jain.com — data centers, connectivity and security, analyzed daily

Author: Deepak Jain

  • Savannah River Pick Shows DOE’s 16 AI Sites Are a Power Bet, Not Just a Land Deal

    Savannah River Pick Shows DOE’s 16 AI Sites Are a Power Bet, Not Just a Land Deal

    TL;DR · 30-second read

    The Short Version

    The United States government is opening some of its own land to companies that want to build the huge computer warehouses that run artificial intelligence. One of the first chosen sites is in South Carolina, at a lab that supports the nation’s nuclear weapons work.

    The twist: these projects are meant to come with their own power plants. The biggest artificial intelligence facilities can need as much electricity as a large power plant produces, and building power on-site is meant to keep that cost off ordinary households’ bills.

    Engineering firm Amentum was picked to build it. No price, size or start date has been announced.

    Microgrid Knowledge reported on July 20, 2026 that the National Nuclear Security Administration (NNSA), part of the U.S. Department of Energy (DOE), has selected Virginia-based engineering firm Amentum to develop a co-located artificial intelligence data center and on-site power generation project at the Savannah River National Laboratory site in South Carolina. The selection followed a competitive request for proposals that DOE began last year, after identifying 16 federal sites it judged could host AI data centers paired with their own energy infrastructure.

    The effort, formally the Savannah River Site Artificial Intelligence Infrastructure and Energy Generation Project, has no announced timeline, generating technology, capacity or cost. DOE says the project could meet its own electricity needs without shifting large costs to utility customers and could also increase the power available to the grid.

    Executive Summary

    NNSA has named Amentum to lead construction of both an AI data center and the power plant that would feed it, on federal land at one of the country’s principal nuclear security sites. It is a concrete step in DOE’s program to put AI computing on government property, and the first selection at Savannah River under that program.

    The significance lies less in the land than in the pairing. DOE did not simply offer acreage; it screened its sites for their ability to host data centers together with co-located generation, and it justified the Savannah River project on the grounds that it would not push large grid-upgrade costs onto utility customers. That makes federal sites a potential route around the slowest and most politically sensitive part of AI buildout: getting enough electricity to the building.

    What has been announced is a selection and a scope, not a financed, sized project. The generating technology, megawatt capacity, tenant, financing and schedule are all still undisclosed, so the model’s claims about cost and grid benefits remain untested.

    Sixteen Sites, One Condition: Bring the Power

    DOE’s list of 16 federal sites was not a catalogue of empty lots. According to the department, it identified sites that could support AI data centers paired with co-located energy infrastructure. The Savannah River award makes that pairing explicit: generation is in the project’s official name, and Amentum is contracted to build the power supply alongside the data center rather than to plug into whatever the local utility can provide.

    The reason is the grid connection itself. A large new electricity user normally joins the grid through an interconnection process, in which the utility studies whether its lines and power plants can serve the new load and what must be built if they cannot. At AI scale, those upgrades are substantial: industry projections cited in the coverage put future AI server racks at up to 1 megawatt each and the largest AI campuses at around 1 gigawatt, roughly the output of a large power plant. When a utility builds for a customer that size, the question of who pays tends to fall on its other customers. DOE’s stated case for Savannah River answers that directly, saying the project could generate its own needs “without shifting huge costs to utility customers.”

    That is where federal land earns its place. A large, secured, government-controlled site offers space, a single landowner and an existing security perimeter for putting generation right next to the load, which many sites closer to population centers struggle to offer. The trade-off is engineering: AI training creates wide, fast swings in power demand, and a self-supplied site has less grid backing to absorb them. Whatever generation Amentum ultimately builds will need to be designed around that volatility, which makes the power system as central to the project as the data center it serves.

    Why a Nuclear Security Site, and Why Amentum

    NNSA Administrator Brandon Williams framed the award as a way to “move faster, apply innovative technologies, and make productive use of federal land while maintaining our commitment to mission delivery.” That last clause matters. Savannah River hosts sensitive defense work, and any private-sector partner operating there has to coexist with it. Amentum already does; it was selected earlier this year to provide design and engineering services for the site’s Tritium Finishing Facility, which supports the nuclear weapons program.

    That existing footprint is a competitive advantage that is hard to replicate. Contractors that already understand federal security, procurement and site operations are positioned to win the first wave of these awards, while conventional data center developers may find their route onto federal land runs through partnerships with such firms. Amentum and NNSA say they will collaborate on expanding private-sector partnerships in the project, which leaves room for tenants, financiers or generation owners still to be named.

    What Is Settled, and What Still Rests on “Could”

    The announcement establishes a selection made through a competitive process, a defined scope covering both compute and generation, and a named site. Those are real commitments of intent from a federal agency. The headline benefits are conditional, however: the project could meet its own electricity needs, and it could increase power available to the grid. Williams himself called it a “proposed partnership.”

    Until the parties disclose the generation type, capacity and commercial structure, the ratepayer-protection claim cannot be measured. Its value depends on details not yet public: whether the site stays connected to the grid for backup, who pays for that connection, and on what terms any surplus power is sold. The first projects under DOE’s program will set the template for the remaining sites, so those terms will matter well beyond South Carolina.

    Background

    The Savannah River Site in South Carolina is a Department of Energy complex overseen by the National Nuclear Security Administration, the DOE agency responsible for nuclear weapons and nuclear security. Its national laboratory supports defense programs including tritium production. Amentum, a Virginia-based firm spun out of AECOM about six years ago, ranks as the second-largest U.S. government services contractor behind Leidos and was already working at the site on the Tritium Finishing Facility.

    The award sits within a wider federal push to use government land for AI infrastructure. DOE identified 16 of its sites as candidates for data centers paired with their own power, and Energy Secretary Chris Wright has described the AI race as “the next Manhattan project.” The pressure behind that push is electricity: an analysis by the Penn State Institute of Energy and the Environment projects AI’s share of global electricity use could rise from 4% to 20% by 2035.

    Sources

    Source: Amentum to Develop Co-Located AI Data Center and Power Infrastructure at Savannah River National Laboratory (Microgrid Knowledge, July 20, 2026): NNSA selects Amentum to build an AI data center with on-site power generation at the Savannah River Site.

  • 55 Oklahoma Data Centers, Zero Water Permits: AI Cooling Runs on Utility Contracts

    55 Oklahoma Data Centers, Zero Water Permits: AI Cooling Runs on Utility Contracts

    TL;DR · 30-second read

    The Short Version

    Oklahoma has about 55 data centers, the giant warehouses of computers behind apps, websites and artificial intelligence. Not one has applied for a state permit to draw water.

