AI Power Weekly

AI Power News 7/27/26

Will McKnight — Mon, 27 Jul 2026 17:45:20 GMT

OpenAI Announces $30 Billion Georgia Data Center With 3.2 GW Power Deal

OpenAI announced plans to spend more than $30 billion on a massive data center complex near Savannah, Georgia, securing 3.2 gigawatts of power from Georgia Power in what may rank as one of the largest single-site AI infrastructure projects to date. The facility, called Project Camellia, will receive several hundred megawatts starting in 2028, with the full buildout extending through 2032. One gigawatt powers roughly 750,000 US homes at any given time, so this single campus could eventually consume as much electricity as 2.4 million homes (Bloomberg).

This marks OpenAI’s first self-built data center, a shift from leasing chip capacity from cloud providers like Oracle and Amazon Web Services. The company also raised its total computing infrastructure spending forecast to $750 billion through 2030, up from $600 billion earlier this year. That increase alone, $150 billion, exceeds the annual GDP of more than 130 countries (Quartz).

The project signals that hyperscalers are moving beyond signing power purchase agreements and into direct ownership of generation and data center infrastructure. OpenAI guaranteed that residential electricity prices in Georgia will not increase and committed to reducing power consumption during periods of high demand before impacting other users. Voluntary pledges have never stopped a rate case before, and this one will get its first real test once Georgia Power’s actual cost allocation filings start showing up.

White House Expands AI Power Pledge to 200 Utilities and 23 States

President Trump expanded the Ratepayer Protection Pledge on July 23, adding nearly 200 utilities, data center developers, and 23 state governors to a voluntary agreement first signed by tech companies in March. The pledge now covers 80 percent of all power delivered to US homes and businesses. Under the agreement, data center operators commit to building, bringing, or buying the new generation resources needed for their facilities and paying for all new power delivery infrastructure, rather than passing costs to residential ratepayers (White House).

The expanded pledge remains non-binding and contains no enforcement mechanism. Tariff rules that govern how grid infrastructure costs are allocated to household bills are set by state public utility commissions and FERC, not by voluntary pledges. The commitment does, however, create political pressure on data center developers at a time when 71 percent of Americans oppose AI data center construction in their local communities, according to a recent Gallup poll (Tech Times).

The White House announcement comes as the Ratepayer Protection Act passed the House Energy and Commerce Committee unanimously on July 22, though it still needs full House and Senate votes before becoming law.

DOE’s Genesis Mission Taps Oklo and X-Energy to Speed Up Nuclear Reactors for AI

The Department of Energy said July 22 that its Genesis Mission, a federal push to apply artificial intelligence across scientific research, is directing a $60 million grant to Idaho National Laboratory, Nvidia, Amazon Web Services, and X-energy to use AI to cut the cost of building nuclear reactors in half and speed up construction timelines. A day earlier, advanced reactor developer Oklo joined a related $200 million DOE effort aimed at shortening how long it takes to design, license, and build new plants (Bloomberg).

Genesis Mission drew more than $800 million in commitments from private partners on top of the federal money, part of a broader effort Energy Secretary Chris Wright says will apply AI to problems ranging from reactor design to materials science across nearly 300 total awards. The bet is that AI can shave years off the decade-plus timelines that make new nuclear plants a hard sell for utilities racing to power AI data centers, and pairing reactor developers directly with government supercomputers is a new model for how Washington is trying to speed that up (HPCwire).

Rolls-Royce SMR Adds Two More Czech Sites for New Reactors

The Czech Ministry of Industry and Trade, national utility ČEZ, and Rolls-Royce SMR signed a memorandum of understanding on July 21 to begin preparatory work on small modular reactors, compact nuclear plants built largely from factory-made parts rather than assembled on site, at two additional locations: the Tušimice site in the Ústí nad Labem region and the Dětmarovice site in Moravian-Silesia. Both sites currently host aging coal plants that the Czech Republic is working to replace with lower-carbon generation (World Nuclear News).

Central and Eastern Europe has become one of the more active regions for early SMR deployment, and this deal extends a partnership between the three parties that is already advancing SMR plans elsewhere in the country. Watch whether the Czech Republic becomes a proving ground that Rolls-Royce can point to when it pitches the same reactor design to customers in the UK and US.

ERCOT Flags Preliminary Record for Texas Summer Power Demand

The Electric Reliability Council of Texas, the operator that keeps the lights on for most of the state, said Tuesday, July 21, that Texas may have set a new preliminary peak demand record, topping the prior high by roughly 2,000 megawatts during a heat wave. ERCOT had already been warning that summer 2026 demand could top 92 gigawatts, one gigawatt powers around 750,000 average US homes, with data centers and cryptocurrency mining adding large, constant loads on top of the usual air conditioning spike (Houston Public Media).

Despite the record, ERCOT put the odds of an actual grid emergency at just 0.21 percent for the month, a reminder that record demand and grid failure are not the same thing. Texas has more data center capacity under construction than any other state, so this kind of record is a preview of what summer peaks will look like for years to come.

India Announces $43 Billion Energy Storage Investment by 2032

India’s Ministry of Power informed Parliament on July 27 that the government plans to invest Rs 3.49 lakh crore ($43 billion) in battery energy storage systems and another Rs 1.29 lakh crore ($16 billion) in pumped storage plants by 2031-32. The plan targets 47.24 GW / 236 GWh of BESS capacity and 26.69 GW / 175 GWh of pumped storage by the end of the decade (ProKerala).

India’s energy storage buildout is driven by renewable integration and the variability of solar generation. The government is backing deployment with $1.09 billion in viability gap funding for 43.2 GWh of BESS capacity, interstate transmission charge waivers for storage co-located with renewables, and production-linked incentives for 50 GWh of advanced chemistry cell battery manufacturing (DC&T Global).

Solar-plus-storage auctions in India have recently cleared at tariffs of $32 to $39 per MWh for four-hour battery systems, and the country’s first six-hour BESS auction in early 2026 came in at $34 per MWh. Those prices are now competitive with new coal power, which is seeing rising tariffs across recent auctions. India’s aggressive storage targets position it as one of the fastest-growing battery markets outside China, which added 167 GWh of storage in 2025 and is expected to add another 203.5 GWh in 2026.

Have fun this week,

Will

Thanks for reading AI Power Weekly! This post is public so feel free to share it.

AI Power News Roundup 7/17/26

Will McKnight — Fri, 17 Jul 2026 18:48:54 GMT

PJM Capacity Auction Adds $6.3 Billion to Electric Bills as Data Centers Drive Demand

PJM Interconnection, the nation’s largest grid operator serving 67 million people across 13 states, released results on July 14 showing its latest capacity auction will add $6.3 billion in costs to ratepayer bills over the next three years. The auction, which secures power supply commitments two years out, hit a total cost of $16.4 billion for the 2028/2029 delivery year, matching the record set in late 2025. Data centers accounted for roughly $6.3 billion of that total, according to Monitoring Analytics, PJM’s independent market monitor. (Philadelphia Inquirer)

The capacity market, which pays power plants to be available during peak demand, exists to ensure grid reliability. When projected demand rises faster than supply, prices spike. PJM’s auction failed for the third consecutive time to procure enough generation to meet its reliability target, falling 6.8 gigawatts short, the equivalent of nearly seven nuclear reactors. (Insurance Journal)

Total data center costs embedded in PJM auctions since 2024 now amount to approximately $29 billion, a sum that gets passed to households and businesses through higher electricity rates. In Washington DC, Pepco residential customers saw bills rise by an average of $21 per month starting in June 2025, with roughly $10 of that increase tied directly to capacity price increases. The numbers underline a structural mismatch: AI infrastructure is scaling at a pace the existing grid cannot match without massive cost increases to everyone else. (Fortune)

Pennsylvania Governor Josh Shapiro negotiated a price cap that prevented even higher costs, but the cap expires in 2030. Until then, PJM ratepayers are funding a build-out they have no control over and receiving none of the revenue AI generates. This is not a local story. Every major grid market faces versions of the same question: who pays when new load arrives faster than new supply?

Virginia Enacts First US Electricity Consumption Tax on Data Centers

Virginia Governor Abigail Spanberger signed the Commonwealth’s 2026 budget on June 30, creating the first state tax in the US charged directly on data center electricity consumption. The tax, which took effect July 1, charges $0.011 per kilowatt-hour on all power consumed by data centers, whether supplied by utilities or generated on-site. Legislative estimates project the tax will generate $600 million annually for Virginia’s general fund over the next two years. (Kiplinger)

The tax applies to both grid power and behind-the-meter generation, including solar and natural gas turbines, closing the loophole data center operators might otherwise exploit. If annual collections exceed $600 million, the excess gets refunded to data centers on a pro rata basis. The tax sunsets on July 1, 2028, giving Virginia a two-year experiment in metered revenue extraction from the largest power users in the state. (Williams Mullen)

A 500 MW data center running continuously would owe roughly $48 million per year, a 10 percent increase in effective electricity costs. Virginia preserved its sales tax exemption on data center equipment while adding this consumption charge, a compromise that keeps the state competitive for new projects while clawing back some revenue through operational costs. The tax does not discourage data centers, it just ensures the state extracts value from the electrons they burn. Watch for other states to copy this model. (Data Center Knowledge)

New York Enacts First State Moratorium on Hyperscale Data Centers

Governor Kathy Hochul signed an executive order on July 14 pausing state environmental permits for one year on new “hyperscale” data centers, defined as facilities that draw 50 megawatts or more of power, enough to run roughly 40,000 homes. The order directs Empire State Development to publish a Community Investment Framework within 60 days, giving local governments a template for negotiating benefits before new projects break ground. New York becomes the first state to freeze hyperscale data center permitting outright, rather than simply studying the grid strain. (NY Governor’s Office)

The move follows months of rising electricity bills across the Northeast tied to new data center demand, the same dynamic driving PJM’s cost spike above. Hochul framed the pause as protecting ratepayers and the environment while New York builds a permanent regulatory framework instead of approving projects case by case. Watch whether other states under similar grid strain, from Georgia to Ohio, follow with pauses of their own. (CNBC)

White House Plans to Expand Ratepayer Protection Pledge to Utilities

The White House is organizing an event in the coming weeks to expand its Ratepayer Protection Pledge, first signed in March by Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI, to include electric utilities, data center developers, and state governors, according to sources cited July 13. The original pledge commits signatories to cover the cost of new power plants, grid upgrades, and unused reserved capacity themselves, instead of passing those costs to existing electricity customers. A White House official said the pledge “has been so impactful that additional stakeholders also want to sign it.” (US News)

Pulling utilities into the pledge would mark a shift from voluntary tech company commitments toward something closer to an industry norm. The White House was still finalizing the guest list and event date as of this writing. If utilities sign on, it could become the template other states point to when writing their own data center cost allocation rules, including New York’s new moratorium above. (Virginia Business)

Four US Microreactors Achieve Criticality by July 4 Deadline

Four advanced nuclear microreactors reached criticality, the point where a reactor sustains a self-supporting chain reaction, by July 4, 2026, exceeding President Trump’s goal of three reactors under the Department of Energy’s Reactor Pilot Program. Aalo Atomics became the fourth company to hit the milestone early July 4 at Idaho National Laboratory, following Antares Nuclear on June 4, Valar Atomics on June 22, and Deployable Energy on July 1. (World Nuclear News)

Aalo’s Critical Test Reactor went from groundbreaking to criticality in less than eight months, one of the fastest reactor builds in modern American history. The reactor includes a full-scale core demonstrating the nuclear components of Aalo’s 10 MWe reactors, which will be deployed in 50 MWe Aalo Pods designed specifically to power AI data centers. The company has already begun work on a second reactor at INL, dubbed Project Ascension, which will produce electricity for an on-site data center in the coming months. (Business Wire)

The Reactor Pilot Program, launched in June 2025, allows reactor developers to partner with DOE and use DOE authorization instead of the traditional NRC licensing route, dramatically accelerating development timelines. Deployable Energy’s Unity reactor achieved criticality in roughly 150 days, a benchmark for execution speed that proves the US can still build nuclear reactors quickly when regulatory pathways are clear. Energy Secretary Chris Wright praised the achievement as proof America can “execute again” on nuclear innovation. (DOE)

These are test reactors, not commercial plants. But they demonstrate the technology works and the supply chain exists to manufacture fuel, components, and reactor systems at scale. The goal is 400 GW of US nuclear capacity by 2050, with microreactors serving military bases, data centers, and industrial facilities. Every major hyperscaler has signed nuclear deals totaling more than 9.8 GW of committed capacity. The July 4 deadline was symbolic, but the execution was real.

Eight States Move to Sue Over Offshore Wind Lease Buyouts

Eight East Coast states notified the federal government on July 16 that they plan to sue over the Trump administration’s cash buyouts of offshore wind leases held by developers Invenergy and Bluepoint Wind. California filed its own separate notice, challenging a $111 million buyout of Invenergy’s Morro Bay lease, a site permitted for up to 2 GW of offshore wind capacity, enough to power more than half a million homes. The administration has now spent more than $2.5 billion buying out eight offshore wind leases nationwide. (California DOJ)

California argues the buyouts violate federal law by using taxpayer funds to settle litigation that was never actually filed. The administration is redirecting some of that buyout money toward new natural gas plants in Indiana, Wisconsin, Iowa, and Kansas, a direct trade of offshore wind capacity for gas generation. The fight adds a new legal front to an offshore wind pipeline that keeps shrinking even as AI data centers keep pushing power demand higher. (E&E News)

Microsoft Opens $3.3 Billion Wisconsin Data Center

Microsoft opened its first data center at the Mount Pleasant campus in Wisconsin, a $3.3 billion facility that began limited operations and is now fully online. The project, which started construction years ago, has become one of Microsoft’s highest-profile AI infrastructure developments. The campus is designed to support multiple data center buildings as Microsoft scales capacity to meet Cloud and AI demand. (Data Center Knowledge)

Wisconsin continues to attract hyperscale investment, but the state also faces growing tension over data center costs. Oracle recently asked a Wisconsin court to overturn state utility regulators’ decision requiring hyperscale developers to post hundreds of millions of dollars in financial security before connecting to the grid. The case represents an early legal test of how states allocate the financial risks of powering AI infrastructure, particularly when data centers may abandon projects or demand levels change after grid upgrades are built. (Data Center Knowledge)

Have fun this weekend,

Will

Sources

Monday News Roundup 7/6/26

Will McKnight — Mon, 06 Jul 2026 16:59:18 GMT

Chevron Signs 20-Year Deal to Build Microsoft a Private, 2.67 GW Power Plant

On June 22, Chevron subsidiary Energy Forge One signed a 20-year power purchase agreement with Microsoft to build a co-located gas power plant in West Texas dedicated entirely to a Microsoft data center. The project, called Project Kilby, will deliver roughly 2.67 gigawatts of capacity in phases, using gas turbines from GE Vernova and Caterpillar. The power will never touch the public grid. First delivery is expected in 2028, as Microsoft plans $190 billion in capital spending for 2026, a 61% jump from 2025.

A fossil fuel major just decided selling electricity beats selling oil. When Chevron sees better returns building a power plant for one customer than pumping crude, the market has already moved.

Watch whether other oil majors copy Chevron’s model instead of just selling gas to utilities.

