, can deliver round-the-clock carbon-free power without waiting on a reactor's decade-long permitting process. Fervo Energy signed a 396MW power deal with Google to feed a planned Utah data center, with an option to expand to nearly 1 GW by 2030, building on a smaller 115MW deal the two signed in 2024.
Why it matters. This is the first geothermal deal sized like the nuclear and gas deals AI labs sign, proof that a power source running around the clock without fuel shipments or reactor safety review can compete for the same contracts.
What's overstatedGeothermal gets pitched as an already-proven alternative to nuclear waiting games. Fervo's own contract is not due online until 2028, and the option to nearly triple it to 1 GW is not due until 2030, so this remains a very well-funded construction project, not power on the grid yet.
Fervo EnergyGoogle
2 receipts
- Fervo Energy signed a 396MW PPA with Google for a Utah geothermal project expected online in 2028.
announced a 396-megawatt (MW) power purchase agreement (PPA) with Google to enable the continued development of the Cape Station enhanced geothermal systems (EGS) GeoCluster, expected to come online in 2028
Fervo Energy, 1 Sept 2026 ↗ - Google has an option to expand the deal to nearly 1 GW by 2030, building on a 115MW PPA the two signed in 2024.
Fervo signed a 115 MW PPA with Google and NV Energy in June 2024 that helped pioneer the Clean Transition Tariff
Fervo Energy, 1 Sept 2026 ↗
- Fervo Energy signed a 396MW PPA with Google for a Utah geothermal project expected online in 2028.
- 6
AI weather models quietly become the forecast billions of people already see
How it works. Google DeepMind's WeatherNext models have replaced older, physics-only supercomputer simulations as the base layer behind Search, Maps and Gemini weather. The newest version, WeatherNext 3, ingests live satellite data every hour instead of every six, producing forecasts roughly five times sharper than its predecessor, and independent evaluator Brightband confirmed it is now the most accurate model on the market for 2-meter temperature on 26 of the last 30 days tested.
Why it matters. Forecasts are the invisible infrastructure behind flight delays, farming decisions and disaster planning, and control of that layer is consolidating into whichever lab's model wins the accuracy leaderboard this month, not staying with the national weather agencies that used to own it.
What's overstatedGoogle's launch post reads like consumer forecasting has replaced the need for official weather services. The same page tells users to check their local meteorological agency or national weather service for actual severe weather warnings, because a sharper phone forecast is not a certified warning system.
Google DeepMindBrightbandNOAA
3 receipts
- WeatherNext 3 ingests live satellite data hourly and produces forecasts roughly five times sharper than its predecessor.
Overall, this provides a global weather picture roughly five times sharper than our previous model, WeatherNext 2, which produced forecasts on a 25-kilometer grid in 6-hour increments.
Google, 3 Sept 2026 ↗ - Independent evaluator Brightband found WeatherNext 3 had the lowest 2-meter temperature forecast error on 26 of the last 30 days.
WeatherNext 3 had the lowest error for 26 of the last 30 days.
Brightband, 2 Sept 2026 ↗ - Google's own launch page tells users to rely on local meteorological agencies for official severe weather warnings, not the AI model.
For official weather forecasts, severe weather warnings, and public safety advisories, please refer to your local meteorological agency or national weather service.
Google, 3 Sept 2026 ↗
- WeatherNext 3 ingests live satellite data hourly and produces forecasts roughly five times sharper than its predecessor.
Who's doing it
12 companies to know, 7 with a round we can source. Newest first where dated.
Search these companies in the startup database ↗ · Explore all research tools
Commonwealth Fusion Systemsnuclear fusion power
Building SPARC, a fusion demonstration machine, and ARC, its first planned commercial fusion power plant, aimed partly at AI-era power demand.
Latest · 23 Oct 2026CFS said it would receive $85 million from the US government to develop advanced chips for fusion power. TechCrunch ↗
$1B (total raised $4B) Additional equity financing, 30 July 2026 source ↗
Helion Energynuclear fusion power
Fusion power startup building its Orion plant, with a power deal with Microsoft targeting first commercial electricity later this decade.
Latest · 22 Oct 2026Helion signed a letter of intent for up to $175 million in CHIPS funding for fusion power semiconductors. Helion ↗
$465M Series G, 4 June 2026 source ↗
X-energysmall modular reactors and HALEU fuel
Builds Xe-100 small modular reactors and their fuel supply chain, with commercial projects underway for Amazon, Dow and Centrica.
Latest · 19 Oct 2026Amazon, AEP and X-energy are exploring a 960 MW Indiana project built from 12 Xe-100 reactors, with no timing given. Utility Dive ↗
Kairos Powersmall modular nuclear reactors
Building a 50MW demonstration molten-salt reactor in Tennessee under a long-term deal to supply Google's data centers.
Latest · 16 Oct 2026Kairos lined up $643 million in private and government financing for its Tennessee demonstration reactor. Neutron Bytes ↗
$643M Private ($368M) and US Energy Department ($275M) financing, 16 Oct 2026 source ↗
Valar Atomicsmodular nuclear microreactors
Nuclear startup proposing 'Project Beehive', a data center campus with some 456 small reactors on Bureau of Land Management land near Price, Utah.
Latest · 30 Sept 2026Proposed 'Project Beehive' on federal land in Utah, with some 456 small reactors, in a proposal to federal regulators. NPR ↗
Crusoe Energyvertically integrated AI infrastructure and power
Builds and powers its own AI data centers end to end, increasingly using stranded gas and now nuclear microreactors.