    They buy water from local utilities like any other customer, so the state does not track how much they use. One Google site uses as much water each year as 10,000 homes, and most of it evaporates.

    With a $15 billion campus proposed near Yukon in a state hit by years of drought, neighbors are asking questions developers are not yet required to answer.

    None of the 55 data centers that advisory firm Baxtel lists in Oklahoma has applied for a state water permit, according to reporting by Investigate Midwest published in The Oklahoman. They buy water as customers of municipal utilities, rural water districts or industrial parks that already hold water rights, so their consumption sits largely outside state oversight. Senate Bill 259, signed by Gov. Kevin Stitt in May, allows groundwater permits for data centers only if they use water-conserving, recycling cooling systems, but it does not reach facilities that buy water instead of seeking permits.

    The gap is drawing attention as Atlanta-based Beltline Energy proposes a $15 billion, 1,000-megawatt campus of 15 buildings on 480 acres straddling Oklahoma City and Yukon. Yukon is considering selling the project up to 2.5 million gallons of wastewater per day, while Beltline has requested extensions on the water agreement and has not said which cooling system it would use.

    Executive Summary

    Oklahoma’s newest data center water rule targets the one route data centers do not use. State permits govern taking water directly from aquifers and rivers; data centers instead contract with local suppliers. “So far, we have not had any permits issued specifically to data centers,” said Julie Cunningham, director of the Oklahoma Water Resources Board.

    The practical consequence is that the decisions that determine how much water an AI campus consumes, including its volume, its source and whether it evaporates or recycles cooling water, are being made in municipal supply agreements rather than state permits. Yukon’s pending contract with Beltline is a live example.

    Closing the permitting gap would not by itself produce usage data. The state relies on self-reported figures, metering requirements have failed twice, and the water board reported fewer than 20 staff in 2025. Rep. Carl Newton, the author of SB 259, says a new bill next year will address data centers drawing groundwater through a secondary source.

    Why the Permit Rule Misses the Meter

    SB 259 attaches a sensible condition to groundwater permits: data centers can get one only if their cooling systems are designed to conserve and recycle water. But a permit is only required when a user withdraws water itself. A data center that buys from a city, a rural water district or an industrial park holding industrial water rights is treated as an ordinary customer of that system, and the state never sees an application. That is why a statewide count of zero permits sits alongside 55 listed facilities.

    The real regulatory instrument is therefore the supply contract. Google’s Pryor data center, the state’s largest, buys water from the state-owned MidAmerica Industrial Park, which draws and treats it from the Neosho River. Beltline’s Yukon purchase agreement hinges on a separate water agreement with the city. Whoever negotiates those contracts, usually a municipal utility or city council, effectively decides volume, source and whether the water is potable or treated wastewater. The state’s conservation standard never enters the room.

    This matters for everyone on the supply side. City water utilities are becoming the de facto gatekeepers for AI cooling, often with small staffs and without statewide guidance. Developers face a patchwork of local terms rather than one state rulebook. And residents who want answers, like Yukon-area neighbor Pencia Truhan, who said developers “don’t have to tell us what they’re going to do about water,” find no state filing to consult.

    The Cooling Choice Is Also a Power Choice

    Cooling design is where water and electricity trade off. Open-loop, or evaporative, systems like Pryor’s lose water to the atmosphere: Google reports the site consumes 1.1 billion gallons a year, about the annual use of 10,000 American households, and returns roughly 275 million gallons to the Neosho River. That leaves about 825 million gallons evaporating each year. Closed-loop systems recirculate water, and air-cooled systems use little water but significantly more electricity.

    For Beltline’s proposal, that trade-off lands on an unusually large base. A 1,000-megawatt campus would, at full capacity, draw roughly as much electricity as Oklahoma City’s average use and more than double the demand of Google’s Pryor facility. Choosing air cooling to spare local water would push that electricity figure higher; choosing evaporative cooling would do the reverse. Because the cooling system has not been disclosed, neither the water nor the full power burden can yet be estimated.

    Yukon’s Math: Wastewater, Wells and a 2035 Deadline

    Yukon’s planning memo describes selling Beltline up to 2.5 million gallons of wastewater a day, and officials have promised the project will not rely on drinking water. At full volume that is about 900 million gallons a year, comparable in scale to what Pryor evaporates annually. Using treated wastewater is a common way to spare potable supply, but the sale still commits a resource the city may want for other purposes.

    The city’s supply position is already tight. Yukon relies on 15 leased wells in the Garber-Wellington Aquifer plus purchases from Oklahoma City, and those leases expire in 2035. A city-commissioned study projects its 27,300 residents will grow by about 23% over 25 years, in Canadian County, the state’s fastest-growing county. Against that, Beltline-commissioned studies project 450 permanent jobs and $543.12 million in tax revenue over 25 years, figures that potential tax abatements would significantly reduce.

    Measurement Is the Harder Fix

    Even a closed loophole would leave a data gap. The Oklahoma Water Resources Board relies on self-reported water use. Newton’s original bill would have required commercial groundwater users to install meters; that provision was removed after opposition from Panhandle irrigation groups, and Stitt vetoed similar legislation in 2023 after the Oklahoma Farm Bureau and others argued it infringed on private property rights.

    Capacity is the other constraint. In 2025 the board reported fewer than 20 staff handling an average of 250 permit applications and amendments a year, plus investigations and fieldwork, and its 140-page Water Demand Forecast from last October does not mention data centers. Cunningham said the agency is working to build expertise. Until measured data exists, planners in a drought-stressed state are forecasting water demand without one of its fastest-growing users in the model.

    Background

    Data centers house the servers behind everyday online services, and the newest hyperscale facilities built for artificial intelligence concentrate enough computing to generate heat on an industrial scale. Removing that heat often uses large volumes of water, which is why cooling design has become a siting issue alongside land and power. Oklahoma already hosts one of Google’s largest facilities in Pryor, where the company outlined plans addressing power and water for new sites, and the state is drawing further proposals.

    Beltline Energy, based in Atlanta, began as a renewable energy developer that secured land and developed solar projects before selling them to investors, and it is applying a similar model to data centers. It has developed more than a dozen solar farms but lists no completed data center projects. Its Oklahoma proposals, in Yukon and Oklahoma City, Piedmont and Luther, arrive as the state’s water regulator, which relies on self-reported usage data, works to develop expertise in the sector.