(Chevron)

Texas Approves Batch Zero Process for a 438 GW Data Center Queue

On June 18, the Public Utility Commission of Texas approved rules for ERCOT to process large-load interconnection requests, the applications developers file to connect new demand to the grid, in batches instead of one at a time. The queue has swelled past 438 GW, with nearly 90% of it from data centers. ERCOT’s all-time peak demand record is 85,508 MW, so the queue is more than five times the grid’s actual capacity. Historically, only about 49% of claimed data center load ever materializes, so grid planners are building rules for demand that may never show up.

(Utility Dive)

DOE Offers $17.5 Billion in Loans to Rebuild the Nuclear Supply Chain

On June 23, the Department of Energy’s Office of Energy Dominance Financing issued a conditional loan commitment of $17.5 billion to finance long-lead items, the specialized components like reactor vessels that take years to manufacture and usually cause construction delays, for five utility-led projects. The goal is 10 new large reactors, mostly Westinghouse AP1000 units. The loan implements Executive Order 14302’s target of 300 GW of new nuclear by 2050, but $17.5 billion covers only about 9% of the roughly $200 billion total cost for ten AP1000s. Watch whether utilities actually sign construction contracts once the full price tags land.

(Utility Dive)

FERC Orders Six Grid Operators to Justify or Rewrite Large-Load Rules

On June 18, FERC issued show-cause orders to the six largest U.S. grid operators, PJM, MISO, Southwest Power Pool, CAISO, ISO New England, and NYISO, directing each to justify its existing rules for connecting gigawatt-scale power users or propose reforms within 60 days. FERC went region by region instead of writing one national rule, a move built to survive legal challenges from individual states. The orders lean hard on cost transparency: who pays for the transmission upgrades a new data center requires. The rules due back by August will decide which data center projects actually get built, and where.

(Data Center Knowledge)

Supreme Court Strips Removal Protections From FERC Commissioners

On June 29, the Supreme Court ruled 6-3 in Trump v. Slaughter that a president can remove heads of independent agencies without cause, overturning the 90-year-old Humphrey’s Executor precedent. FERC’s five commissioners set the rules for the interstate power grid, transmission costs, and increasingly how data centers connect to it. Legal experts say there is now no doctrine stopping a president from firing a commissioner over a decision he dislikes, ending decades of independence regulators used to referee fights between utilities, generators, and data center developers. Watch whether this administration tests that power on FERC directly, given the agency is mid-fight over the same large-load rules above.

(E&E News)

Vattenfall Starts Construction on Germany’s Largest Offshore Wind Project

On July 1, Vattenfall installed the first monopile, the steel foundation that anchors a turbine to the seabed, for Nordlicht I in the German North Sea. Combined with the planned Nordlicht II, the cluster will reach 1.6 GW and produce about 6 terawatt-hours a year, roughly enough for close to two million German households. Both phases are due online by 2028. Europe is still building offshore wind at scale even as the U.S. cancels leases outright, below.

(Vattenfall)

Rolls-Royce SMR Wins Sweden’s First New Nuclear Contract in 40 Years

On June 15, Swedish utility Videberg Kraft selected Rolls-Royce SMR to build three small modular reactors, compact nuclear plants built from factory-made sections instead of poured on site, at the Värö Peninsula next to the existing Ringhals plant. Each reactor produces 470 MW, for roughly 1,410 MW combined and about 12 terawatt-hours a year. It is Sweden’s first new nuclear build in over 40 years, with first power expected in the mid-2030s.

(UK Government)

DOE Orders Colorado Coal Plant to Keep Running for a Third Time

On June 26, Energy Secretary Chris Wright issued a third emergency order forcing five utilities to keep the 427 MW Craig Unit 1 coal plant in Colorado running through September 26, six months past its original retirement date. It is a small plant, but three straight emergency extensions show how tight grid margins have gotten across the Southwest Power Pool region as demand climbs. Watch whether a fourth extension follows in September, or whether replacement capacity finally shows up.

(American Public Power Association)

Have fun this week,

Will

Sources

Do we really still need coal?

Will McKnight — Thu, 18 Jun 2026 18:18:09 GMT

Coal’s share of U.S. electricity has fallen to historic lows. Wind and solar together surpassed coal for the first time in history. Nearly half of the coal capacity that existed in 2010 has already been shut down.

And yet, the federal government has repeatedly used emergency authority to keep aging fossil plants running. There have been 43 total Section 202(c) orders and extensions and 22 plant-retirement deferral orders covering eight plants. Those orders have already added an estimated $300 million in extra ratepayer costs through mid-May 2026, with monthly costs still rising (IEEFA).

Today we’ll dive in to whether we can manage without coal today, and which communities get hurt if we ditch it.

What Would the Grid Look Like Without Coal Tomorrow?

Not good, for certain regions.

The North American Electric Reliability Corporation, which monitors grid reliability across the U.S. and Canada, projects capacity shortfalls across most of its assessment areas over the next decade, with summer peak demand growth expected to significantly outpace new generation coming online (NERC).

PJM is the clearest example. PJM is the grid operator covering 13 states from New Jersey to Illinois plus D.C., managing electricity for about 67 million people. In December 2025, PJM ran its 2027/2028 capacity auction and came up 6,623 MW short of its reliability target, described in trade coverage as the first time PJM failed to meet its reserve-margin target. FERC Commissioner Judy Chang said publicly: “We’re hitting the grid reliability crisis that has been brewing for years” (Utility Dive).

The auction also showed where replacement capacity stands. Data center-driven demand added a 5,250 MW increase to PJM’s forecast, while the auction drew in only 774 MW of new generation. The gap between what is being demanded and what is being built is the core of the reliability problem (Utility Dive).

That does not mean coal needs to stay forever. It means the replacement pipeline has to be moving faster than it currently is.

Which States Are Most Exposed?

The risk is not distributed evenly. A handful of states still depend on coal for most of their electricity and have little replacement capacity waiting to come online.

West Virginia, Kentucky, Missouri, Wyoming, and Indiana generate the largest shares of their electricity from coal, in some cases well above half their total generation. Per EIA 2024 state profiles, West Virginia ran roughly 85% coal and Indiana fell from 85% in 2014 to about 42% in 2024. Many of these states have fewer near-term replacement options than gas- and renewables-heavy regions (EIA).

Indiana is the sharpest near-term case. Two of its plants are currently under federal emergency orders requiring them to stay open. Pennsylvania has more coal plants than any other state by count, with 14 operational facilities as of 2025 according to Statista, and anchors much of the PJM capacity that grid operators are watching most closely.

California has almost fully exited coal in its power mix, with coal at roughly 2% of generation including imports in 2024. Texas has reduced coal sharply but was still around the mid-teens at about 15% in 2024. The coal-heavy states in Appalachia and the Midwest face a much steeper near-term replacement challenge.

Which Plants Is Washington Keeping Alive by Force?

The mechanism is Section 202(c) of the Federal Power Act, a wartime-era authority that lets the Department of Energy order power plants to keep running during grid emergencies. Historically, DOE used the authority sparingly. Since May 2025, use has accelerated sharply, with POWER’s running log counting 43 orders and extensions as of April 2026 and more added since (Power Magazine).

Centralia Unit 2 in Washington state, approximately 730 MW, operated by TransAlta. Washington’s last coal plant, scheduled to close under a 2011 state law. The DOE ordered it to keep running in December 2025 and has extended the order. It is also the largest source of mercury and greenhouse gas emissions in the state (Earthjustice).

J.H. Campbell in Michigan, operated by Consumers Energy. IEEFA and EIA put the affected facility at 1,331 MW. The most recent DOE order was issued May 18, 2026, running through August 16, 2026.

Eddystone Units 3 and 4 in Pennsylvania, 760 MW, operated by Constellation Energy. These are oil and gas units, not coal, but are part of the broader Section 202(c) fight. DOE’s May 21, 2026 order runs May 25 through August 22, 2026.

(Constellation)

R.M. Schahfer Units 17 and 18 in Indiana, operated by NIPSCO, with affected capacity reported as roughly 700 to 850 MW depending on whether you use EIA’s affected-unit figure or POWER’s. Under order through mid-June 2026. Continuing to operate would require more than a billion dollars in expenditures according to the groups challenging the order.

F.B. Culley Unit 2 in Indiana, operated by CenterPoint Energy. Under DOE order through June 21, 2026. The plant costs $615,000 per day to run during the extension period.

Craig Station Unit 1 in Colorado, approximately 446 MW, operated by Tri-State Generation and Transmission. Order extended through June 28, 2026.

The orders are being challenged in the U.S. Court of Appeals. Opponents argue Section 202(c) is limited to imminent, unexpected emergencies. DOE has defended a broader view of its authority. That argument is unresolved (Earthjustice).

What Does the Job Count Look Like?

The total depends on how broadly you count. Direct coal mining employment averaged about 44,000 per the EIA 2024 Annual Coal Report. Broader coal-linked employment estimates run much higher, into the hundreds of thousands when power plant workers and transportation are added, though those figures rely on older methodology (Global Energy Monitor).

Power plant operators earn a median annual wage of around $100,000, with BLS 2023 OES data showing roughly $97,000 median and $107,000 mean for the occupation (BLS). The jobs concentrate in West Virginia, Kentucky, and Pennsylvania, and they are not evenly distributed. They sit in rural counties where coal is often the largest private employer. That is the real constraint on speed, not the megawatts.

The coal question is really two separate questions that keep getting tangled together: do we need it for the grid right now, and what do we owe the people whose livelihoods depend on it. Grid operators are giving a fairly clear answer on the first one, at least for specific regions over the next several years. The second question is going to outlast the plants.

What to watch: whether the Court of Appeals strikes down the 202(c) orders this summer, and what PJM’s next capacity auction looks like when results come out later this year.

Have fun this weekend,
Will

Sources

AI Power News 6/8/26

Will McKnight — Mon, 08 Jun 2026 16:50:56 GMT

FERC Sets June Deadline for Large Load Interconnection Rules

FERC committed on April 16 to issue new regulations by the end of June governing how data centers and other large electrical loads connect to the interstate transmission system. The move responds to Energy Secretary Chris Wright’s October 2025 directive asking FERC to accelerate interconnection procedures for loads exceeding 20 megawatts, a category that covers most AI data centers but also includes many big box stores and industrial facilities (POWER Magazine).

The rulemaking tackles three questions that will shape billions of dollars in infrastructure spending. Who pays when a data center requires transmission upgrades? Should costs be socialized across all ratepayers or assigned directly to the customer causing them? And does FERC even have jurisdiction over load interconnection, which has historically been a state matter?

FERC Chair Laura Swett called large load integration “the most important and pressing problem in contemporary American public policy” and signaled she’s willing to push federal authority to its legal limits. The agency has already reviewed more than 3,500 pages of comments. State regulators, meanwhile, warned that federal overreach could disrupt retail rate cases that fall under state jurisdiction (E&E News).

Watch for the June ruling. It will either create a clear national framework for data center development or trigger a wave of litigation that delays projects for years.

Oklahoma Passes Data Center Ratepayer Protection Act

Oklahoma Governor Kevin Stitt signed the Data Center Consumer Ratepayer Protection Act into law in early May, making Oklahoma the first state to mandate that data centers cover their own infrastructure costs. The law requires facilities adding 75 megawatts or more of demand to pay for grid upgrades directly rather than passing costs to residential and commercial ratepayers. It also imposes a 60-day notice requirement before land purchases, giving local communities advance warning of major projects (FOX23).

The bill passed both chambers unanimously with bipartisan support from 36 co-authors. It takes effect July 1, though utilities still need to develop implementation rules. Senator Grant Green framed it as protecting farmers and ranchers from surprise developments. “Data centers won’t be able to buy land without notifying the community,” he said (Oklahoma Energy Today).

Other states are watching. Maryland passed similar legislation in May 2025, and Virginia regulators revised permitting guidance in May 2026 to challenge assumptions that backup generators are rarely used.

SoftBank Announces $85 Billion France Data Center Investment

SoftBank Group announced on May 30 it will invest up to €75 billion, approximately $85 billion, to develop 5 gigawatts of AI data center capacity in France. The first phase targets 3.1 GW in the Hauts-de-France region by 2031, with sites in Dunkirk, Bosquel, and Bouchain. State-owned utility EDF is providing the site of a former power plant in Bouchain and partnering on power supply (Data Center Knowledge).

The deal highlights why access to reliable electricity is becoming the decisive factor in data center site selection. France offers competitive nuclear power and existing grid infrastructure at decommissioned industrial sites. SoftBank founder Masayoshi Son said France’s energy position “was absolutely decisive in our decision.”

SoftBank is simultaneously funding the Stargate program in the U.S. and proposing a 10 GW site in Ohio. Analysts from TD Cowen noted the company has over $130 billion in debt and questioned how it will fund all announced projects.

Google Signs 1 GW Texas Solar Deal with TotalEnergies

Google and TotalEnergies signed two 15-year power purchase agreements on February 9 to deliver 1 gigawatt of solar capacity to Google’s Texas data centers. The electricity will come from TotalEnergies’ 805 MW Wichita and 195 MW Mustang Creek solar projects, with construction starting in the second quarter of 2026. The deal will deliver 28 terawatt-hours of electricity over the contract term (TotalEnergies).

TotalEnergies called it the largest renewable PPA it has signed in the U.S. The agreement adds to separate 1.2 GW PPAs that Clearway, a company 50% owned by TotalEnergies, signed with Google for projects across Texas, the Northeast, and Central U.S.

The scale is impressive. The tension is that solar only generates power during daylight while data centers run 24/7. Google will still rely on the broader Texas grid for balancing at night and during cloudy periods.

NRC Approves NuScale’s Uprated SMR Design

The Nuclear Regulatory Commission approved NuScale Power’s uprated 77 MWe small modular reactor design on May 28, 2025, completing its technical review ahead of schedule and under budget. This is NuScale’s second approved design, the previous 50 MWe version was certified in 2020, and makes it the only SMR company with NRC design approval (Utility Dive).

A standard design approval means the reactor meets NRC safety requirements. Companies still need separate permits to construct and operate plants using the design. NuScale said the larger modules will better serve hyperscale data center customers. The design runs on conventional low-enriched uranium rather than scarce high-assay fuel, which gives it a supply chain advantage.

The approval came weeks after the Trump administration issued executive orders requiring NRC to review new reactor applications within 18 months, down from current multi-year timelines.

Colorado Breaks Ground on NextEra Battery Project

NextEra Energy Resources and Platte River Power Authority held a groundbreaking on June 5 for the Weld Energy Center, a battery energy storage project in Ault, Colorado, north of Greeley. The $141 million facility will provide grid reliability for Fort Collins, Loveland, Longmont, and Estes Park. Construction will create over 200 jobs and generate tax revenue for local schools and emergency services (CBS Colorado).

Battery storage is rapidly moving from niche technology to core grid infrastructure. Over 10 gigawatts of utility-scale battery storage were installed in 2024, and 2025 exceeded that total by mid-year. The expansion is driven by the need to balance intermittent solar and wind generation as those sources grow.

Trump Administration Announces $850M to Modernize US Coal Capacity, Build 2 New Plants

The Trump administration announced $850 million in Department of Energy funding to upgrade existing coal-fired power plants and build two new coal facilities, the most significant federal investment in coal in decades. Energy Secretary Chris Wright framed the announcement as a grid reliability measure, arguing the U.S. cannot phase out dispatchable generation, power sources that can run on demand regardless of weather, while AI data center electricity demand continues to climb (Utility Dive).