Latest · 17 Sept 2026Crusoe says it has over $140 billion in total contracted value across its vertically integrated AI infrastructure platform. Crusoe ↗
$3.9B Series F, 17 Sept 2026 source ↗
Karman (formerly Utilidata)AI power orchestration for data centers and grids
NVIDIA-backed software and hardware that senses and controls power draw at the rack level, so AI data centers can squeeze more compute out of the power they already have.
Latest · 15 Sept 2026Utilidata rebranded as Karman as AI data centers became the frontline of the power grid, expanding its board with executives from Verrus and VoltServer. Karman ↗
$40M extension ($100M total Series C) Series C, 28 May 2026 source ↗
BrightbandAI weather and climate forecasting
Builds AI weather forecasting and data tools, and independently benchmarks models like Google's WeatherNext against each other and against reality.
Latest · 2 Sept 2026Brightband's independent benchmark found Google's new WeatherNext 3 model has the lowest 2-meter temperature forecast error on 26 of the last 30 days. Brightband ↗
Fervo Energyenhanced geothermal power
Drills enhanced geothermal wells for round-the-clock carbon-free power, now selling gigawatt-scale capacity directly to AI data center developers.
Latest · 1 Sept 2026Fervo signed a 396MW PPA with Google for a Utah geothermal-powered data center project expected online in 2028. Fervo Energy ↗
Emerald AIgrid flexibility software
Software that throttles AI data center power draw during grid stress events, then ramps back up, so data centers connect to the grid faster.
Latest · 25 Aug 2026Emerald AI says its grid-flexibility approach could unlock over 100 gigawatts of capacity on the existing US grid. Emerald AI ↗
$150M Series A, 25 Aug 2026 source ↗
Elementl Powernuclear project development
Develops sites for new nuclear plants and lines up utility and hyperscaler customers before a reactor vendor is even chosen.
Latest · 21 Aug 2026Elementl submitted the first 600MW of a planned 1.5GW nuclear plant in Ohio into PJM's interconnection queue. New Project Media ↗
Aalo Atomicsmodular nuclear microreactors
Builds small modular nuclear reactors (XMRs) designed to sit next to and directly power AI data centers.
Latest · 30 July 2026Crusoe and Aalo Atomics are partnering to deploy the first nuclear-powered AI factory data center. Crusoe ↗
over $136M Cumulative, multiple investors, 30 July 2026 source ↗
Your move
Our read, built on the receipts above.
The open lane is software that makes existing power capacity flexible or visible, in the model of Emerald AI and Karman, cheaper to build than a reactor and it ships in months, not years. Reactor and fusion hardware itself is already crowded: Kairos, X-energy, Aalo Atomics, Commonwealth Fusion and Helion are all chasing the same handful of hyperscaler customers.
The signal to watch is whether a deal has an actual in-service date attached; most new-reactor headlines are dated 2027 to 2034 or carry no date at all, like the 960 MW Indiana plan. The red flag to probe in a pitch is a gigawatts-under-exclusivity number with no interconnection filing behind it: Elementl has only put 600MW of its claimed 1.5GW plant into PJM's actual queue so far, and Valar's 456-reactor Utah plan is a proposal, not a permit.
If you work at a utility or a large power user, test flexible load curtailment software this quarter the way Salt River Project and Portland General Electric already have with Emerald AI. A demonstrated 25% cut over three hours is a cheaper first move than requesting a new gas interconnection.
The valuable skill is reading a power purchase agreement and an interconnection queue filing, not just a model card. Utilities and AI labs are hiring people who can tell a signed PPA with a near-term delivery date from one that is really a press release for the 2030s. Start by pulling one recent PPA or interconnection filing from a public company like Constellation, Vistra or Fervo.
What's next
- 10 Dec 2026 Whether ERCOT and the Texas PUC file their reports on the governor's data center interconnection audit by 10 December, as they said in August they hoped to insideclimatenews.org ↗
- Jan 2028 Fervo Energy's Cape Station geothermal GeoCluster in Utah, backing Google's 396MW PPA, is expected to come online fervoenergy.com ↗
- Jan 2028 Helion's Orion plant faces a 2028 deadline to deliver grid electricity to Microsoft techcrunch.com ↗
Other industries
- FinanceBanks move agentic AI from chatbots to running back-office and dealmaking work
- Software engineeringSpaceX owns Cursor and starts rewiring it: one Grok bundle, and OpenAI's models on the way out
- HealthcareAmbient scribes go from pilot project to clinical infrastructure
- LegalLegal AI vendors go from research assistant to agent that produces the deliverable
- Real estateLeasing AI graduates from answering the phone to acting inside the property system
- Sales and marketingCustomers start sending their own AI agents, and vendors race to build the front door for them
- Jobs, HR and recruitingAI runs the first-round interview, not just the resume filter
- InsuranceClaims move from adjuster-first to AI-first, humans supervise the exceptions
- CybersecurityAttackers hand whole intrusion campaigns to AI agents, at about $50 a breach
- Consulting and accountingAI-written client reports become a recurring, public Big Four embarrassment
- Retail and e-commerceAgent checkout retreats to a cart, and the banks guard the last tap
- EducationNYC and LA ban AI for students, not for teachers
- Media and creativeAI video's money goes to the app that resells the models, not the lab that trains them
- Manufacturing and roboticsHumanoid robots move from demo livestreams to paid warehouse shifts
- Transport and robotaxisRobotaxis go from demo to metered fare, and Waymo starts borrowing to scale
- GovernmentGSA turns AI adoption into a government-wide subscription, then swaps 'free' for metered contracts
- Agriculture and foodDriverless tractors learn to run the tools behind them, through open platforms and retrofit kits