    Sources

    Source: Loophole leaves data center water consumption in Oklahoma largely untracked (The Oklahoman), on how Oklahoma data centers buy water from utilities outside state permitting, and the proposed Beltline Energy campus near Yukon.

  • 70% Oppose AI Data Centers Nearby. Their Power Lines Hinge on In-State Benefit

    70% Oppose AI Data Centers Nearby. Their Power Lines Hinge on In-State Benefit

    TL;DR · 30-second read

    The Short Version

    Power companies building electricity lines to feed new data centers, the warehouses of computers that run tools like ChatGPT, are using the government’s power to force landowners to sell.

    That power is only supposed to be used when the public benefits. Courts have usually allowed it when a line makes local electricity more dependable. One state’s top court blocked a line that crossed the state without serving anyone who lived there.

    With 7 in 10 Americans against data centers near them, expect more fights over farms and backyards.

    Power companies in states including Georgia and Pennsylvania have turned to eminent domain, the government power to take private property without the owner’s consent, to secure land for transmission lines serving the data center buildout. The practice is drawing opposition as polling shows 7 in 10 Americans oppose AI data centers in their own communities, University of Dayton law professor Aaron Walayat wrote in an essay published by The Conversation and republished by Fortune on July 19, 2026.

    The United States has more than 3,000 data centers, with another 1,500 in development, according to a Pew Research Center analysis. Data centers accounted for more than 4% of national electricity use in 2024. The legal question now in play is whether a line built to serve a private data center meets the Constitution’s “public use” requirement for taking land.

    Executive Summary

    Local opposition to data centers has mostly focused on the buildings themselves: utility bills, noise, pollution, water use and lost green space. It is now spreading to the transmission lines that carry power to those buildings. Transmission lines often have to cross private land, and when owners refuse to sell an easement, which is a right to use a strip of their property, utilities can ask a court to force the sale through condemnation.

    This matters because transmission is already one of the slowest parts of bringing new data center capacity online. A line that faces a credible legal challenge on public-use grounds adds legal risk to a project’s schedule. Precedent points to a consistent dividing line. Courts have upheld takings when a line delivers power or improves grid reliability for customers inside the state. They have been more skeptical when a line serves nobody in the state it crosses.

    For developers, utilities and regulators, how a line is justified in the record, and whom it demonstrably serves, is becoming as consequential as where it is routed.

    The Siting Fight Moves From the Building to the Wire

    A data center can be sited on a single parcel bought from a willing seller. The power it needs usually cannot. High-voltage transmission lines run for miles across many owners’ land, and every parcel on the route has to be secured. With more than 3,000 US data centers operating and another 1,500 in development, and with data centers already using more than 4% of the country’s electricity in 2024, utilities are planning new lines at a pace that makes holdout landowners a practical problem, not a rare one.

    The standard sequence is negotiation first and condemnation second. A utility offers to buy an easement. If the landowner declines, the utility, exercising power delegated by the state, asks a court to compel the sale at a court-determined “just compensation.” That is how power companies in Georgia and Pennsylvania have approached land acquisition for new lines. The 7-in-10 opposition figure measures attitudes toward data centers rather than transmission specifically, but the objections voters cite, higher bills and loss of green space, apply directly to a line crossing someone’s farm.

    Why In-State Benefit Is the Test That Decides These Cases

    The federal “public use” standard is permissive, so the contest shifts to state courts and state law, where a clear pattern emerges. The supreme courts of South Dakota and Vermont upheld utility takings because the lines provided at least some energy and improved grid reliability for in-state customers. In 1984, the Mississippi Supreme Court rejected a condemnation for a line that would have run from Mississippi into Louisiana without serving any Mississippi customers. In Texas, the state Supreme Court has held that a common carrier’s project must “serve the public” and “cannot be built only for the builder’s exclusive use.”

    Put together, those rulings describe the mechanism. A line documented as reinforcing reliability for a state’s own ratepayers fits well-established precedent, and the growing strain data centers place on the grid strengthens that case. A line whose practical function is to feed one private campus, or to carry power across a state border to load elsewhere, gives landowners a specific legal argument. Walayat’s reading is that reliability-driven takings will likely qualify as public use, while disputes over whether a line actually serves in-state customers remain the opening for challenges.

    The people most affected are the ones who write the justification. Utilities and their regulators decide how a line’s purpose is described in planning and certificate proceedings. Developers decide whether to take power from a dedicated feeder or from broader network upgrades. Landowners’ lawyers will look for evidence in that record that the public benefit is incidental. Projects that can show real benefit to in-state customers stand on firmer ground. Projects that cannot are carrying legal risk that belongs in their schedules.

    Kelo’s Long Shadow Makes State Lines Matter

    The backdrop is Kelo v. City of New London, the 2005 decision in which the US Supreme Court allowed a city to take homes for private redevelopment near a Pfizer facility on economic-development grounds. The redevelopment never happened, and Pfizer later left the city. The backlash was lasting: 45 states enacted eminent domain reform laws, and the supreme courts of Michigan, Ohio and Oklahoma have barred taking private property to hand it to another private party purely for economic development.

    For data center infrastructure, that history means a landowner’s odds depend heavily on the state. Challenges under a restrictive state constitution may succeed where a federal claim would not. Courts still routinely permit utilities to condemn land, so this is not a veto over transmission. It does mean that justifying a line as good for local economic growth, the argument Kelo accepted, is a weaker footing in many states than justifying it as necessary for reliable service to local customers.

    Weighing the Claims on Both Sides

    Each side’s argument deserves scrutiny. The industry’s position, that new lines are needed to keep the grid reliable as demand rises, is supported by the demand figures. It is also the same argument courts reward, which gives utilities an incentive to frame data center lines in reliability terms whether or not that is their main purpose. The opposition’s position, that private companies are borrowing a public power for private gain, echoes Kelo, where the promised public benefit never arrived. It has to be tested line by line, though, because many transmission projects do serve broad customer bases.

    The 70% figure is a strong signal of political headwind. It is not a legal standard. Public use is decided on the facts of each project, not by polling. The practical question for each contested line is the one courts have already asked: who in this state actually gets power from it?

    Background

    Eminent domain rests on the Fifth Amendment, which allows the government to take private property for public use if it pays just compensation. Most condemnations are carried out by state and local governments, which can delegate the power to utilities such as power and water companies. The US Supreme Court’s 2005 Kelo v. City of New London ruling read “public use” broadly to include economic development. The resulting backlash led 45 states to pass reform laws, and several state supreme courts, including those of Michigan, Ohio and Oklahoma, interpret their own constitutions more strictly.