The move reverses a decade of federal policy that prioritized coal plant retirements. The administration has now combined direct investment with emergency regulatory orders to slow the coal exit across multiple states. For energy investors, the signal is clear: plants that were priced for closure may have extended operational life.

Watch for legal challenges from states and environmental groups. Federal authority to fund new coal construction sits on shakier legal ground than the emergency orders keeping existing plants online.

DOE Orders OUC’s 465-MW Coal Unit in Florida to Continue Running

The Department of Energy ordered Orlando Utilities Commission to keep its 465-megawatt Stanton Unit 1 coal plant operating rather than proceed with its planned retirement. The DOE invoked Section 202(c) of the Federal Power Act, an emergency reliability provision that gives the federal government authority to require a plant to stay online when its shutdown would threaten grid stability (Utility Dive).

OUC, the municipal utility serving Orlando, had scheduled the unit for retirement as part of its clean energy transition plan. The order overrides that schedule and is part of a broader pattern: the Trump administration has now issued multiple 202(c) orders across several states to prevent coal retirements ahead of peak summer demand.

Municipal utilities have fewer regulatory pathways to contest these orders than investor-owned utilities. OUC’s options are limited, and the forced operation adds cost that will eventually land on Orlando ratepayers.

Have fun this week,

Will

Sources

AI Power News 5/26/26

Will McKnight — Tue, 26 May 2026 16:48:33 GMT

Trump Cabinet Joins Utah Governor at Operation Gigawatt Energy Summit

Governor Spencer Cox hosted the Operation Gigawatt Summit in Park City, Utah on May 22, drawing Energy Secretary Chris Wright, EPA Administrator Lee Zeldin, White House Science Advisor Michael Kratsios, and NRC Chairman Ho K. Nieh for a day-long session on America’s AI energy buildout. The gathering produced one signed commitment: Utah and the White House Permitting Improvement Steering Council inked a memorandum of understanding to coordinate state and federal approvals for critical energy infrastructure, a template the administration wants to replicate in other states.

The competitive thread running through every session was China. “China understands that whoever can produce power the fastest and the cheapest will own the future of AI,” said investor Sal Churi of Trust Ventures at the summit. The US leads on AI capability right now, and energy access is the variable that changes that.

Faster nuclear permitting was the central thread, with the NRC chairman at the table alongside governors and private developers. That combination is unusual. Watch whether the Utah-federal permitting MOU produces measurable timeline reductions or stays in the category of signed-but-not-enforced agreements. (Deseret News)

SpaceX Pays $185M for xAI’s First Data Center

SpaceX paid $185 million for a 785,000-square-foot data center in Memphis, Tennessee, marking xAI’s first data center. The property previously traded in December 2023 for $35 million when Phoenix Investors acquired it from Electrolux.

(Memphis Business Journal)

In its IPO filing, SpaceX disclosed xAI will buy another $2.8 billion worth of turbines for AI infrastructure over the next three years, including $2 billion specifically for mobile gas turbines. That is the same type of turbine generating legal trouble. The NAACP sued xAI last month for operating dozens of unregulated gas turbines, with xAI using 46 turbines despite permits for only 15. The EPA ruled earlier this year that xAI was operating the turbines in violation of federal law.

xAI separately received permission from Mississippi regulators on April 29 to install 41 natural gas turbines generating 1.2 GW at a former Duke Energy power plant site in Southaven to power its Colossus 2 and Colossus 3 data centers. The company is building its own shadow grid rather than wait for utilities. Watch whether other hyperscalers follow this model or if permitting battles slow the approach. (Commercial Property Executive)

NRC Accepts NANO Nuclear Microreactor Construction Permit

The Nuclear Regulatory Commission formally accepted NANO Nuclear Energy’s construction permit application on May 20 for its KRONOS microreactor at the University of Illinois Urbana-Champaign, initiating formal safety, environmental, and technical review. The application was submitted March 31 by the University of Illinois, NANO Nuclear’s partner for the planned full-scale reactor.

The KRONOS system is now among a small group of Generation IV advanced nuclear reactors, and the company believes it is the first commercially-ready microreactor to progress to the construction permit application stage of the NRC’s formal licensing process. Based on NANO Nuclear’s current understanding, the company estimates the NRC formal review will be completed in 2027, providing the opportunity to begin nuclear construction activities at the university.

This matters because microreactors are the only nuclear tech that can potentially match data center deployment timelines. Traditional large reactors take a decade or more. The KRONOS microreactor is designed to provide energy for applications including data centers, industrial facilities, remote communities, mining projects, and military bases. If NANO Nuclear can build and operate by late 2027, it will set the commercial pace for an entire category. (Globe Newswire)

Ontario’s 250 MW Battery Project Goes Live Ahead of Schedule

Ameresco’s joint venture with Atura Power began commercial operations at a 250 MW/1,000 MWh battery energy storage project in Napanee, Ontario on May 19, approximately five weeks ahead of schedule and on budget. Ontario’s Minister of Energy and Mines Stephen Lecce announced the project’s operational status.

The project was a winner in the Ontario Independent Electricity System Operator’s Long-Term 1 competitive tender in 2023, which awarded over 850 MW of battery storage capacity contracts. Energy storage is the mechanism that makes intermittent renewables work with AI data center loads that run 24/7. Batteries smooth the gap.

Ontario’s grid operator launched a Long Lead-Time request for proposal earlier this month seeking to procure up to 800 MW of long-duration energy storage. Canada is building storage faster than most U.S. states. (Energy Storage News)

DOE Awards $94M to Eight Companies for SMR Deployment

The Department of Energy on May 14 selected eight companies to receive more than $94 million in federal cost-shared funding to support near-term deployment of advanced light-water small modular reactors in the United States by addressing gaps in licensing, supply chain, and site preparation. Small modular reactors are factory-built nuclear plants with smaller output than traditional gigawatt-scale reactors.

Recipients include Constellation SMR Development ($17.3 million) to pursue an NRC Early Site Permit in New York, Nebraska Public Power District ($27.9 million) for a permit in Nebraska, and BWXT Nuclear Energy ($21.4 million) to procure equipment for reactor pressure vessel assembly in Mount Vernon, Indiana. DOE announced $800 million in Tier 1 awards in December 2025 to Tennessee Valley Authority and Holtec Government Services to advance initial projects in Tennessee and Michigan.

These are preparation grants, not construction money. But site permits and supply chain investments are the long poles in the tent for nuclear deployment. Watch whether these projects can secure power purchase agreements with tech companies before breaking ground. (Department of Energy)

DOE’s $1.9B SPARK Transmission Program Closes Applications

The Department of Energy’s SPARK program closed its full application window on May 20, wrapping up competition for approximately $1.9 billion in grid upgrade funding. SPARK, which stands for Speed to Power through Accelerated Reconductoring and other Key Advanced Transmission Technology Upgrades, finances replacing old transmission lines with advanced conductors that carry more power on the same towers without building new right-of-way, one of the fastest and cheapest ways to expand grid capacity.

The program prioritizes projects that can move power to large new loads, which in 2026 means data centers. It is part of the Grid Resilience and Innovation Partnerships program, authorized at up to $10.5 billion over five years through the Infrastructure Investment and Jobs Act.

Reconductoring rarely makes headlines. It is also typically faster to permit and deploy than new nuclear or gigawatt-scale batteries, and it uses infrastructure that already exists. The question now is who applied and whether DOE selects projects that actually close existing bottlenecks rather than upgrade lines that were already unconstrained. Awards are expected later this year. (Department of Energy)

Have fun this week,

Will

Sources

AI Power News 5/18/26

Will McKnight — Mon, 18 May 2026 17:25:11 GMT

Anthropic Buys xAI’s Entire Colossus 1 Data Center Capacity

Anthropic announced Wednesday it is buying all the compute capacity at xAI’s Colossus 1 data center in Tennessee, roughly 300 megawatts that allowed Anthropic to immediately raise its usage limits for Claude. The deal is worth billions of dollars and immediately monetizes one of xAI’s most impressive technical accomplishments. Elon Musk said on X that xAI had already moved training to a newer data center, Colossus 2, and simply did not need both facilities. xAI is now expanding its Memphis footprint to nearly 2 gigawatts with a third building purchase and a planned natural gas power plant.

(xAI)

The arrangement raises questions about xAI’s strategy. Most AI companies prioritize using compute for their own model training rather than renting it out. Google, Meta, and Microsoft are all building more data centers even as they train models, because they view computing power as strategic for building tomorrow’s products. By focusing on data centers rather than models, xAI is positioning itself more like a neocloud business: buying GPUs from Nvidia and renting them to model developers. That is a harder business, squeezed by chip suppliers and shifting demand cycles. CoreWeave, which oversees comparable computing power, is valued at less than a third of xAI’s $230 billion valuation from its January funding round.

The deal also suggests xAI’s Grok chatbot is not burning through the compute resources the company built. Grok usage has plummeted since image generation controversies earlier this year. Becoming a supplier to Anthropic is lucrative, but it sends an unusual message about where Musk’s priorities actually lie.

(TechCrunch)

NextEra Energy to Acquire Dominion Energy for $67 Billion

NextEra Energy announced it will acquire Dominion Energy in an all-stock deal valued at approximately $67 billion, creating the world’s largest regulated electric utility. The combined company will serve roughly 10 million customers across Florida, Virginia, North Carolina, and South Carolina. Dominion shareholders get 0.8138 NextEra shares for each share they hold, and Dominion customers receive $2.25 billion in bill credits over the two years after closing.

Virginia is the largest data center market in the world, and Dominion is the primary utility serving it. NextEra already has 21 gigawatts of data center load requests in its Florida pipeline alone. Combining the two utilities puts a single company in control of the electricity supply for the most data center dense region in the country, plus one of the fastest growing.

The deal still requires federal and state regulatory approval and is not expected to close until mid-to-late 2027. Watch whether state regulators in Virginia push back on the return on equity NextEra can charge, since higher utility profits ultimately flow through to data center power costs and residential bills.

(CBS News)

FERC Chair Signals June Rulemaking on Data Center Grid Access

Federal regulators will take “widespread action” next month to accelerate how data centers and other large loads connect to the electric grid, FERC Chair Laura Swett announced Tuesday at a United States Energy Association event in Washington. The move comes as interconnection delays have become what Swett called “the most important and pressing problem in contemporary American public policy.” FERC staff have reviewed more than 3,500 pages of comments on how to handle large load interconnections, examining recent actions in PJM and Southwest Power Pool.

The stakes are enormous. Data centers now account for 97% of projected load growth in the PJM region, which serves 67 million people across 13 states. Wholesale power prices in PJM jumped 76% in the first quarter of 2026 compared to the same period in 2025, driven primarily by data center demand. That translates to real money for residential customers. If current trends hold, PJM households could face an additional $70 per month in electricity costs by 2033, according to the Natural Resources Defense Council.

FERC has already ordered PJM to create new transmission service options for colocated loads, essentially data centers paired with their own generation. Now the agency is weighing whether to extend federal oversight to all large load interconnections over 20 megawatts, a jurisdiction historically left to states. The Department of Energy set an April 30 deadline for FERC to act, but June appears to be when the rubber meets the road. Watch whether FERC requires full upfront payment for network upgrades or allows costs to be credited back over time.

(Broadband Breakfast)

DOE Awards $94 Million to Eight Companies for SMR Deployment

The Department of Energy handed out $94 million to eight companies on Wednesday to speed deployment of small modular reactors, the light water designs that are the closest to commercial reality. A small modular reactor, or SMR, is a nuclear plant that generates 300 megawatts or less and can be built in a factory rather than entirely onsite. The money targets three bottlenecks: licensing, supply chain, and site preparation. Constellation SMR Development received $17.3 million to pursue an Early Site Permit in New York. Nebraska Public Power District got $27.9 million for similar work in Nebraska. On the manufacturing side, Framatome secured $8.8 million to expand fuel pellet production in Washington, adding capacity for 200 metric tons of uranium annually.

This is the second round of federal SMR funding. In December, DOE awarded $800 million to Tennessee Valley Authority and Holtec for first mover projects. Those two deals alone represent more than 1 gigawatt of planned nuclear capacity. The $94 million announced this week is smaller but strategically important because it funds the supply chain that will make follow on orders possible. Secretary Chris Wright framed the awards as critical to powering data centers and manufacturing, saying advanced light water SMRs will provide “reliable, round the clock power” for AI growth.

The subtext is speed. Light water SMRs use proven reactor technology, not exotic designs still under development. That matters when hyperscalers are signing 20 year power purchase agreements for nuclear electricity they need online by 2030.

(Department of Energy)

PJM Wholesale Power Costs Jump 76% on Data Center Demand

Wholesale electricity prices on the PJM grid averaged $136.53 per megawatt hour in the first quarter of 2026, up from $77.78 in the same period last year, according to Monitoring Analytics, PJM’s independent market monitor. That is a 76% increase in three months, driven almost entirely by data center load. The report confirms what grid operators have been warning about: data centers are not just another customer, they are reshaping the economics of the entire Mid Atlantic and Midwest power market. Data center forecasts now account for $21.3 billion, or 45%, of capacity costs across PJM’s last three auctions.

Capacity prices tell the story even more starkly. PJM’s capacity auction cleared at $329.17 per megawatt day for the 2026 to 2027 delivery year. Two years ago, the same auction cleared at $28.92. That is more than a tenfold increase. Utility supply rates across the region have already risen between 5% and 44%, and that was before this year’s auction results hit customer bills.

The grid itself is struggling to keep up. PJM’s most recent capacity auction fell 6,500 megawatts short of its reliability target for the first time. Meanwhile, data centers represent 97% of new load growth in the region. This is not a temporary spike. It is a structural shift in electricity demand happening faster than generation can be added.

(Bloomberg)

Battery Storage on Track for Record 24 GW of Additions in 2026

Developers plan to add 24 gigawatts of utility scale battery storage to the U.S. grid in 2026, a 60% jump from the 15 gigawatts added last year, according to the Energy Information Administration. By the end of the first quarter of 2027, total U.S. battery storage capacity will surge from 44.6 gigawatts to over 67 gigawatts. Three states dominate the buildout: Texas, California, and Arizona. Major projects slated for this year include the 621 megawatt Lunis Creek battery in Jackson, Texas, and the 500 megawatt Bellefield 2 Solar and Energy Storage Farm in Kern County, California.

Battery storage has become essential grid infrastructure, not an optional add on. Utilities need batteries to manage the rapid expansion of solar generation, which floods the grid with cheap power at midday but disappears in the early evening when demand peaks. Data centers are also driving storage deployment. On site batteries allow data centers to buffer against grid instability and store renewable energy during surplus periods. The U.S. data center market is expected to grow to 60 gigawatts in 2026, and battery storage is integral to energy management for virtually all new large scale facilities.

Costs are falling fast. Levelized cost of storage has reached approximately $140 per megawatt hour in California solar plus storage configurations, according to Fluence. That makes storage economically competitive for an expanding set of grid services, from energy arbitrage to frequency regulation.

(PV Magazine USA)

Lake Tahoe Faces Power Loss as Utility Prioritizes Data Centers

NV Energy told Liberty Utilities it will stop providing power to the Lake Tahoe region after May 2027, citing a need to redirect capacity to data centers in Northern Nevada. Nearly 49,000 residents in the Lake Tahoe area now face the prospect of finding a new power source within a year. The reason is straightforward: Google, Apple, and Microsoft have built or are planning data center facilities around the Tahoe Reno Industrial Center east of Reno. Those facilities are pulling massive amounts of electricity. Data centers consumed 22% of Nevada’s electricity in 2024, and that share could hit 35% by 2030.

The scale of new demand is staggering. Twelve data center projects in Northern Nevada alone could drive 5,900 megawatts of new demand by 2033, according to the Desert Research Institute’s analysis of NV Energy’s 2024 Integrated Resource Plan. About 75% of major project load growth in NV Energy’s own filing is attributed to data centers. Liberty Utilities asked California regulators to authorize an expedited request for proposals for replacement energy beginning June 2027. Short term power is likely available from elsewhere in the West, but Lake Tahoe’s small customer base gives it little leverage when competing against data centers and mining companies for long term contracts.

The situation underscores a broader dynamic. When utilities face capacity constraints, data centers with guaranteed revenue streams and 20 year contracts win. Residential customers with variable usage and month to month billing lose.

(Fortune)

NERC Issues Highest Level Alert on Data Center Grid Risks

The North American Electric Reliability Corporation issued a Level 3 alert, its highest warning level, about threats to the power system from large data centers. The alert was developed after reports of data centers abruptly going offline in Virginia and Texas, raising concerns about blackouts. NERC warned that “computational loads, such as data centers, could increase exponentially in the next four years” and that “significant risks” to the bulk power system “need to be addressed through immediate action.” The alert recommends seven essential actions for transmission owners, grid operators, and utilities to ease risks from large loads.

AI workloads can swing power demand dramatically in milliseconds, far faster than traditional industrial loads. A paper from Nvidia, Microsoft, and OpenAI scientists warned last year that AI power swings can “cause physical damage” to grid infrastructure. NERC’s alert calls for transmission planners to develop detailed modeling data from computational loads and study grid stability margins at least annually in areas with large AI infrastructure. Transmission owners should install fault recording devices to assess data center performance during grid disturbances.

The actions are not mandatory yet, but NERC is clearly laying groundwork for mandatory reliability standards for data centers, which would require FERC approval. Part of the challenge is scale. Multiple gigawatt complexes are now being planned across the country. A gigawatt is enough to power roughly 750,000 homes.

(E&E News)

Have fun this week,

Will

Sources

Are Technological Advances Reducing the Health Side Effects of Living Near a Data Center Yet?

Will McKnight — Thu, 14 May 2026 17:42:50 GMT

Most people picture a data center as a quiet building. Glass facades, humming servers, precision climate control. The impression is of something contained and sterile.

People who live next to one describe something different. A low vibration they feel in their chest at 3am. Headaches that started the month construction finished. Windows that stay closed year-round. Not because of weather, but because of noise. A U.S. News investigation published in April quoted neighbors using the phrase “living in hell.”

The question is whether any of this is improving. New cooling technology is genuinely reducing noise inside and around these buildings. But the biggest source of health complaints near data centers is not the servers. It is the gas turbines.

(NPR)

What are people near data centers experiencing, and what is actually causing it?

Neighbors living near large data center campuses report a consistent set of symptoms: sleep disruption, headaches, dizziness, vertigo, nausea, and ear pain. Some have documented hypertension and panic attacks. The complaints fall into two categories, and they come from two different sources.

The first is air quality. Most large data centers run gas turbines or diesel generators on-site to provide power independently of the grid. These are not emergency backup units that kick on during outages. Many run 24 hours a day, seven days a week, producing the nitrogen oxide and particulate matter emissions the American Cancer Society links to heart disease, lung disease, and cancer. A Virginia Commonwealth University study found that data centers in Northern Virginia, which has the highest concentration of data centers anywhere in the world, produce 11% of the region’s total nitrogen oxide emissions and 3% of its particulate matter as point source pollutants. A separate study used EPA modeling to estimate that one Vantage facility alone causes between $53 million and $99 million in annual health damages to surrounding communities (Piedmont Environmental Council).

The second is noise. Cooling systems that keep servers from overheating run continuously and generate two kinds of sound. Audible fan noise from air-cooled facilities reaches 85 to 100 decibels at the source. That is roughly the equivalent of a motorcycle running outside your bedroom window around the clock. Below that is infrasound, sound under 20 hertz that sits below the threshold of human hearing. You cannot hear it. You can feel it. Research published in Frontiers in Behavioral Neuroscience this year found that infrasound elevates cortisol levels and causes nausea and physical discomfort even in people who have no idea they are being exposed (TechRadar).

Most coverage of the data center health problem focuses on the servers. Turbines and generators, running 24/7 to power those servers, are driving most of the air quality harm.

What is infrasound, and why can’t regulators do anything about it?

Infrasound comes from the data center itself, not the power supply next to it. Large HVAC units, cooling towers, and industrial fans produce sub-20 hertz vibrations as a byproduct of continuous operation. The problem for residents trying to get relief is that standard noise meters are not designed to capture frequencies below 20 hertz. When neighbors file complaints, inspectors arrive with equipment that physically cannot detect what is bothering them. The complaint gets dismissed. The hum continues (Tom’s Hardware).

The 2026 Frontiers study is worth flagging here because it tested subjects who had no expectation of being exposed and no auditory awareness of it. Infrasound still elevated their cortisol and triggered aversive responses. The harm does not require the person to know it is happening.

There is no federal measurement standard for sub-audible industrial noise. Until there is, regulators have no baseline to enforce against.

How bad is the turbine problem, and is anyone tracking it?

The turbines exist largely because the grid cannot keep up. Data centers need power now. Interconnection queues at PJM, the nation’s largest grid operator, run five to seven years. So operators build their own generation on-site. They are functioning as power plants. Many operate under air permits that were not designed for continuous industrial generation (Inside Climate News).

Virginia’s regulators have been tracking what happens if this continues to scale. If Northern Virginia’s data center turbines ran at their full permitted capacity, they would account for 49% of the region’s total nitrogen oxide point source emissions. Virginia’s Department of Environmental Quality launched a dedicated air quality monitoring project specifically because of data center density in that corridor. The state is currently weighing whether to expand the conditions under which facilities can run on-site generators (Grist).

The grid problem creates the health problem. That chain matters because grid-tied power, not better turbines, is the only real fix for the air quality side.

Is liquid cooling actually making a dent in the noise?

Liquid cooling addresses the fan noise problem directly. Traditional air-cooled data centers push roughly 10,000 cubic feet of air per minute through server banks, generating about 95 decibels of ambient noise and consuming 25 kilowatts just for fan power. Immersion cooling, where servers are fully submerged in non-conductive fluid, is virtually silent. Direct-to-chip cooling, which pipes liquid directly to processors rather than blasting air across them, eliminates most facility-level fans. Microsoft began fleet-wide deployment of direct-to-chip cooling across Azure campuses in July 2025. Direct-to-chip now holds a 47% share of new data center cooling deployments (Introl).

Most existing data centers were not built for liquid cooling, and retrofitting them is expensive. Liquid cooling is being adopted in new hyperscale builds. The legacy stock already sitting in residential corridors in Virginia, Phoenix, and Chicago’s suburbs will stay air-cooled for years. The technology is improving at the leading edge; the installed base is not.

Liquid cooling also does nothing for the infrasound problem. The sub-audible hum comes from building-level mechanical systems: cooling towers, HVAC units, and large pumps. Swapping server cooling from air to liquid does not eliminate those.

Are regulations actually forcing the industry to change?

Regulations are starting to move, but slowly. Montgomery County, Maryland adopted a 50-decibel noise limit at property lines for data centers, stricter than Maryland’s standard 65-decibel industrial threshold and among the toughest in the country. Chandler, Arizona’s city council voted unanimously against a new data center application in 2025, with noise concerns cited explicitly. More than 200 data center-related bills were introduced in state legislatures in 2025 alone, and moratorium proposals have appeared in 11 states in 2026 (MultiState).

The infrasound gap remains unaddressed at every level of government. No federal standard exists for measuring sub-audible industrial frequencies. Communities cannot file effective complaints for something that available equipment cannot record. Researchers can document the health effects; regulators have no mechanism to act on them.

Fuel cells are a more promising near-term answer to the turbine problem than nuclear or renewables. Goldman Sachs estimates fuel cells could cover 6 to 15 percent of incremental data center power demand, with modular systems deployable in under a year, far faster than grid interconnection. They produce fewer emissions than gas turbines and generate substantially less noise (Goldman Sachs). Adoption is early.

The technology to reduce harm near data centers exists and is improving. Liquid cooling works. Fuel cells are cleaner and quieter than gas turbines. The problem is that adoption moves in years and the buildout is moving in months. For the communities sitting next to today’s infrastructure, they need improvements now.

Have a nice weekend

— Will

Sources

AI Power News 5/11/26

Will McKnight — Mon, 11 May 2026 17:42:34 GMT

Nvidia and PulteGroup Are Mounting AI Data Centers on the Sides of Homes

Nvidia, homebuilder PulteGroup, and energy startup Span announced a pilot program last week to mount small liquid-cooled AI computing units on the exterior walls of newly built homes, turning residential neighborhoods into distributed data center networks. Each unit, called an XFRA node, holds an Nvidia RTX PRO 6000 Blackwell GPU and draws power from the home’s electrical panel. Span claims it can replicate the capacity of a 100-megawatt data center by deploying nodes across 8,000 homes in roughly six months at $3 million per megawatt, well below what traditional data center construction costs. PulteGroup has deployed the system in a single home so far in early testing, with plans to integrate XFRA units into new subdivisions.

Homeowners pay roughly $150 per month for electricity and internet service; installation is free. Span sells that compute to AI cloud providers and keeps the revenue. The homeowner hosts the hardware, takes the electricity bill reduction, and otherwise does nothing.

This is a genuinely new model, and the economics could work at scale. The friction is regulatory: utilities have no framework for millions of residential nodes acting as compute sellers, and zoning, insurance, and grid interconnection rules were all designed for a different world. Whether this scales before those frictions stop it is the real question.

Watch for hyperscaler response. If a major cloud provider signs a formal agreement to buy compute from residential XFRA networks, this goes from pilot to industry story fast. (CNBC)

Meta Bets on Space-Based Solar and 100-Hour Energy Storage

Meta reserved 1 gigawatt of space-based solar energy from Overview Energy, a startup building satellites in geosynchronous orbit that collect sunlight around the clock and beam it to Earth as near-infrared light. An orbital demonstration is planned for 2028; commercial delivery to the US grid could begin as early as 2030. Simultaneously, Meta reserved 1 gigawatt of ultra-long-duration storage from Noon Energy, using reversible solid oxide fuel cells that hold power for more than 100 hours. Meta now has staked positions in space solar, long-duration storage, nuclear, and geothermal simultaneously. That is not a single bet. That is a hedge against every possible failure in the grid buildout. (Meta)

California Batteries Cover 40% of Statewide Demand at Peak

California’s grid-scale battery fleet discharged a record 12,000 megawatts during peak evening hours in late March, briefly covering more than 40% of statewide electricity demand at the moment the grid was most strained. California now has over 17,000 megawatts of installed battery storage, up nearly 2,000% since 2019. The milestone proves what utility planners have argued for years: battery storage can function as genuine dispatchable capacity, not just a short-duration buffer. A dispatchable resource is one the grid operator can call on demand, which was previously only possible with gas peakers or hydro. (Inside Climate News)

NextEra Gets Federal Approval for 10 GW of New Natural Gas

NextEra Energy received federal approval to develop up to 10 gigawatts of new natural gas generation in Texas and Pennsylvania, including a 5.2 GW plant in Texas and a 4.3 GW facility in southwestern Pennsylvania, both designed to serve AI data center hubs. NextEra’s Florida utility separately has 21 gigawatts of data center load requests in its interconnection queue, with more than half in advanced discussions. The company’s upside forecast for new generation to serve data center customers now reaches 30 gigawatts by 2035, making NextEra one of the clearest direct beneficiaries of the AI power buildout. (Data Center Dynamics)

Fermi America Implodes: CEO and CFO Out, Stock Down 81%

Fermi America, the startup backed by former US Energy Secretary Rick Perry with plans to build 11 gigawatts of nuclear, solar, and natural gas for a Trump-branded AI campus near Amarillo, Texas, lost its CEO and CFO within two days of each other in late April. The stock debuted at $21 per share in October 2025 and now trades near $5, an 81% decline from its IPO price. No confirmed hyperscaler tenant ever signed a contract. Fermi is the first significant public failure of a company built entirely on the AI power buildout story, a reminder that proximity to a real trend does not guarantee a real business. (TechCrunch)

MARA Holdings Acquires 505 MW Ohio Gas Plant for $1.5B

Bitcoin miner MARA Holdings signed a deal to acquire Long Ridge Energy, a 505-megawatt combined-cycle natural gas power plant in Hannibal, Ohio, plus 1,600 acres of adjacent land, for approximately $1.5 billion including assumed debt. MARA plans to build a 200-megawatt AI data center at the site starting in 2027. The acquisition marks a clear inflection point: bitcoin miners and AI data center operators are now competing directly for the same power assets, and the miners who already own generation facilities have a structural cost advantage that most data center developers cannot quickly replicate. (Data Center Dynamics)

Massachusetts Signs 2,400 MW of Offshore Wind, Saving $1.4B

Massachusetts signed contracts for 2,400 megawatts of offshore wind power, saving an estimated $1.4 billion compared to prior procurement rounds for equivalent capacity. The deals demonstrate that offshore wind costs continue to fall even as federal lease activity has slowed under the current administration. Massachusetts is targeting 3,200 megawatts of offshore wind capacity by 2035, and cheaper procurement costs improve the odds that the state hits that number. (Electrek)

Have fun this week,

Will

Sources

AI Power News 5/6/26

Will McKnight — Wed, 06 May 2026 15:13:17 GMT

Meta Plans to Beam Solar Energy Down from Space

On April 27, Meta announced two new energy partnerships that are either a glimpse of AI’s power future or an expensive hedge, depending on how you read it. The first deal, with Overview Energy, reserves up to 1 gigawatt of space solar power for Meta’s data centers. Overview places satellites in geosynchronous orbit, roughly 22,000 miles above Earth, where they collect sunlight around the clock and beam it back as near-infrared light to ground-based solar installations. The first orbital demonstration is planned for 2028. Commercial delivery to the U.S. grid could start as early as 2030.

The second deal is with Noon Energy, reserving up to 1 GW and 100 gigawatt-hours of ultra-long-duration storage. Noon’s technology uses reversible solid oxide fuel cells to store carbon-based energy for more than 100 hours, which is far beyond what lithium-ion batteries can do today. A 25 MW pilot project is targeting 2028.

Both of these are pre-commercial. Meta is not plugging servers into satellites next year. But the company has now staked positions in space solar, long-duration storage, nuclear, and geothermal simultaneously. That is not a bet on one technology. That is a hedge against every possible failure in the grid buildout. Watch how other hyperscalers respond. The race for firm power is producing some genuinely strange infrastructure.