    The AI boom has brought a matching surge in data center construction, to more than 3,000 facilities in operation and about 1,500 in development. Those facilities used more than 4% of US electricity in 2024, which pushes utilities to build new transmission. Those lines increasingly cross land whose owners do not want them.

    Sources

    Source: Power companies are using eminent domain to seize land for data centers as 70% of Americans say not in my backyard (Fortune, republished from The Conversation). An essay by University of Dayton law professor Aaron Walayat on the legal limits of taking land for data center transmission lines.

  • GPU Liquid Cooling Cuts Fans, Not Heat: Ventiva’s 1-2 CFM Bet on the Back of the Rack

    GPU Liquid Cooling Cuts Fans, Not Heat: Ventiva’s 1-2 CFM Bet on the Back of the Rack

    TL;DR · 30-second read

    The Short Version

    Data centers are full of computers that run hot, and fans have long blown that heat away. Fans alone use about 15% of a data center’s electricity, according to one company’s chief executive.

    Newer servers pipe liquid directly over their hottest chips, so many fans can be removed. The catch: smaller parts nearby, like network and memory cards, can then overheat and slow the whole machine down.

    Ventiva makes a tiny, silent air mover with no moving parts that uses electrically charged air to cool those parts. It is still being tested, and laptops get it before data centers do.

    Ventiva, a cooling technology company founded about 15 years ago as Thorn Micro Technologies, is qualifying a fanless “ionic air mover” for use inside data center servers. CEO Carl Schlachte told Facilities Dive that the device is aimed at a specific problem created by direct-to-plate liquid cooling. When operators remove bulk fans after moving processors onto liquid-cooled plates, lower-value components such as network interface cards (NICs), boot-optimized storage solution (BOSS) cards and memory modules (DIMMs) can overheat and throttle server performance.

    According to Schlachte, fans consume about 15% of the energy flowing into data centers, and the component needs only one or two cubic feet per minute (CFM) of airflow to cool those parts. Ventiva is qualifying the device with server manufacturers for the rest of 2026. First shipments go into laptops and mini PCs next year, and data center deployments follow.

    Executive Summary

    The announcement is less about a new way to cool a data center and more about a gap in an existing one. Direct-to-plate liquid cooling runs coolant through metal plates mounted on the hottest chips, the CPUs and GPUs, and it lets operators strip out many of the bulk fans that used to push air through a server. Removing those fans saves energy. It also removes the incidental airflow that kept everything else in the chassis cool. Ventiva is selling a product for that residual heat.

    This matters because liquid cooling adoption is being driven by dense GPU servers, and every liquid-cooled server still contains air-cooled parts. If those parts throttle, operators lose some of the performance they paid for. Ventiva’s pitch is that a solid-state air mover built into the card can recover it without bringing the fans back.

    The 15% figure attached to the story needs careful reading. It describes fans’ share of data center energy as Schlachte characterised it. It is not a measured saving from Ventiva’s device, which targets component temperature and performance rather than facility power.

    Liquid Cooling Moves the Heat Problem Rather Than Ending It

    A cold plate cools only the chip it touches. In a conventional air-cooled server, bulk fans pull a large volume of air front to back, and that stream cools the processors as well as everything downstream of them: network cards, boot drives, memory modules and power components. When direct-to-plate cooling takes over the processors, operators can reduce or remove those fans. The processors are fine. The components that were riding along in the airflow lose their cooling.

    This is the mechanism behind the headline. Schlachte described servers that were “throttling due to overheating at the back”, meaning the chips automatically slowed down to protect themselves. He said the fix does not require much air: one or two CFM moving between DIMMs is enough for things to “start to get better almost immediately.” For comparison, a single conventional server fan is designed to move far more air than that. The operational point is that a liquid-cooled server needs very little airflow for its residual components, but it still needs some, and it needs that airflow in the right place. A device small enough to sit on the card itself could supply it.

    The people affected are server manufacturers designing liquid-cooled platforms and operators retrofitting air-cooled halls with direct-to-plate systems. In both cases, the residual air-cooled load is a design problem. It does not disappear when the coolant loop goes in.

    Reading the 15% Correctly

    Coverage of the launch attached a 15% energy efficiency figure to the device. As Schlachte described it, 15% is the approximate share of data center energy consumed by fans. That figure is the reason operators want to remove fans, and removing them is a benefit of liquid cooling. Ventiva’s product does not produce that saving. At most, it makes the saving easier to keep by solving the thermal side effect.

    The company’s own performance claims are softer and not yet quantified in public. Schlachte said upgraded servers would run “some double-digit percentage better” and suggested operators would get “15% more uptime.” Neither figure came with a test configuration, workload or baseline. Uptime is normally measured as the share of time a system is available, so a 15% improvement implies a starting point that few production data centers would accept. The claim may mean fewer throttling events or fewer thermal faults, but it has not been defined. These are plausible directions of benefit, but they are not substantiated results, and buyers should treat them as hypotheses until qualification data is published.

    An OEM Channel, Not a Facilities Purchase

    Ventiva’s route to data centers runs through server manufacturers. Schlachte described OEMs approaching facilities managers to propose an off-cycle upgrade: technicians would swap existing BOSS cards, NICs or DIMMs for versions with the air mover built in. He framed this as “no hit to you in terms of bottom line.” That framing leaves open who absorbs the cost of the new cards and the technician time.

    This channel has advantages. The unit of adoption is a card, not a cooling plant, so an operator can trial it in a handful of servers without facility engineering work. It also depends on OEMs qualifying and warrantying the parts, which is why Ventiva is spending the rest of 2026 on qualification and thermal simulation. The sequencing reflects that dependency: laptops and mini PCs first next year, and data centers after that.

    Old Physics, New Operating Conditions

    The underlying principle, electrohydrodynamic flow, is well established. A charged wire creates a small plasma field that ionises nitrogen and oxygen molecules in the air. The ions drift toward a negatively charged collector and drag neutral air along with them, producing a breeze with no blades or motor. Schlachte noted that people have observed the effect for about 300 years. Removing moving parts removes a common failure point, and small size allows placement directly on a component.

    The engineering question is whether that principle holds up for years inside a server, in a hall with its own dust, humidity and maintenance regimes. Ventiva has said it has cards running and simulations under way. It has not yet published the reliability, efficiency and service-life data that operators and OEMs use to decide whether a new component belongs in production racks.