(Bloomberg)

IEA: AI Data Center Electricity Surged 50% in 2025

The International Energy Agency published its “Key Questions on Energy and AI” report on April 24, showing that electricity use from AI-focused data centers grew 50% in 2025, against overall global electricity demand growth of 3%. Total data center electricity consumption rose 17% to around 415 terawatt-hours, roughly equivalent to the entire electricity consumption of France. Consumption is on track to double by 2030. The five largest tech companies spent more than $400 billion on capital expenditures in 2025 and are expected to increase that by 75% in 2026.

The IEA noted that planning and regulatory systems worldwide are being “stretched” by the wave of project applications. Supply chain bottlenecks across chips and energy equipment have tightened since the agency’s 2025 report. The numbers confirm what every utility earnings call this quarter has already said: the grid was not built for this, and the gap between AI load growth and infrastructure response is widening.

(IEA)

NextEra Q1 Earnings: 9.5 GW Gas Buildout, 21 GW Data Center Pipeline in Florida

NextEra Energy reported Q1 2026 earnings on April 23, with adjusted EPS up 10% year over year to $1.09. The headline was its data center pipeline. CEO John Ketchum said the company has 21 gigawatts of large-load interest in Florida alone, with 12 GW already in advanced discussions. NextEra is simultaneously developing 9.5 GW of new natural gas capacity in Texas and Pennsylvania to serve data center hubs, part of a Japan-backed initiative. The company added a record 4 GW of new renewables and storage to its contracted backlog in the quarter, bringing total backlog to 33 GW.

The 15-by-35 goal, securing 15 GW of new generation for large loads by 2035, now has an upside case of 30 GW. NextEra is positioning itself as the utility that builds power plants for AI. (Motley Fool)

Fermi America Implodes After CEO Exit and Missing Its First Tenant

Fermi America, the startup that raised investor excitement with plans to build 11 gigawatts of nuclear, solar, and gas for a Trump-branded AI campus near Amarillo, Texas, is in serious trouble. The company’s co-founder and CEO Toby Neugebauer and CFO Miles Everson departed abruptly in April. The stock ended the week 81% below its October 2025 IPO price. Construction has stalled. Fermi never secured a confirmed hyperscaler tenant. The Washington Post reported April 28 that shareholder lawsuits accuse the company of overhyping its prospects.

This is the first significant public implosion of a company built on the AI power buildout story. The failure reveals the gap between investor appetite for the sector and the operational reality of actually building new generation. (Washington Post)

MARA Holdings Acquires 505 MW Gas Plant in Ohio for $1.5 Billion

On April 30, MARA Holdings announced it will acquire Long Ridge Energy and Power from FTAI Infrastructure for approximately $1.5 billion. Long Ridge owns a 505-megawatt combined-cycle natural gas plant in Hannibal, Ohio, along with over 1,600 contiguous acres of land for a digital infrastructure campus. MARA is primarily known as a Bitcoin mining company and says the deal increases its digital infrastructure capacity by 65%. The transaction requires FERC approval and is expected to close in the second half of 2026.

Bitcoin miners and AI data centers increasingly compete for the same power infrastructure. MARA is turning a generation asset into a long-term platform. (CoinDesk)

Massachusetts Signs Contracts Saving $1.4 Billion on Offshore Wind

Massachusetts announced on May 1 that it has signed contracts with a developer consortium for up to 2,400 megawatts of new offshore wind capacity, saving the state an estimated $1.4 billion compared to prior offshore wind procurement costs. The deal reflects falling offshore wind construction and financing costs. Offshore wind capacity is growing even as some federal lease activity has been frozen under the current administration’s energy policy. Massachusetts is the first state to publicly quantify the savings from the latest round of procurement.

(Electrek)

California Battery Storage Hits 12,000 MW During Peak Hours

California’s grid-scale battery system discharged a record 12,000 megawatts during peak evening hours last week, a level the state says is equivalent to the output of 12 large nuclear power plants and covered more than 40% of statewide demand at that moment. The milestone reflects explosive growth in battery storage deployment, driven by the need to capture daytime solar and release it after sunset when demand peaks. California’s installed battery storage has grown by nearly 2,000% since 2019.

The context for the rest of the country: California is proving batteries can carry meaningful grid load at scale. Whether that translates to regions with less solar and more AI compute concentrated in single campuses remains an open question. (Inside Climate News)

Have fun this week,

Will

Sources

AI Power News 4/27/2026

Will McKnight — Mon, 27 Apr 2026 17:38:52 GMT

Microsoft Loses Its Exclusive Hold on OpenAI

Microsoft and OpenAI announced Monday they have rewritten their seven-year partnership, ending Microsoft’s exclusive right to sell OpenAI models. OpenAI can now run its products on Amazon, Google, or any other cloud provider, though Azure stays the default unless Microsoft cannot deliver what a workload needs. Microsoft keeps a non-exclusive license to OpenAI intellectual property through 2032 and continues to collect a capped revenue share through 2030. Microsoft shares fell nearly 3% on the announcement. Alphabet and Amazon ticked up.

Every grid forecast and substation buildout for the past three years assumed OpenAI compute would concentrate inside Microsoft’s data center footprint. That assumption just expired. Power planners now have to model OpenAI workloads landing in any of three hyperscale clouds, which means more parallel power deals, not fewer.

Watch which AWS or Google Cloud site brings the first OpenAI workload online, because the hyperscaler that moves first resets the price for every AI power deal that follows. (Bloomberg)

Duke Energy Files the Largest Utility Capital Plan in U.S. History

Duke Energy unveiled a $103 billion five-year capital plan on April 24, the largest spending program ever filed by a regulated U.S. utility. Roughly 60% of the budget funds new generation, including 5 GW of new natural gas in the Carolinas and Indiana and 4.5 GW of new battery storage by 2031. Since November, Duke has signed 1.5 GW in new data center service agreements with customers including Microsoft and Compass, lifting committed data center load to 4.5 GW. Data centers will drive roughly 75% of the economic-development load growth Duke projects through 2030. (Fortune)

Siemens Energy Raises 2026 Outlook on AI Power Demand

Siemens Energy raised its full-year guidance on April 23, citing AI data center demand. The company now expects revenue growth of 14% to 16%, net profit around €4 billion, and free cash flow before tax of roughly €8 billion, nearly double the prior forecast. Second-quarter order intake hit a record €17.75 billion, with Grid Technologies orders up 41% and Gas Services up 32% year over year. Grid Technologies is the unit that builds the transformers, switchgear, and high-voltage equipment that connects new generation to large industrial loads. (Bloomberg)

Kevin O’Leary’s 7.5 GW Wonder Valley Project Nears Approval in Utah

The Stratos Project, a hyperscale data center campus on 40,000 acres in Box Elder County, Utah, moved closer to final approval this week. The buildout proposes 7.5 gigawatts of new power generation, almost twice the entire state’s current electricity consumption of about 4 GW. Kevin O’Leary’s O’Leary Digital is the developer, with West GenCo handling permitting and natural gas access through interstate pipelines. Financing terms, anchor tenants, and the actual gas supply for a project this size remain unannounced. (Salt Lake Tribune)

FERC Punts Large-Load Interconnection Rule to June

The Federal Energy Regulatory Commission confirmed last week it will rule on data center interconnection reform in June, missing the April 30 deadline that Energy Secretary Chris Wright set in February. The proposed rule would let large loads of 75 MW or more file joint interconnection requests alongside new generation, compressing what is now a multi-year sequential process into a single review. State regulators at NARUC have warned the federal proposal encroaches on traditional state authority over retail load and rate design. The two-month delay is short, but it pushes any actual queue impact into late 2026 at earliest. (Utility Dive)

U.S. in Talks with Engie to Cancel Three More Offshore Wind Leases

Engie confirmed on April 26 that it is in talks with the U.S. government to cancel its three remaining offshore wind leases, following TotalEnergies’ lease surrender in March. The discussions stem from the Interior Department’s December 2025 stop-work order on all five offshore wind projects under construction, an order currently blocked by temporary injunctions in five federal courts. Every canceled lease removes potential offshore generation from the East Coast queue, where regional grid operator PJM has flagged a capacity shortfall heading into the early 2030s. (Maritime Executive)

The AI Power Bill Has Reached Voters

A CBS News investigation on April 25 profiled Atlanta resident Carolyn Kayne, whose monthly electricity bill has nearly doubled in two years as Georgia Power filed six rate hikes in three years, citing data center load growth. A Fortune poll published the same week found a majority of U.S. households now connect data center expansion with rising electricity costs. Public sentiment is becoming the new constraint that hyperscalers will spend the rest of 2026 trying to manage. (CBS News)

Have fun this week,

Will

Sources

AI Power News 4/20/2026

Will McKnight — Mon, 20 Apr 2026 19:04:25 GMT

Oracle and Bloom Energy Expand Fuel Cell Deal to 2.8 GW

Oracle agreed Monday to purchase up to 2.8 gigawatts of power from Bloom Energy, the California fuel cell maker, to supply AI data centers across the United States. Fuel cells generate electricity from natural gas through a chemical reaction rather than combustion, producing power on-site without the grid connections and permitting timelines that can stretch for years. An initial 1.2 GW is already contracted and deploying now, with the rest following through next year. Bloom secured $7.65 billion in data center contracts across just 90 days in early 2026.

This deal is not really about power supply. It is evidence that the grid interconnection queue is so broken that building your own power plant on-site is now the faster path. Bloom delivered Oracle’s first fuel cell installation in 55 days, against a 90-day target. That is the benchmark grid utilities now have to compete with. The hyperscalers are quietly becoming their own utilities.

Watch whether Amazon and Microsoft accelerate their own off-grid strategies in response to Oracle’s lead. (Bloomberg)

ERCOT Projects Texas Grid Could Quadruple to 368 GW by 2032

Texas grid operator ERCOT told state lawmakers this week that peak electricity demand, currently at a record 85,508 megawatts, could nearly quadruple to 367,790 megawatts by 2032, driven almost entirely by data center construction requests. ERCOT, the independent system operator that manages the power grid for most of Texas, is responsible for matching electricity supply with demand in real time. CEO Pablo Vegas added a key caveat: these numbers represent every company that has filed to connect to the grid, not every company that will actually build. ERCOT is now replacing its old interconnection review process with a system called Batch Zero, which groups large new requests of 75 MW or more into batches for faster evaluation. The Federal Reserve Bank of Dallas estimates that wholesale power prices in Texas could rise as much as 50% from data center demand growth alone. (Insurance Journal)

EIA Projects US Power Demand Will Hit Records in Both 2026 and 2027

The U.S. Energy Information Administration projected this month that American electricity consumption will climb from a record 4,195 billion kilowatt-hours in 2025 to 4,260 billion kWh in 2026 and 4,388 billion kWh in 2027, driven primarily by AI data centers. If the forecast holds, it would mark the first time since 2007 that U.S. power demand has risen for four consecutive years, and the strongest four-year growth streak since 2000. The Department of Energy separately estimated that data centers accounted for 4.4% of all U.S. electricity use in 2023 and expects that share to reach 12% by 2028. For context: at 12%, data centers alone would consume roughly as much power as every home in the country combined. (EnergyNow)

Judge Orders GE Vernova to Stay in Vineyard Wind Contract

A Massachusetts judge granted Vineyard Wind a preliminary injunction this week, blocking GE Vernova from walking away from its turbine supply and maintenance contract. GE had sent a termination notice in February, citing more than $300 million in unpaid bills. Vineyard Wind is the first utility-scale offshore wind project in the United States and has faced construction setbacks, including a blade failure in 2024. The injunction keeps GE crews on-site through the legal proceedings and prevents a shutdown that would have stalled the project indefinitely. A separate court ruling earlier this month also cleared five major U.S. offshore wind projects to continue construction after the Trump administration missed a final appeal deadline. (Vineyard Gazette)

Eos Energy and TURBINE-X Target AI Data Centers with On-Site Storage

Eos Energy Enterprises and TURBINE-X Energy announced a joint development agreement Tuesday to deploy up to 2 gigawatt-hours of Eos zinc battery storage paired with on-site gas turbines at AI data centers, with first installations planned for 2027. Zinc batteries use more abundant and less expensive materials than lithium-ion cells and are designed for longer discharge cycles, making them suited to the sustained, high-draw loads that AI workloads generate. The pitch is speed: the partnership claims it can deliver hyperscale power capacity in months rather than the years required for traditional grid interconnections. Eos shares rose 12% on the announcement. (Globe Newswire)

Fermi CEO and CFO Exit as Project Matador Loses Momentum

Fermi Inc. lost both its co-founder CEO Toby Neugebauer and CFO Miles Everson in the span of two days this week, sending shares down 22%. Fermi is developing Project Matador, a planned gigawatt-scale AI data center campus in Texas that was positioned as one of the most ambitious buildouts in the country. The departures follow months of friction: a tenant pulled out in December, investors filed a class-action lawsuit, and executives faced pointed questions on the March 30 earnings call about the company’s inability to name a single signed anchor tenant publicly. Fermi announced a pivot to a “Fermi 2.0” strategy, creating an Office of the CEO led by two co-presidents and launching a search for a permanent chief executive. The company also disclosed it no longer expects to hit its original target of 1.1 gigawatts of capacity by end of 2026. (Axios)

Gas Turbine Shortage Is Slowing the AI Power Buildout

The rush to build natural gas power plants for AI data centers has run into a hard physical constraint: there are only three major gas turbine manufacturers in the world, none of which invested to scale during the previous decade of flat power demand. GE Vernova ended 2025 with an 80 GW backlog stretching into 2029. Siemens Energy is carrying a €136 billion order backlog, the largest in its history. Wood Mackenzie estimates turbine prices will rise 195% by end of 2026 compared to 2019 levels. Even if all three manufacturers deliver on their expansion plans, analysts estimate total output will rise only 20% to 25%, nowhere near enough to absorb the current wave of demand. The bottleneck is not just the grid. (Bloomberg)

DOE Directs FERC to Fast-Track Data Center Grid Connections

Energy Secretary Chris Wright directed the Federal Energy Regulatory Commission this month to initiate a fast-track rulemaking that would allow data centers and other large power users to file joint interconnection requests alongside new generation sources. FERC, the federal regulator that oversees interstate electricity transmission, has a deadline of April 30, 2026 to respond to the directive. The proposed rule would allow large load customers to request co-located generation and grid connections simultaneously, cutting study times and reducing the cost of required network upgrades. It would also require those customers to fund the full cost of the grid infrastructure they need, a provision aimed at ensuring other ratepayers are not left with the bill. (Gibson Dunn)

U.S. Utilities Announce $1.4 Trillion in Capital Spending Through 2030

A new analysis of 51 U.S. investor-owned utilities, released this week, found they plan to spend a combined $1.4 trillion on infrastructure through 2030, a 27% jump from last year’s $1.1 trillion projection and roughly double the $700 billion invested across the entire previous decade. Duke Energy has committed $102.2 billion and Southern Company pledged $81.2 billion. The spending covers new power plants, transmission line upgrades, distribution modernization, and grid hardening. The surge reflects utilities trying to keep pace with data center load requests that are arriving faster than steel can be put in the ground. For energy investors, this is the clearest signal yet of where capital is flowing at scale. (Tech Insider)

Have fun this week

Will

Sources

AI Power News 4/13/26

Will McKnight — Mon, 13 Apr 2026 17:57:55 GMT

Meta Funds 10 Gas Plants for Its Louisiana AI Campus

Meta has agreed to finance 10 natural gas power plants totaling 7.5 gigawatts of generation capacity to support its $27 billion Hyperion data center campus in Richland Parish, Louisiana. That figure is more than a 30% increase to Louisiana’s entire current grid capacity, and it does not count the additional 2.5 GW of solar and battery storage Meta also agreed to fund. The deal, struck with utility Entergy Louisiana, also includes roughly 240 miles of new 500-kilovolt transmission lines (the high-voltage highways that carry power long distances across the grid). The original plan, approved just last year, called for three gas plants. Meta tripled it. (Bloomberg)

Meta built a regional power system and handed the operating license to Entergy. At some point the distinction between tech company and energy developer stops meaning anything, and Hyperion is past that point.

Watch whether the Louisiana Public Service Commission approves the gas units on schedule, and whether other states start asking why their utilities do not have the same arrangement.

Microsoft and Nvidia Launch AI for Nuclear Initiative

At CERAWeek in Houston on March 24, Microsoft and Nvidia announced a joint initiative to use AI tools to speed up nuclear power plant construction. The collaboration targets three specific bottlenecks: permitting, engineering design, and plant operations. AI tools will flag documentation errors, unify data across a plant’s full lifecycle, and support digital twins (virtual replicas that let engineers run tests before touching physical hardware). Early results are meaningful: nuclear startup Aalo Atomics cut its permitting timeline by 92% using Microsoft’s AI tools and estimates it saves $80 million per year as a result. (Axios)

Permitting delays are one of the biggest reasons nuclear plants take 10 to 20 years to build in the United States. If AI can compress that meaningfully, it changes the economics of every advanced reactor project in the queue.

Meta Signs 6.6 GW of Nuclear Deals to Power Its AI Supercluster

Meta signed agreements with Vistra, TerraPower, and Oklo to add 6.6 gigawatts of nuclear capacity by 2035, enough to power the Prometheus AI supercluster the company is building. The deals include expanding existing nuclear plant operations, funding advanced reactor development, and backing Oklo, a small modular reactor startup (SMR means a compact, factory-built nuclear reactor designed to be deployed faster and cheaper than a traditional plant). The combined scale makes Meta one of the largest corporate nuclear buyers in US history. (CNBC)

Read alongside the Louisiana gas story, Meta’s energy strategy comes into focus: gas plants to cover the next few years, nuclear contracts locked in for 2035. The company is moving faster on both fronts than most utilities move on either.

TotalEnergies Exits US Offshore Wind, Redirects $1B to Oil and Gas

The Department of the Interior announced on March 23 that TotalEnergies, the French energy major, agreed to surrender its US offshore wind leases. In exchange, the company committed to invest $1 billion, roughly the stated value of those leases, in oil, natural gas, and LNG production inside the United States. (NPR)

Offshore wind, which generates electricity from turbines built in the ocean, has faced a sustained federal effort to slow its development since early 2025. The TotalEnergies agreement is the most significant exit yet. Meanwhile, federal judges this week refused to stop construction on five other wind farms, including Dominion Energy’s Coastal Virginia Offshore Wind project, which began delivering power to the grid on Monday. The two stories running in parallel capture where offshore wind actually stands: federally hostile at the policy level, but too far along at the project level to unwind entirely.

DOE Uses Emergency Powers to Freeze Coal Plant Retirements

The Department of Energy has now issued at least 16 emergency orders since May 2025 directing utilities to keep coal plants running past their planned closure dates, using a rarely invoked authority called Section 202(c) of the Federal Power Act. The most recent orders cover more than 2 GW of coal capacity across the Midwest, Mountain West, and Pacific Northwest. (POWER Magazine)

Earthjustice filed a legal challenge to one Indiana order this month, calling it a federal override of decisions utilities made on their own timelines for sound business reasons. The DOE’s stated justification is grid reliability. The EIA (the federal agency that tracks energy data) now expects retirement delays to continue throughout 2026. For grid planners and utility engineers, the practical problem is real: generation is being kept online without a long-term plan for what replaces it. (EIA)

Xcel Energy Proposes Large-Load Tariff for Data Centers

Xcel Energy filed a proposal with Colorado regulators in early April to create a separate rate tier for customers consuming more than 50 megawatts of power at peak, a threshold that targets data centers specifically. The structure requires a $120,000 upfront deposit plus full payment for load and transmission studies before a project is approved. The goal is to stop residential customers from subsidizing the grid upgrades that large industrial users require. (Colorado Sun)

77 large-load tariffs are now pending or in place across 36 states, up from just 14 approved between 2018 and 2024. State regulators are moving faster than federal policy on this. For utility executives and rate case teams, designing a large-load structure that survives a commission challenge is the immediate problem. The whether is settled.

Crusoe Signs Two Energy Storage Deals for AI Data Center Buildout

AI data center developer Crusoe announced back-to-back storage partnerships this month. The first expands a battery microgrid deployment with Redwood Materials from 4 to 24 Crusoe Spark modular data centers. The second is a deal with Form Energy for up to 12 GWh of multi-day iron-air battery storage, with deliveries beginning in 2027. Multi-day storage is a battery technology that can discharge continuously for 100 hours or more, unlike a standard lithium-ion battery which typically runs for 4 hours. (Data Center Dynamics)

Data centers need power that never stops. Multi-day storage is one of the few technologies that can cover an extended grid outage without a diesel generator. This deal is an early signal that hyperscale operators are starting to treat long-duration storage as infrastructure, not a backup.

Big Tech Accelerates Investment in Next-Generation Nuclear

Microsoft, Google, Amazon, and Meta are all now signing long-term power purchase agreements (PPAs) with nuclear developers before construction begins, giving those developers contracted revenue early enough to attract private financing. That changes the risk math that has historically made nuclear hard to fund privately. The IAEA said this month that nuclear is the only energy source that checks all five boxes AI data centers require: low-carbon, round-the-clock, high power density, grid-stable, and scalable at the speeds the industry needs. (Insurance Journal)

Coastal Virginia Offshore Wind Delivers First Power Despite Federal Headwinds

Dominion Energy confirmed Monday that its Coastal Virginia Offshore Wind project began delivering electricity to the grid, a significant milestone for the largest offshore wind project under active construction in the United States. The Trump administration issued a stop-work order on the project earlier this year. Federal judges declined to enforce it after Dominion showed the project was too far along to halt without serious safety and financial consequences. (NPR)

Dominion has ratepayers and bond markets committed to this project in ways that make reversal genuinely costly. The first power delivery does not end the political risk, but it does make the project harder to stop. Watch for the administration’s next move and whether Dominion faces permit challenges at the federal level as it scales up capacity.

Have fun this week,

Will

Sources

AI Power Weekly News Roundup 3/30

Will McKnight — Mon, 30 Mar 2026 19:37:23 GMT

SoftBank and DOE Announce 10 GW Data Center Campus on Former Ohio Uranium Site

The Department of Energy, SoftBank, and AEP Ohio announced a public-private partnership to build a 10-gigawatt data center campus at the former Portsmouth Gaseous Diffusion Plant in Pike County, Ohio, a Cold War-era uranium enrichment facility now owned by the federal government. SB Energy, SoftBank’s energy arm, will build 9.2 GW of new natural gas generation to power the campus, while SoftBank is investing $4.2 billion with AEP Ohio to build and upgrade transmission lines across southern Ohio.

At full buildout, 10 GW would be more than half the total operating capacity of every existing U.S. data center combined. The project is part of Japan’s $550 billion U.S. investment commitment made during tariff negotiations with the Trump administration.

The Portsmouth site is a genuinely clever move. The federal land is already cleared, the Cold War infrastructure footprint means transmission access exists, and the symbolic resonance of powering AI on a former nuclear weapons site is not lost on this administration. The natural gas-heavy power plan will draw fire from clean energy advocates, but the scale signals the federal government is treating AI infrastructure as a national priority on par with the original Manhattan Project facilities.

Watch for state regulators and environmental groups in Ohio challenging the gas generation buildout, and how this competes with existing PJM interconnection queue projects for equipment and contractors. (DOE)

Trump Administration Pays TotalEnergies $928 Million to Exit Offshore Wind

The Department of the Interior and TotalEnergies signed agreements March 23 to relinquish two offshore wind lease areas, including the Carolina Long Bay lease with roughly 1,300 MW of potential capacity off the North Carolina coast. The government will reimburse TotalEnergies $928 million in lease fees paid in 2022, dollar for dollar, on the condition TotalEnergies invest that amount in U.S. oil, gas, and LNG projects including Rio Grande LNG in Texas. TotalEnergies pledged not to pursue any new offshore wind projects in the United States. This sets a precedent every other offshore wind leaseholder is watching closely, and the loss of 1,300 MW off the Carolinas forces grid planners to rethink Southeast coastal power planning. (Utility Dive)

Seven Hyperscalers Sign White House Ratepayer Protection Pledge

Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI signed the White House’s Ratepayer Protection Pledge in early March, committing to fund all new power generation and grid infrastructure required by their data centers at no cost to residential electricity customers. Signatories must negotiate separate rate structures directly with utilities and pay those rates whether they consume the electricity or not, and must make backup generation available to grid operators during emergencies. This is the first formal, White House-brokered commitment from hyperscalers to bear the full cost of AI’s electricity buildout. It directly changes how utilities like AEP, Duke, and Dominion approach future interconnection negotiations. (White House)

X-energy and Talen Energy Sign SMR Agreement for PJM Market

X-energy and Talen Energy signed a Letter of Intent March 19 to evaluate deploying three or more four-unit XE-100 small modular reactor plants across Pennsylvania and the PJM Interconnection market, the grid serving 65 million people from Illinois to New Jersey. The XE-100 is an 80-megawatt high-temperature gas-cooled reactor, a design that uses helium instead of water as a coolant and can generate heat for industrial processes in addition to electricity. Talen already operates the Susquehanna nuclear station and has direct experience structuring hyperscale power deals. Talen’s stock rose 7.1% on the announcement. (World Nuclear News)

Bipartisan REWIRE Act Would Fast-Track Grid Capacity Upgrades

Senators Dave McCormick (R-PA) and Peter Welch (D-VT) introduced the REWIRE Act in early March, creating a categorical exclusion from full environmental review for grid capacity projects built within existing rights-of-way. Reconductoring, upgrading existing transmission towers with advanced high-capacity wires that can double or triple power flow without new construction, could cut U.S. grid infrastructure costs by $85 billion by 2035 according to industry estimates. The bill also directs FERC to increase financial returns for utilities that deploy advanced conductors, giving them a direct incentive to move faster. Transmission permitting is the single biggest bottleneck between new power generation and the load growth AI is driving. (Senator Welch)

FERC Approves 11 New Cybersecurity Standards for the Power Grid

FERC unanimously approved 11 updated Critical Infrastructure Protection standards March 19, enabling secure use of virtualization technologies across the high-voltage transmission system and tightening baseline cybersecurity requirements for all grid-connected systems. New mandates include updated password protocols for remote access and intrusion detection requirements for lower-impact grid assets that were previously lightly regulated. As AI data centers connect to the grid in ever-larger blocks, a cyberattack on a control system near a major data center cluster is a scenario utilities and operators now model explicitly. The updated standards are the regulatory foundation that lets utilities safely adopt the software-defined grid management tools now in wide deployment. (FERC)

Have fun this week.

- Will

Sources

Top AI Power Weekly News - week of 3/16/26

Will McKnight — Wed, 18 Mar 2026 16:17:21 GMT

Iran Strikes Three AWS Data Centers in UAE and Bahrain

On March 3, Iranian drones struck three Amazon Web Services data centers in the Middle East, hitting two facilities in the United Arab Emirates and one in Bahrain. The attacks caused structural damage, disrupted power delivery, and triggered fire suppression systems that added water damage on top. AWS, or Amazon Web Services, is one of the giant companies that runs cloud computing infrastructure for thousands of businesses worldwide. This was the first confirmed military strike on a hyperscale cloud provider. (Fortune)

Banking apps including ADCB and Emirates NBD went dark. Payments platforms Alaan and Hubpay failed. Ride-hailing service Careem reported outages across the Gulf. Iran’s state media explicitly framed the Bahrain facility as a legitimate military target for “supporting U.S. military and intelligence activities.” AWS told customers to migrate workloads to unaffected regions immediately. (CNBC)

The cloud industry built redundancy for software failures and power outages. It is not clear the same thinking has been applied to drone warfare. Every enterprise with workloads in geopolitically exposed regions now has a harder version of the risk question to answer, and cloud providers do not have a clean answer ready. (Bloomberg)

Watch for whether AWS announces plans to relocate or reinforce Middle East infrastructure, and whether other hyperscalers follow with formal geopolitical risk assessments of their own data center portfolios.

Revolution Wind and Vineyard Wind Reach the Grid in the Same Week

On March 16, two major U.S. offshore wind projects crossed the finish line on the same day. The 704 MW Revolution Wind farm, a joint venture between Ørsted and Skyborn Renewables, began delivering electricity to the New England grid. Connecticut’s energy regulators estimate it will reduce New England’s wholesale power costs by up to $500 million annually once fully operational. The 806 MW Vineyard Wind 1 off Nantucket, Massachusetts, completed installation of its final turbine, becoming the first large-scale offshore wind project ever fully constructed in the United States. (Utility Dive)

The milestones came the same day the U.S. Justice Department declined to appeal the federal injunction blocking the Trump administration’s stop-work order on offshore wind construction. Both projects reached completion despite sustained federal opposition. Together they will power roughly 750,000 homes. (Electrek)

For energy investors, the legal strategy held against an adversarial federal posture. The remaining U.S. offshore wind pipeline is watching closely to see whether that precedent sticks.

Trump Administration Weighs $1 Billion to Kill Offshore Wind Projects

Reports from mid-March indicate the Trump administration is considering offering offshore wind developers up to $1 billion in taxpayer funds to voluntarily abandon their projects. The proposal has no announced mechanism and came the same week Revolution Wind and Vineyard Wind reached the grid. (EDF)

The offer signals the administration may escalate from stop-work orders to financial inducements, a different kind of pressure with direct implications for contract enforceability. Any investor with offshore wind exposure now has to model a scenario where the federal government tries to buy its way out of the energy transition rather than just block it.

Seven Tech Giants Sign White House Ratepayer Protection Pledge

Seven major U.S. tech companies signed a White House commitment on March 4 to “build, bring, or buy” the power generation needed to run their AI data centers rather than drawing from the public grid. Residential electricity bills have risen 36% since 2020, averaging 17.44 cents per kilowatt-hour, and the administration framed the pledge as consumer protection. (The Hill)

The pledge has no enforcement mechanism. Grid rules are set state by state across 50 public utility commissions, and converting a White House announcement into binding regulatory requirements is a different problem entirely. The gap between the commitment and the policy infrastructure to enforce it is large. (Inside Climate News)

Data Center Developers Move to Finance Nuclear Directly

Meta’s deals with Vistra, TerraPower, and Oklo could unlock up to 6.6 GW of nuclear capacity by 2035, combining existing reactor restarts with next-generation designs. AI infrastructure companies are moving past signing nuclear power purchase agreements and toward financing plant construction outright. (CFACT)

The shift reflects the hard reality of nuclear timelines. Power purchase agreements signed today will not deliver electrons for a decade or more. Small modular reactors, compact designs that can be factory-built and deployed faster than conventional nuclear plants, are the main target of this direct financing push. Developers who want firm, always-on power for AI workloads are concluding that if they want it built, they need to help pay for it.