    Background

    Data centers have traditionally been cooled with air: room-level systems chill the air, and fans inside each server pull it across the components. As processors, especially the GPUs used for AI workloads, have grown denser and hotter, operators have adopted direct-to-plate liquid cooling, which carries heat away from the hottest chips through coolant-filled plates. That shift lets operators reduce the number of bulk fans, which Ventiva’s CEO says account for roughly 15% of data center energy use.

    Ventiva, originally Thorn Micro Technologies, has spent about 15 years developing solid-state air movers based on electrohydrodynamic flow, the airflow produced when charged ions travel between a wire and a collector. The company is positioning the technology first in compact consumer devices and then in servers. There, it targets components that are not on the liquid cooling loop but still need a small, steady supply of air.

    Sources

    Source: Ventiva touts fanless cooler for 15% data center energy efficiency gains — Facilities Dive’s interview with Ventiva CEO Carl Schlachte on the company’s ionic air mover for liquid-cooled servers.

  • CISA’s Coming 72-Hour Rule Makes AI Data Center Incident Response a Contract Term

    CISA’s Coming 72-Hour Rule Makes AI Data Center Incident Response a Contract Term

    TL;DR · 30-second read

    The Short Version

    The US government’s cybersecurity agency plans to finish new rules in September 2026. They will require essential services, likely including hospitals, banks and power companies, to report serious hacks within three days and any ransom paid to hackers within one day.

    The catch: many of these organizations run their computers inside buildings owned by someone else, including the huge data centers being built for artificial intelligence. A customer can only meet a three-day deadline if the building’s operator tells it about a break-in fast. Expect the contracts between them to get much stricter.

    The Cybersecurity and Infrastructure Security Agency (CISA) expects to issue its final rule under the Cyber Incident Reporting for Critical Infrastructure Act of 2022 (CIRCIA) in September 2026, Hunton Andrews Kurth reported on July 17. The rule will require covered critical-infrastructure entities to report covered cyber incidents to CISA within 72 hours and ransom payments within 24 hours.

    The date comes from a preview of the 2026 Unified Agenda, the federal government’s schedule of planned regulatory actions. CISA missed the statutory October 2025 deadline, then set an internal target of May 2026. Town halls originally planned for March and April were postponed by a lapse in Department of Homeland Security funding and held June 15 through 18, drawing more than 1,200 stakeholders over four days.

    Executive Summary

    CIRCIA is the first cross-sector federal mandate requiring critical-infrastructure operators to report cyberattacks to a single agency. After a proposed rule in April 2024 and more than a year of slippage, a final rule is now expected in September 2026, with the two headline obligations unchanged: 72 hours to report a covered cyber incident and 24 hours to report a ransom payment.

    For the data center and cloud industry, the most immediate consequence may come less from whether operators are themselves covered and more from their customers’ obligations. Hospitals, banks, utilities and other covered entities increasingly run critical systems in third-party facilities and clouds, and the statute explicitly contemplates incidents that begin with a compromised hosting provider. A customer’s 72-hour clock therefore depends on how quickly and completely its provider tells it what happened, which turns incident notification into a negotiated contract term.

    The practical window is short. Organizations that may be in scope should assess coverage against the proposed criteria, align incident response plans with the reporting deadlines, and review provider and customer agreements before the final text lands.

    A Tenant’s 72-Hour Clock Runs Through Its Data Center

    Under CIRCIA, the reporting duty sits with the covered entity: the hospital system, bank, utility or other organization that meets the final rule’s criteria. The 72-hour window runs from the point that entity reasonably believes a covered incident has occurred. Much of that entity’s computing, however, no longer sits in buildings it owns. It runs in colocation facilities, where a tenant rents space, power and cooling inside someone else’s building, and in public and GPU clouds, where it rents the computers themselves. The 2022 statute anticipates this: among the incidents the rule is meant to capture, it lists unauthorized access facilitated through the compromise of a cloud service provider, managed service provider or other third-party data hosting provider.

    That places the provider’s detection and notification speed directly on the customer’s compliance path. A covered tenant cannot report what it has not been told, and it needs enough detail from its provider (what was accessed, when and how) to file a report that meets CISA’s content requirements. Breach-notification clauses are not new, but a binding federal deadline gives customers a concrete number to negotiate against: notification windows well inside 72 hours, named incident contacts, and obligations to share forensic detail. If a provider is itself swept in by the final size- or sector-based criteria, it will be running its own clock in parallel with its customers’.

    The AI buildout sharpens the problem. New AI capacity is often delivered in layers: a facility owner, a GPU cloud or managed-infrastructure provider operating inside that facility, and the enterprise or model developer consuming the compute. Each layer is a hand-off in the notification chain, and each hand-off consumes hours of the covered entity’s window. Operators whose customer lists include regulated industries should expect incident-notification terms to be negotiated with the same seriousness as power and uptime commitments.

    Three Targets, One Missed Deadline, No Final Text Yet

    The timeline explains why caution is warranted. The statute required a final rule by October 2025. CISA missed that date and set an internal target of May 2026, citing the volume of comments on the April 2024 proposal and an intent to streamline requirements. Town halls meant to inform the final text were then pushed from March and April into mid-June by a Homeland Security funding lapse. September 2026 is the third target date.

    CISA has signalled that it wants to streamline the rule but has not said how. Streamlining could mean narrower coverage, fewer required data fields, clearer thresholds for what counts as a reportable incident, or some combination. Until the text is published, the April 2024 proposal remains the best available guide, which is why the practical advice is to assess coverage against its proposed sector-based and size-based criteria rather than wait. The turnout of more than 1,200 stakeholders at the June sessions suggests how many organizations believe they may be affected.

    Unified Agenda dates are projections, not commitments, and this rule has slipped before. Publication of a final rule is also not the same as the start of reporting obligations; the effective date and any transition period will matter as much as the publication date. Operators should treat September as a planning anchor, not a certainty.

    The 24-Hour Ransom Report Tests Governance, Not Just Tooling

    The 24-hour window for reporting a ransom payment is the tighter of the two deadlines. Ransom decisions typically involve executives, legal counsel and cyber insurers, not the security team alone. A one-day reporting clock after payment means the documentation behind that decision (who approved it, how much was paid, to whom, and why) has to be assembled as the decision is made, not reconstructed afterward. Organizations without pre-assigned authority and a documented escalation path will find the deadline hard to meet.