EIA Projects Record 24.3 GW of Battery Storage Additions in 2026

The U.S. Energy Information Administration projects 24.3 GW of new battery energy storage will come online in 2026, part of a record 86 GW of total new generating capacity expected this year. Battery storage, which captures electricity when supply is high and releases it when demand spikes, is now the second-largest category of new capacity behind solar. (EIA)

A 24 GW single-year addition would roughly double cumulative U.S. storage capacity. Utilities are deploying storage at scale to firm up intermittent renewables and absorb the load spikes driven by data center buildout, particularly in ERCOT (Texas) and CAISO (California). For grid reliability, storage at this scale is the piece that makes renewable-powered data centers viable at scale.

Sources

How are the Iran war and the closure of the Strait of Hormuz affecting the energy industry?

Will McKnight — Fri, 13 Mar 2026 17:35:01 GMT

When US and Israeli forces launched Operation Epic Fury on February 28, striking thousands of targets across Iran in the opening days, the strikes hit military sites, nuclear facilities, and Iranian energy infrastructure. Iran retaliated across the region. Twelve days into a war neither side has shown signs of ending quickly, the question for the global economy is what happens to the Strait of Hormuz, the 21-mile-wide channel between Iran and Oman through which roughly 20 million barrels of oil per day transit the global economy.

The answer, so far, is that the strait is effectively shut to most commercial traffic. Insurers, shipowners, and crews are unwilling to take the risk. Tanker transits have collapsed, and the downstream effects are cascading through oil markets, natural gas systems, European electricity grids, Asian economies, and hundreds of billions of dollars in Gulf AI investment simultaneously (Bloomberg).

What has the war hit, and what is the impact?

Both sides have struck energy infrastructure. The initial US and Israeli strikes targeted Iranian oil refineries and export terminals alongside military sites, hitting Iran’s own production capacity in the opening hours. Iran’s retaliatory strikes have spread across the region: drone attacks hit QatarEnergy’s LNG production facilities, forcing Qatar to declare force majeure on its gas exports. Force majeure is a legal term for “we cannot perform our contracts due to events outside our control,” and it means buyers around the world are suddenly without supply they were counting on. Qatar supplies 20 percent of global liquefied natural gas (LNG is natural gas that has been chilled to liquid form for transport by ship), and this shutdown happened with essentially no warning (Al Jazeera).

Strikes also hit Amazon Web Services data centers in the UAE and Bahrain, damaging two facilities and hitting a third with debris. This is believed to be the first known deliberate military strike on commercial hyperscale data centers (Yahoo News).

The pattern across all of this is the same: modern conflict does not stay on the battlefield. The same war taking lives in the region is now degrading energy supply and targeting the physical infrastructure of the global AI industry. That is a new reality, and it matters for how you think about what comes next.

Which energy sectors are bleeding fastest?

Oil markets took the first and hardest hit. Brent crude, the global benchmark, has surged more than 57 percent since the war began, jumping from roughly $70 to over $110 per barrel driven almost entirely by supply fear, not actual shortfall. Rystad Energy puts a four-month Hormuz disruption scenario at $135 per barrel (CNBC). The IEA announced a coordinated release of 400 million barrels from member nations’ emergency reserves, which put a ceiling on the panic for now. US gasoline prices have risen roughly 60 cents per gallon since the war began, reaching a national average of around $3.58 (NPR).

LNG markets are arguably worse. Oil has strategic reserves and diversified shipping lanes. LNG does not have the same slack. European LNG benchmark prices (the TTF hub, Europe’s main gas trading point) climbed from €31.9 to €54.3 per megawatt-hour in less than a week. Asian LNG futures jumped 51 percent. European prices jumped 77 percent (CNBC). These are energy-price moves that normally take months or years.

Electricity grids feel it through gas. In most countries, gas-fired power plants are what the grid calls on last, when demand is highest. The final unit of power on the grid in any given hour sets the price for everyone. When gas prices spike 77 percent, electricity prices follow. In Spain, wholesale electricity prices jumped nearly 700 percent in a single week, from roughly €18 to €137 per megawatt-hour (The Olive Press).

Who is most exposed, and who has cushion?

Asia is the most exposed region by far. South Korea maintains a mandatory nine-day minimum LNG stockpile requirement and has announced emergency fuel price caps, the first since the late 1990s, alongside a 100 trillion won ($68 billion) stabilization fund. The country is highly exposed: any sustained disruption to Gulf LNG supply pushes it toward that threshold fast. Government offices across Southeast Asia are cutting work weeks, limiting travel, and imposing alternating driving days (Al Jazeera). Japan and South Korea are scrambling for spot LNG from Australia, Canada, and the US, but alternative cargoes take weeks to arrange.

India is the most vulnerable major economy. The country imports 90 percent of its crude oil, holds roughly 45 days of strategic reserves, and has 9 million citizens working in Gulf countries affected by the conflict. The combination of oil import costs, remittance disruption, and inflation risk is severe (New Lines Institute).

China has the most buffer. Beijing holds an estimated 1.2 billion barrels of strategic crude reserves, roughly 108 days of imports at current refinery levels. More importantly, oil and gas make up a relatively small share of China’s power mix compared with most other Asian economies. China’s aggressive solar and wind buildout, often criticized as overbuilt, turns out to be strategic insulation (CNBC).

Europe is in an uncomfortable middle ground. The EU has diversified its gas supply significantly since Russia invaded Ukraine in 2022, and storage levels are healthier. But the bloc is re-exposed through Qatar, which became a key LNG supplier precisely because it replaced Russian pipeline gas. EU officials are quick to say 2026 is not 2022. What they mean is supply is more diversified. What they cannot say is that pricing will not hurt (Euronews).

What does this mean for the AI industry’s energy buildout?

The Gulf was in the middle of a historic AI construction boom when the war started. Saudi Arabia, the UAE, and Qatar collectively committed hundreds of billions of dollars to data centers, chips, and AI infrastructure over the coming years. That buildout now has a targeting problem: the AWS strikes in the UAE demonstrated that data centers are viable military targets, something no hyperscaler had factored into its Gulf infrastructure risk calculus (The Information).

Existing projects will likely continue. Walking away from sunk capital is expensive, and the hyperscalers have long-term government relationships in the region. But the next wave of capacity decisions, the facilities that were still in site-selection and negotiation phases, may quietly shift to locations in Europe, Southeast Asia, or North America. The war has introduced a physical security variable that no lease rate or power deal can easily offset (CNBC).

For the US AI buildout specifically, the pain is indirect but real. Electricity accounts for roughly half of a data center’s operating costs. When energy prices spike globally, the economics of expansion tighten. And the conflict is straining semiconductor supply chains: the Middle East supplies critical inputs including helium and bromine used in chip manufacturing, and logistics disruptions are adding supply-side cost pressure to data center operators at exactly the wrong time (CNBC).

This conflict is twelve days old. The Strait of Hormuz, which has never been fully closed in modern history, is effectively closed now. How long it stays that way will determine whether this is a price shock or a structural reshaping of global energy systems.

Stay safe and have a nice weekend,

Will

Sources

AI Power Update March 2nd-9th

Will McKnight — Mon, 09 Mar 2026 14:28:53 GMT

Tech Giants Sign White House Ratepayer Protection Pledge

Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI gathered at the White House on March 4 to sign what the Trump administration called a “Ratepayer Protection Pledge.” Each company committed to build, buy, or bring their own power supply for new AI data centers, rather than drawing from the shared grid and passing those costs along to homes and businesses. The political backdrop matters here: data centers have already added $7.7 billion in transmission costs for ordinary Americans over the past two years, and electricity prices are forecast to rise 6% through 2026 as AI demand outpaces supply.

The problem is the pledge is voluntary, with no enforcement mechanism and no clarity on what actually counts as “building your own power.” No Congressional role was specified. No state regulator role was specified. Trump’s statement at the signing amounted to: “They can build their own power plants as part of their factory, so that no one’s prices will go up.” One energy attorney described the implementation as “complicated...much more complicated than Trump made it sound.”

This is a PR commitment that puts real public pressure on these companies in a way that will be hard to walk back when rate cases come up at state utility commissions. That’s not nothing. But the enforcement gap is wide enough to drive a data center through, and every one of the seven signatories has ongoing grid interconnection requests that this pledge does nothing to change.

Watch: Whether state utility commissions treat this pledge as binding evidence when tech companies next apply to connect new facilities to the grid.

(Inside Climate News)

California Considers Lifting Its 50-Year Nuclear Moratorium

California lawmakers introduced legislation last month that would allow the state to approve advanced nuclear reactor technologies already licensed by the federal government since 2005. California’s moratorium on new nuclear construction has been in place since 1976. The immediate driver is AI: data centers requesting grid connections in the state have already triggered questions about costly grid upgrades, and California’s goal of 90% clean electricity by 2035 leaves very little room for new gas plants. Separately, a push to regulate data center energy use was reduced to a study requirement after Big Tech lobbied against stronger rules.

(Insurance Journal)

INL and NVIDIA Launch Prometheus Project to Halve Nuclear Deployment Timelines

The Idaho National Laboratory and NVIDIA announced a partnership called Prometheus in February, the first project under the Department of Energy’s Genesis Mission, a national initiative to accelerate discovery science and energy innovation. The goal is to use generative AI and digital twins to design, license, manufacture, and operate nuclear reactors. Digital twins are virtual replicas of physical systems that let engineers test scenarios in software before committing to expensive real-world builds. The targets are ambitious: at least 2x faster deployment schedules and more than 50% reduction in operational costs compared to today’s nuclear projects.

(POWER Magazine)

Five States Challenge Midwest Grid’s $22B Transmission Portfolio

Utility regulators from Arkansas, Louisiana, Mississippi, Montana, and North Dakota filed a complaint with FERC, the Federal Energy Regulatory Commission, asking it to unwind a $22 billion set of high-voltage transmission projects approved by MISO, the grid operator covering the Midwest and parts of the South. MISO calculated the portfolio’s benefits at well above its cost. The states’ own consultant put the actual benefits at $4.3 to $7.2 billion, a fraction of MISO’s estimate. FERC has not yet ruled, and the complaint is facing pushback from MISO itself, six other state utility commissions, Ameren, Xcel Energy, and the Data Center Coalition.

(Utility Dive)

FERC to Finalize Data Center Grid Connection Rules by April 30

FERC has set an April 30 deadline to finalize rules governing how large power loads connect to the transmission grid. The rules apply to any facility pulling more than 20 megawatts, which covers virtually every serious data center. The decisions will determine how much security deposits new data centers must put up, how quickly they can advance through the interconnection queue, and how grid costs are allocated when a large load triggers expensive upgrades. Data centers are now the dominant driver of new interconnection requests across every major U.S. grid region.

(CSIS)

AI Transmission Buildout Draws Organized Landowner Pushback

As utilities race to build the high-voltage lines needed to connect new AI data centers, landowners and local communities are pushing back across the country. In the Midwest, the $22 billion MISO transmission package is caught in a monthslong political fight. In other regions, opposition groups are filing complaints against dozens of proposed routes, citing damage to farmland, waterways, and property values. Transmission spending is projected to nearly double to $50 billion per year by 2028, and without a faster permitting process, the buildout that AI infrastructure depends on may be slower than the industry expects.

(AP via U.S. News)

Go forth and have fun this week,

Will

Sources

Using AI Agents in the Power Industry: A Beginner’s Guide

Will McKnight — Thu, 05 Mar 2026 19:08:48 GMT

Most people think building software requires knowing how to code. That was true a year ago. It is not true anymore.

Claude Code is Anthropic’s AI system that writes, runs, and debugs code directly on your computer. You describe what you want in plain English. Claude figures out the code. You get a working tool to do almost any manual task that you’re doing in your day-to-day work.

I have been using it to build agents for my own business, and the results have surprised me in ways I did not expect. Some things worked immediately. Some things took five iterations. This is an honest account of both.

What actually is an agent, and why is it different from chatting with ChatGPT?

A regular AI conversation is like asking a smart friend for advice. You ask a question. They answer. Done.

An agent is different. An agent takes a goal, breaks it down into steps, uses tools to complete those steps, and keeps going until the job is finished. It does not just answer questions. It does things.

Think of it this way: asking Claude “who should I follow up with this week?” gets you advice. An agent actually searches your inbox, finds past contacts, groups them by topic, and hands you a prioritized list. That is the difference. Advice versus action.

Agents can browse the web, read and write files on your computer, pull data from external services, send emails, and chain all of those steps together without you touching anything after the first setup.

What can agents do for people in the energy industry?

The energy industry runs on data, documents, and repetitive processes. That is exactly where agents shine.

A load forecasting agent could pull weather data, historical usage patterns, and grid demand numbers every morning and produce a plain-English summary for your operations team. No spreadsheet wrangling. No manual data pulls.

A regulatory filing monitor is one of the most valuable builds I can imagine for energy companies. FERC (Federal Energy Regulatory Commission) governs wholesale electricity markets, natural gas pipelines, and hydropower licensing. NERC (North American Electric Reliability Corporation) issues reliability standards, violation notices, and enforcement actions. State PUCs handle rate cases and tariff filings. ISO and RTO filings from grid operators like PJM, MISO, CAISO, and ERCOT contain market rule changes that affect dispatch and pricing. EPA rulemakings affect emissions reporting and permitting for new generation. An agent that monitors all five simultaneously, filters by keywords relevant to your business, and surfaces only what matters is worth a significant amount of someone’s time each week.

An RFP comparison agent could scan incoming vendor proposals, extract the key specs and pricing from each document, and produce a comparison table automatically. What used to take a junior analyst two days takes the agent twenty minutes.

None of these require a full-time developer. They require knowing what problem you want to solve, and maybe an hour or two of back-and-forth with an LLM.

What have I built, and what actually worked?

I want to share my actual experience here, because most “here’s what AI can do” posts skip the failures.

SEO and ranking monitor. Watches my website’s search ranking and surfaces content ideas based on keywords I am not ranking for. Drafts content outlines automatically. Took two sessions to get right. Saves several hours every week now.
Pitch deck and website creator. I describe the company, audience, and message. Claude generates the full deck structure, the slide copy, and a matching website mockup with layout and design. This one worked on the first try. UI/UX and design-focused tasks are where Claude is genuinely one-shot capable right now.
Email history scraper. Scraped my entire email history, found contacts I had not followed up with in months, and grouped them by keyword. Clusters like “grid storage” or “data center development” showed up immediately. Runs every Sunday and gets me a prioritized outreach list every Monday morning. Took five iterations to handle all the edge cases.

Complex agents that chain multiple steps together rarely work perfectly on the first build. The email agent worked, but it took four or five conversations with Claude to handle edge cases: duplicate contacts, unusual email formatting, contacts with no clear topic keyword. That is just the reality of software. Even experienced developers spend most of their time debugging, not writing new code. The difference now is that you do not need to understand the code to fix it. You describe what broke and Claude fixes it.

How do you actually build one from scratch?

Here is the exact process I use. It works for complete beginners.