    CIRCIA also lands on top of existing obligations. Public companies already face the SEC’s requirement to disclose material cybersecurity incidents within four business days of determining materiality, and several sector regulators have their own notification rules. The statute provides an exception where an entity reports substantially similar information to another federal agency under an agreement with CISA, but how widely that exception applies will depend on agreements not yet described. A data center operator serving customers across several sectors may find that its customers each face a different combination of deadlines, and that each will expect its provider’s notification terms to fit.

    The intended payoff is real. CISA’s stated purpose is to deploy help to victims quickly, analyze reports across sectors for patterns, and share warnings with other defenders. For infrastructure operators, better cross-sector visibility into attack trends is a genuine benefit; the cost is process discipline that has to be in place before the first incident, not after.

    Background

    The Cybersecurity and Infrastructure Security Agency is the component of the Department of Homeland Security responsible for coordinating the protection of US critical infrastructure against cyber and physical threats. Before CIRCIA, cyber incident reporting to the federal government was largely voluntary or governed by a patchwork of sector-specific rules. CIRCIA, signed in March 2022, directed CISA to create a single cross-sector reporting regime, and CISA published its proposed rule in April 2024, setting out proposed coverage criteria, the scope of reportable incidents, and required report contents.

    Data centers and cloud platforms sit underneath much of that critical infrastructure, hosting systems for healthcare, finance, energy and government customers. As AI demand drives new capacity, more of those workloads run in shared facilities and layered provider arrangements, which is why the reporting rules matter to operators whether or not they end up directly covered.

    Sources

    Source: CISA Plans to Finalize Cyber Incident Reporting Regulations in September 2026 (Hunton Andrews Kurth), on CISA’s expected timeline for the CIRCIA final rule and its 72-hour and 24-hour reporting requirements.

  • Fireworks’ $1.5B Round Shows AI Inference Is Now a Layer Above the GPU Clouds

    Fireworks’ $1.5B Round Shows AI Inference Is Now a Layer Above the GPU Clouds

    TL;DR · 30-second read

    The Short Version

    A young company called Fireworks just raised $1.5 billion from investors. It doesn’t invent its own artificial intelligence. It runs freely available artificial intelligence programs for other businesses, using computer chips spread across more than a dozen cloud companies.

    Its yearly sales pace reportedly jumped from $100 million to over $1 billion in 16 months. Shopify and Twilio took four to five years to make the same climb.

    Why it matters: the money in artificial intelligence isn’t only going to chipmakers and tech giants. The middleman who keeps these systems fast and cheap is now big enough to fund on its own.

    Fireworks, the training and inference platform for open-weight AI models, has raised a $1.5 billion Series D, according to Bessemer Venture Partners, which said on July 17, 2026 that it had joined the round as an existing investor. Bessemer said Fireworks has passed $1 billion in annual recurring revenue (ARR), up from $100 million 16 months earlier. It also said the company processes about 43 trillion tokens a day. Tokens are the word-fragments AI models read and write.

    The company runs on more than a dozen cloud providers across over 20 regions. It is a first-party partner shipping natively inside Microsoft Foundry on Azure. It is led by CEO Lin Qiao, who led the Meta engineering organization that built the PyTorch framework, and by President George Hu, who joined in April 2026.

    Executive Summary

    Fireworks does not sell a chatbot or a proprietary model. It runs other people’s open-weight models, and the fine-tuned versions customers build on them, as fast and as cheaply as possible. It charges by the token. A $1.5 billion round for that business, at more than $1 billion in ARR, is a large bet that inference is a market of its own. Inference means running a trained model to answer requests, and on this bet it is neither a feature of the chip clouds nor of the model labs.

    For infrastructure operators, the important detail is the architecture. Fireworks describes its platform as a virtual GPU cloud. It pools graphics processors from more than a dozen providers and sits between those providers and the enterprises consuming AI. The more demand flows through layers like this, the more GPU owners compete as suppliers to software platforms rather than selling directly to end customers.

    The growth figures are striking. So far, though, they are headline numbers without the detail needed to judge durability. That detail includes margins, customer concentration, how much capacity is contracted, and how ARR is defined for a usage-based business.

    A Platform That Sits Above the Chips

    What Fireworks sells is not compute capacity in the usual sense. Its platform is described as a virtual GPU cloud. That is software pooling graphics processors (GPUs, the chips that do most AI work) from more than a dozen cloud providers in over 20 regions. On top sits a proprietary inference engine that decides how to run each model at maximum speed and efficiency. Customers buy output, measured in tokens, not servers. At roughly 43 trillion tokens a day and more than $1 billion in ARR, that software layer is now a business of real scale. The $1.5 billion Series D finances it as one.

    That is the operational shift. In the familiar cloud model, whoever owned the machines owned the customer. In this model, GPU owners become suppliers. Those owners include hyperscalers (the largest cloud providers, such as Azure) and neoclouds (newer GPU-specialist clouds). The platform holds the customer relationship, decides where each workload runs, and keeps the margin from running it efficiently. Bessemer states the ambition directly: Fireworks wants to be the software layer that abstracts away inference running on “heterogeneous silicon, hyperscalers, neoclouds, and air-gapped environments” — the inference platform “running on every GPU in the world.”

    For GPU cloud and data center operators, this cuts both ways. An aggregator pools demand from thousands of customers, which can make it a large and steady buyer of capacity. But demand routed through a multi-cloud layer is also portable: it can move toward whichever provider offers the best price, availability or location. The Microsoft relationship shows the layer can coexist with hyperscalers rather than simply competing with them. Azure distributes Fireworks natively through Foundry. Still, this is one company’s round. It shows the independent inference layer can be financed at scale, not yet that it will hold its margin once hyperscalers price their own inference services against it.

    Open Models Move the Spend From the Model to the Serving

    The investment thesis rests on a claim about models, not chips. Open-weight models, whose trained parameters are published for anyone to run, now match closed frontier models on the workloads most enterprises care about, at a fraction of the cost. On that premise, enterprises will increasingly post-train an open model on their own proprietary data instead of paying per call for a closed API. They would, in Bessemer’s phrase, stop wanting to “rent” their core intelligence.

    If that premise holds, value shifts toward whoever serves and tunes those open models well. Fireworks’ pitch is that it does both in one place. Its platform combines fine-tuning and reinforcement learning (methods for adapting a model with examples and feedback) with production serving. Every interaction can feed the next version of a customer’s model. That combination also creates switching costs. A customer whose tuned model, data pipelines and serving setup all live on one platform has reasons to stay. That stickiness is what justifies valuing the layer as more than a GPU reseller.

    The load-bearing premise, however, is asserted rather than demonstrated. No benchmarks or customer workloads accompany the claim that open models match frontier quality. The answer will vary by task. Buyers weighing an open-model strategy should test it on their own workloads rather than assume parity.

    What the Growth Numbers Establish, and What They Don’t

    Growth from $100 million to more than $1 billion in ARR in 16 months is, by Bessemer’s comparison, a climb that took Twilio and Shopify four to five years. It establishes that demand for hosted open-model inference is very large right now. For a usage-based business, however, ARR usually annualizes current consumption. It can fall if customers optimize their prompts, switch to smaller models, or move volume to another provider. That makes it a different quantity from contracted subscription revenue.

    Revenue is also not margin. A platform built on capacity drawn from other clouds has to pay for that capacity. Its economics depend on how many tokens its engine can produce per GPU-hour relative to what it pays for that hour. The inference engine is the core asset for that reason: efficiency is the margin. The broader market figures cited alongside the round are context, not evidence about Fireworks. They include hyperscaler capital spending of more than $800 billion in 2026 and token consumption expected to rise more than 30-fold by the end of the decade, and no source is given for either projection.

    Background

    Fireworks was founded by engineers who built and scaled PyTorch at Meta, the open-source framework on which much of today’s AI software is built. CEO Lin Qiao led that engineering organization. The company positions itself as a training and inference platform for open-weight models. It combines a proprietary inference engine, a multi-cloud GPU pool and tools for fine-tuning and reinforcement learning. Bessemer Venture Partners was already an investor before the Series D.

    The round lands as AI spending shifts from building models toward running them. Enterprises that first used closed frontier-model APIs are weighing open models they can tune on private data and serve more cheaply. A growing set of providers now competes to run those models: hyperscalers, GPU-specialist neoclouds and independent inference platforms.

    Sources

    Source: Fireworks raises $1.5B Series D for open model inference – Bessemer Venture Partners: Bessemer’s announcement of its participation in Fireworks’ Series D and its investment thesis.

  • Gold Eagle’s Four-Agency AI Bug Hunt Shows Why Patching Is the Grid’s Real Limit

    Gold Eagle’s Four-Agency AI Bug Hunt Shows Why Patching Is the Grid’s Real Limit

    TL;DR · 30-second read

    The Short Version

    The White House has launched Gold Eagle, a program that uses the most advanced artificial intelligence systems to hunt for hidden flaws in the software that runs power grids, banks and the buildings that house the internet.

    Four parts of the federal government share the lead, working with private companies. Finding flaws faster is the easier half. Fixing them is harder: a power grid cannot simply switch off to install an update. The real test is how quickly operators can make repairs.

    The White House on Tuesday, July 14, launched the Gold Eagle Initiative, a cybersecurity vulnerability coordination clearinghouse meant to speed the identification, prioritization and remediation of software flaws across U.S. critical infrastructure, Industrial Cyber reported. The initiative was established under Executive Order 14409, the June 2 order on advanced artificial intelligence innovation and security.

    Gold Eagle is led jointly by the White House, the Treasury Department, the Department of Homeland Security through the Cybersecurity and Infrastructure Security Agency (CISA), and the Department of War, working with open-source software partners and critical infrastructure companies. It will use frontier AI, meaning the most capable current AI models, for faster exploit detection and prioritized threat sharing, and it has already begun collecting and prioritizing reported vulnerabilities across sectors.

    Executive Summary

    Gold Eagle is a coordination model, not a new regulation. The administration says it will draw on existing federal authorities and resources to cut duplicative vulnerability scanning, verify reported flaws once, and push prioritized, actionable remediation information to defenders in government and the private sector. The pitch is efficiency: many organizations currently scan for, triage and verify the same weaknesses independently.

    The significance for infrastructure operators lies in what happens after a flaw is found. AI-assisted discovery can increase the volume and speed of credible vulnerability reports reaching utilities, data center operators and their equipment vendors. For information technology systems, patching is routine. For the operational technology that controls power, cooling and physical processes, every fix competes with uptime, safety testing and vendor certification.

    That makes remediation capacity, not detection, the measure that will determine whether Gold Eagle materially reduces risk in the grid and data center sectors. The announcement describes the front half of that pipeline in detail and the back half only in general terms.

    A Clearinghouse, Not a Mandate

    The structure the administration describes is a central intake point for vulnerabilities: reports are collected, de-duplicated, verified and ranked, then routed to the organizations that must act on them. Frontier AI is positioned as the accelerant for both discovery and triage. Nothing in the announcement creates new reporting obligations for private operators; it frames Gold Eagle as a way to use “existing federal authorities and resources” more effectively and to “amplify collaboration between government and industry.”

    That design choice matters. A voluntary, collaboration-based model can move quickly because it does not wait on rulemaking, and the announcement says work is already under way. The trade-off is that its reach depends on which companies choose to participate and how much they share. The officials’ statements lean heavily on rhetoric, with references to a “wartime footing” and “unprecedented coordination,” but the announcement provides no participant list, volume figures or performance targets against which those claims can be measured yet.

    Why Patching, Not Finding, Is the Constraint

    The mechanism Gold Eagle describes covers identification, prioritization and verification, and then delivers “prioritized and actionable threat and remediation information to defenders.” The actual fix still lands on two groups outside the clearinghouse: the vendors who must write and validate a patch, and the asset owners who must install it. If frontier AI does what the administration expects and surfaces more credible flaws faster, the four federal leads will be increasing the flow into a pipeline whose downstream end they do not control.

    In operational technology, that downstream end is narrow. The programmable controllers, protection relays and supervisory systems that run a substation, and the building management, power distribution and cooling controls that keep a data hall online, are designed to run continuously for years. Updating them typically requires firmware validated by the equipment maker, testing against safety functions, and a scheduled maintenance window in which redundancy is temporarily reduced. A utility or colocation operator cannot absorb an unlimited number of those windows, and some legacy devices cannot be patched at all, leaving compensating controls such as network segmentation as the only near-term option.

    This is where Gold Eagle’s prioritization promise has real value, and also where it will be tested. A ranked list that tells an operator which of many findings is most exploitable in their environment helps them spend scarce maintenance windows well. An unranked or poorly contextualized flood of findings would do the opposite. The people most affected are control-system engineers at utilities, facility and critical-environment teams at data centers, and the product-security staff at industrial equipment vendors, all of whom have finite capacity to act on what Gold Eagle produces.