Go to Claude, ChatGPT, or any LLM and describe the problem in plain English. I use Claude personally, but any capable LLM works for this. Do not think about code yet. Write out exactly what you want to happen. “I want something that looks at my inbox every Monday, finds emails from people I have not replied to in 90 days, and makes a list grouped by what we were talking about.” The more specific you are, the better the result.
Ask whether it can build it and what the best approach would be. Paste your description and ask: “Is this something you can help me build as an agent? What would be the best approach?” Claude will tell you what is realistic, what tools it would use, and whether there is a simpler way to get the same result.
Ask for extremely detailed step-by-step instructions. Tell Claude: “Give me every single step, including how to install anything I need, how to set up any accounts or API keys, and how to run the code. Assume I have never done this before.” Claude will write instructions detailed enough that you do not need to figure anything out yourself.
Follow the instructions and keep the conversation open. Things will break. When they do, copy the error message and paste it back to Claude. Say: “This happened when I ran step 4. What do I do?” Claude will diagnose and fix it. You do not need to understand why it broke.
Iterate one feature at a time. After the basic version works, ask Claude to add improvements one at a time. Add one thing, test it, then add the next. Adding five features at once is how you end up with something broken in five different ways at once.

That is the whole process.

When a general-purpose agent is not the right tool

Claude and ChatGPT are excellent at text, research, email, and document workflows. But some tasks in the power industry require models trained specifically on engineering data, domain conventions, or industrial signal patterns that generic LLMs were never built for.

Dirac generates detailed, animated work instructions automatically from CAD files for complex mechanical assemblies. What used to take twelve hours of manual authoring now takes ninety minutes. Anduril selected Dirac as its core manufacturing documentation platform in January 2026 and reported an 87.5% reduction in authoring time.

SparkCognition builds AI trained specifically on SCADA data and sensor patterns from industrial energy equipment: wind turbines, power generation assets, oil and gas facilities. Their predictive maintenance system identifies equipment failures a median of 37 days in advance with over 90% confidence and no false positives. In one deployment on a fleet of new turbines, it identified a manufacturing defect that would have caused catastrophic damage to a $100 million asset months before traditional monitoring would have caught it.

OpenDrawing generates bills of materials automatically from electrical single-line diagrams. Engineers upload a drawing and the system extracts every component and quantity without manual data entry. It is built specifically for the symbol libraries and tagging conventions used in electrical construction drawings, which a general LLM cannot interpret reliably.

AutoGrid, now part of Schneider Electric’s software portfolio, manages distributed energy resources in real time using AI trained specifically on grid signal patterns and DER (distributed energy resource) behavior data. It currently manages over 5,000 MW of distributed energy resources across more than 100 utility and industrial partners. A general-purpose agent can summarize a DER report. AutoGrid is making real-time dispatch decisions at grid scale.

When does it make sense to use a purpose-built platform instead of building your own agent?

When the task requires a model trained on specific domain data: engineering drawings, industrial sensor streams, power market signals, or equipment failure patterns that general LLMs have no reliable grounding in.
When your organization requires on-premises deployment. Cloud-based LLM APIs are not always acceptable in regulated environments, for NERC CIP (Critical Infrastructure Protection) compliance, or where data sovereignty is a hard requirement.
When you need enterprise-wide deployment with SSO, role-based access controls, audit logs, and compliance certifications. Building your own agent gets one person a useful tool. A purpose-built platform gets your whole organization a vetted, supported system.
When accuracy needs to be mission-critical rather than “good enough.” A regulatory filing monitor built on Claude is genuinely useful. A platform certified for NERC compliance tracking is a different category of reliability.
When you need deep integration with existing enterprise systems like SAP, OSIsoft PI historian, or SCADA platforms that require validated data pipelines and purpose-built connectors.

Building your own agents and using purpose-built platforms are not competing approaches. Most energy teams will end up doing both. Start with one problem you deal with every week. Ask Claude if it can help. See what happens.

Good luck this week.

Will

Sources

Claude Code overview

Anthropic AI-Driven BOM Creation from Single Line Diagrams

OpenDrawing

Dirac

Production Planning, Automated

Anduril Selects Dirac to Power AI-Driven Work Instructions | PR Newswire

SparkCognition Industrial AI

SparkCognition suggests improved methods to maximize turbine health

Wind Power Engineering AutoGrid

Distributed Energy Resource Management

Schneider Electric to acquire AutoGrid | PV Magazine USA

FERC: What We Do | Federal Energy Regulatory Commission

About NERC North American Electric Reliability Corporation

The AI Power Update Feb 23 - March 1, 2026

Will McKnight — Mon, 02 Mar 2026 14:57:44 GMT

Trump Announces “Ratepayer Protection Pledge” at State of the Union

At Tuesday’s State of the Union, President Trump unveiled the “Ratepayer Protection Pledge,” a commitment by major technology companies to fund their own electricity needs rather than passing costs onto utility customers. Energy Secretary Chris Wright confirmed that “all the brand-name hyperscalers” have signed on, covering companies including Google, Microsoft, and Amazon. Anthropic separately committed to covering 100% of electricity cost increases its data centers cause for residential customers. Trump told the audience no electricity prices would rise from AI data centers and that “in many cases, energy prices will go down.”

The pledge has no enforcement mechanism. Energy experts quickly called it “meaningless” without binding legal force. What it does signal clearly: the White House understands that AI’s electricity costs are becoming a political liability, and it wants Big Tech on record as the responsible party.

Watch whether Trump follows with an executive order that gives these commitments actual legal weight, or whether legislators at every level decide voluntary pledges are not enough. (Axios)

Amazon Commits $12 Billion to Louisiana Data Centers

Amazon announced a $12 billion investment to build data center campuses across Caddo and Bossier Parishes in northwest Louisiana, creating 540 direct jobs and roughly 1,700 additional positions in the region. The company pledged to cover 100% of costs for new energy infrastructure and grid upgrades needed to serve the facilities, working with utility Southwestern Electric Power Company (SWEPCO), with no costs passed to ratepayers. Amazon also committed $400 million to local water infrastructure to avoid straining municipal supplies. (GeekWire)

Google and Xcel Energy Announce 1.9 GW Clean Energy Deal with World’s Largest Battery

Google and Xcel Energy announced a 1.9 gigawatt clean energy package for a new data center in Pine Island, Minnesota, headlined by a 300 MW, 30 gigawatt-hour iron-air battery from Form Energy. Iron-air batteries store electricity by oxidizing iron and can hold power for up to 100 hours, covering multi-day weather events at a fraction of lithium-ion’s manufacturing cost. The full package adds 1,400 MW of wind and 200 MW of solar, and Google covers the total cost under Minnesota’s large-load customer framework. This is the arrangement regulators and advocates are pushing for everywhere: a tech company funding a massive clean energy buildout without shifting costs to residential customers. (TechCrunch)

Hawley and Blumenthal Introduce Bipartisan GRID Act in Senate

Senators Josh Hawley (R-MO) and Richard Blumenthal (D-CT) introduced the Guaranteeing Rate Insulation from Data Centers Act, the first federal bipartisan bill requiring data centers to fund their own electricity rather than raise residential utility bills. New data centers over 20 megawatts, enough to power roughly 15,000 homes, would be required to source all power outside the public grid, with existing facilities getting a 10-year transition window. The bill also mandates public disclosure of current and projected electricity demand. A populist Republican and a liberal Democrat landed on the same basic answer. (NBC News)

NRC Accepts Holtec SMR Construction Permit Application

The Nuclear Regulatory Commission accepted Holtec International’s application to build two SMR-300 reactors at the Palisades Energy Center in Michigan, each generating approximately 340 MW. This is the first major licensing application for a dual-unit small modular reactor (SMR) plant under the Part 50 framework, with an 18-month NRC technical review ahead. Holtec is simultaneously working to restart the existing 800 MW Palisades reactor, decommissioned in 2022, which would make it the first U.S. nuclear plant brought back from decommissioning status. (Federal Register)

Offshore Wind Projects Resume After Court Victories

Five major offshore wind projects resumed construction after federal courts issued preliminary injunctions blocking the Trump administration’s stop-work orders, including Vineyard Wind 1, Sunrise Wind, Empire Wind, and Coastal Virginia Offshore Wind. Dominion Energy reported $228 million in costs from delays and idle equipment during the BOEM suspension, plus $580 million in losses tied to Trump tariffs on steel and materials for the 2.6 GW Coastal Virginia project. Vineyard Wind 1 is now 95% complete with 60 of 62 turbines installed, targeting commercial operations by December 2026. The Trump administration has filed appeals, and broader legal challenges to the executive order suspending wind energy approvals remain pending. (Clean Technica)

xAI Expands Colossus Cluster to 2 GW

Elon Musk’s xAI confirmed it purchased a third building to expand its Colossus supercomputer complex to approximately 2 GW of compute power near Southaven, Mississippi, representing more than $20 billion in investment in the state. The expansion brings xAI’s total GPU count to approximately 555,000 NVIDIA chips across three facilities, making it the largest AI training cluster in the world by a significant margin. Colossus runs primarily on on-site natural gas turbines, and environmental groups have filed lawsuits claiming xAI operated 35 gas turbines at the original Memphis site despite permits for only 15. (Data Center Dynamics)

FERC Orders PJM to Create New Rules for Data Center Co-Location

FERC ordered PJM Interconnection, the nation’s largest grid operator, to establish new transmission service options for data centers co-located with generating facilities. FERC found that PJM’s existing behind-the-meter rules lack clarity on rates and conditions for these arrangements, and the new services will let data centers limit grid withdrawals while paying only for infrastructure they actually use. This ruling signals what FERC may require across other regions, with operators in MISO, SPP, and other RTOs expected to file similar proposals in 2026. (FERC)

Good week to follow the money and the meter. Make it a great March.

Will

Sources

How is AI advancing nuclear R&D today?

Will McKnight — Thu, 22 Jan 2026 19:47:46 GMT

If you follow nuclear energy through headlines, it feels like we’re living in the future already. SMRs seem like they’re right around the corner, fusion is always announcing another breakthrough, and big tech keeps signing new nuclear power deals for their data centers.

But if you step back and look at the grid, and the reality is that the U.S. has barely added net-new nuclear capacity in decades.

This week I sat down with Kevin Kong of Everstar to understand how companies like his are using AI to quietly change the slow, painful parts of nuclear development to help SMRs, advanced fission, and eventually fusion get deployed to our grid faster.

What are the main ways AI can help?

The biggest misconception about AI in nuclear is that it’s here to “design reactors.”

That’s not where the leverage is.

Nuclear already has solid physics, proven designs, and decades of operational experience. What it doesn’t have is speed. And most of that slowness has nothing to do with engineering itself.

Permitting is the obvious one. Nuclear projects live and die by regulatory approvals, and the work required to get there is enormous. Teams spend months digging through prior filings, regulatory guidance, environmental reviews, and historical precedent just to justify why a specific design or site should be allowed. AI doesn’t replace regulators or human judgment, but it can dramatically reduce the time spent searching, synthesizing, and drafting the first pass of compliant documentation.

“Nothing happens without approvals. You can have sites ready, designs ready, money ready, but it all comes down to getting green lights.” -Kevin Kong

Simulation and engineering workflows are another major area. Nuclear engineers already rely on physics-based models, but running those models and iterating on designs is slow and expensive. AI is increasingly being used as a way to accelerate iteration, explore more design scenarios, and reduce the overhead around setting up and validating each run.

Siting is another challenge where AI is useful. Choosing where to put a reactor requires decades of weather data, seismic risk, soil behavior, water access, environmental protections, local and tribal regulations, and emergency planning. That’s a massive research and synthesis problem. AI is well suited to pull together fragmented information and surface what actually matters for a given site.

That brings us to how AI can also help the industry deal with legacy information. Nuclear operators are constantly bouncing between outdated documents, poorly searchable government portals, and internal systems that were never designed to talk to each other. AI helps translate that mess into something usable, searchable, and consistent, which sounds boring until you realize how much time it saves.

“People are jumping between documents from a file cabinet in room C from the 1970s, government portals with terrible search, Google just to find the right document, and then back into Word redrafting things that have already been rewritten dozens of times.” - Kevin Kong

In short, AI is helping nuclear today by removing friction from everything around the reactor, while simulation and physics-based AI are also helping engineers get closer to nuclear fusion. Firms like NVIDIA working alongside university research labs on the problem (NVIDIA).

How is it being used today?

One of the clearest early use cases is in permitting and licensing, the part of nuclear development that most people outside the industry never think about but that eats years off every project. Nuclear labs and tech giants are already experimenting with this. In late 2025, Idaho National Laboratory announced a collaboration with Microsoft to lean on Azure cloud and AI tools specifically to help streamline the licensing process for new nuclear technologies (INL), speeding up regulatory research and documentation that traditionally takes huge amounts of time.

Permitting isn’t the only place AI is being embedded. On the operational side, utilities and vendors are beginning to use AI-driven scheduling and workflow tools to tighten up the mess of tasks that go into plant upgrades and construction plans. A recent report on collaborations between Aalo Atomics and Microsoft highlighted how generative AI and “AI agents” are being used to organize and optimize workflows around permitting and project planning (Nuclear Newswire) for a demonstration reactor, cutting cost and time by identifying the most impactful tasks to tackle first.

AI is also creeping into simulation and engineering, although it’s less flashy than the fusion headlines you see but arguably just as meaningful. Researchers and labs are feeding advanced algorithms with physics-based models to reduce iteration time and explore design spaces more quickly. These hybrid approaches, where AI augments deterministic models, are part of broader industry efforts to advance reactor design and safety analysis that organizations like the International Atomic Energy Agency have been talking about for years (NUCNET).

Even outside direct plant design or permitting, AI is being adopted in adjacent areas that matter down the road. Legacy reactors and academic facilities are starting to experiment with AI for remote monitoring, cybersecurity, and predictive maintenance, which future commercial designs will depend on for safety and reliability. These efforts don’t make big headlines yet, but they are essential groundwork for any grid-scale deployment.

Where are the gaps?

Despite the progress, there are still major problems AI hasn’t solved yet.

Construction and manufacturing remain huge challenges. The cost gap between building reactors in the U.S. versus countries like China is massive.

“Gigawatt-class reactors in the U.S. cost around ten billion dollars. In China, it’s closer to two and a half.” - Kevin Kong

Regulatory acceptance is another open question. AI can assist with analysis and drafting, but regulators still need to trust the outputs, understand how conclusions were reached, and feel confident that safety isn’t being compromised.

Fusion has its own suite of unsolved problems. AI can help with materials discovery and simulation, but the industry is still far from physics break-even, let alone commercial viability. Many fusion announcements are more about fundraising than deployment, and AI doesn’t change the underlying physics constraints. It can accelerate progress, but it can’t skip steps.

“There’s a big difference between physics break even and commercial break even, and we’re nowhere near commercial today.” - Kevin Kong

There’s also the human factor. Nuclear is a deeply conservative industry for good reason. Introducing new tools, even helpful ones, requires training, cultural buy-in, and clear accountability. AI that isn’t trusted or understood won’t get used, no matter how powerful it is.

Why does this matter?

The reason that the industry is so excited about nuclear is that it can potentially help our world reach a state of energy abundance.

“As energy costs come down, you start unlocking things that were previously too expensive like clean hydrogen, new transportation systems, more compute.” - Kevin Kong

The idea of energy abundance should be exciting to everyone and the main takeaway for me after this conversation was that AI will help make the nuclear industry move at a pace closer to the rest of the modern economy and thus help us reach abundance faster.