    What Grid and Data Center Operators Should Prepare For

    The operators best positioned to benefit are those that already know what they run. Prioritized advisories are only actionable if an organization can quickly match a flaw to a specific firmware version on a specific device, which requires an accurate asset inventory and, increasingly, software bills of materials from vendors. Operators without that inventory will receive the same intelligence but struggle to use it.

    Procurement is the second lever. Contracts for switchgear, uninterruptible power systems, cooling plant and building management platforms can specify how quickly a vendor must deliver a validated patch after a credible vulnerability is reported. If Gold Eagle accelerates disclosure, those terms become more consequential. Smaller utilities and regional data center operators with thin security staff are the most exposed to a gap between what they are told and what they can fix, and they may lean more on managed security providers and vendor support as a result.

    Open Source and the Question of Trust

    The inclusion of open-source software partners is notable, because much critical infrastructure software depends on open-source components maintained by small teams. AI-scale discovery could send those maintainers more verified reports than they can resolve, so the value of deduplication and verification upstream is substantial if it works as described.

    Participation also raises questions that private operators will weigh carefully. The Department of War is a co-lead alongside civilian agencies, and Treasury’s statement focuses on the financial sector. Companies deciding how much to share will want clarity on how data is handled, who sees it, how long vendors get before details circulate more widely, and what legal protections apply. None of that undercuts the initiative’s aims, but the answers will shape how much of the private sector’s visibility Gold Eagle can actually draw on.

    Background

    Coordinated vulnerability disclosure is the long-standing practice of reporting a software flaw privately to the vendor, giving time for a fix, and then publishing details so users can protect themselves. In the United States, CISA, created in 2018 within the Department of Homeland Security, plays a central role: it supports the CVE system that assigns identifiers to known flaws and maintains the Known Exploited Vulnerabilities catalog, which flags weaknesses attackers are actively using.

    Critical infrastructure spans energy, finance, communications, water and other sectors, and increasingly includes the data centers that support cloud and AI services. Much of it runs on operational technology, the industrial control systems that manage physical equipment, which historically lags office IT in patching because uptime and safety come first. Gold Eagle is the administration’s attempt to apply AI to the discovery and triage end of that process at national scale.

    Sources

    Source: US launches Gold Eagle initiative to boost cybersecurity vulnerability coordination, bolster critical infrastructure defense (Industrial Cyber): the White House’s launch of an AI-driven vulnerability clearinghouse under Executive Order 14409.

  • Michigan’s 2M-Gallon Water Cap Would Push Data Centers to Closed-Loop Cooling

    Michigan’s 2M-Gallon Water Cap Would Push Data Centers to Closed-Loop Cooling

    TL;DR · 30-second read

    The Short Version

    Michigan lawmakers have proposed a set of rules for data centers, the giant warehouses of computers that run apps, cloud services and artificial intelligence.

    The biggest of these buildings can use up to 5 million gallons of water a day, about seven Olympic swimming pools. One bill would limit the water a site can permanently use up to about three pools’ worth a day. Another would require cooling systems that reuse the same water over and over.

    Developers would also need a signed deal with the local town before they could build or connect to the power grid.

    Michigan Democrats introduced a package of bills in early July 2026 to regulate data centers’ energy use, water use, noise, labor standards and use of nondisclosure agreements, MultiState reported. The centerpiece would require operators to sign legally binding community benefit agreements with local governments before they receive site plan approvals, permits or utility interconnection. An interconnection is the approval to connect a facility to the power grid.

    The package contains competing Senate and House versions on energy and water. Senate Bill 1047 would require 20-year utility contracts and 90% clean energy procurement. House Bill 6135 would create a new large-load rate class with 15-year contracts and no clean energy mandate. On water, Senate Bill 1046 would cap consumptive use at 2 million gallons per day. The House is considering a requirement for closed-loop cooling or municipal water only.

    Executive Summary

    Michigan’s package goes beyond one-off siting fights. It would write conditions on energy, water, labor and local benefits into state law. Most permitting disputes are settled project by project in township meetings. These bills would instead set a statewide floor that every large facility would have to clear. None of the bills has passed, and the Senate and House versions differ in important ways.

    The water provisions matter most for how facilities are designed. Large data centers can use up to 5 million gallons a day. The Senate would cap consumptive use at 2 million gallons, and the House would require closed-loop cooling or municipal water. The two chambers take different routes, but both would steer developers toward cooling that evaporates less water. The community benefit agreement requirement would change the order of development. A signed local deal would come before the grid connection, not after it.

    Michigan is not alone. Six other states have considered community benefit agreement mandates this year. Twelve have weighed closed-loop cooling requirements, and twelve have weighed decommissioning rules. What stands out in Michigan is how many of these ideas arrive together in a single package.

    Two Competing Water Bills, One Direction

    Senate Bill 1046 would create a new state permit for anyone using, or expecting to use, 550,000 gallons of water or more per day. Applicants would have to hold at least three public hearings before they apply. Permit holders would face a cap of 2 million gallons per day of consumptive use. Consumptive use means water that leaves the local supply and is not returned, mostly through evaporation in cooling towers. For comparison, large data centers can use up to 5 million gallons a day. At that scale, a design that depends on evaporating large volumes of water would struggle to fit under the cap.

    The House approach is blunter. New data centers would have to use closed-loop cooling, which recirculates the same water instead of drawing and discharging fresh supply. The only alternative would be to take water solely from a municipal system. The mechanisms differ: the Senate sets a volume limit, while the House mandates a technology or a water source. The engineering result is similar. Developers planning large campuses in Michigan would be pushed toward low-evaporation cooling, whichever version prevails.

    This follows a shift the industry has already started. Operators are actively adopting less water-intensive cooling. The effects spread beyond the building itself. Cooling equipment suppliers would gain a more predictable market. Municipal water utilities could become the default supplier under the House option. Closed-loop and dry cooling generally save water but use more electricity. That puts water policy in direct contact with the energy bills moving alongside it.

    The Local Deal Would Come Before the Grid Connection

    House Bill 6137 and Senate Bill 1050 would reorder the development process. No local site plan approval, no local permit and no Public Service Commission approval of an interconnection could be issued until the local government approves a community benefit plan. For developers, the negotiation with the host community would move onto the critical path, ahead of the utility process that usually sets project timelines.

    The agreements would have to cover at least half of five listed topics, which in practice means three. The topics are local hiring, water use, cost allocation for electrical or water upgrades, community investment funding, and home efficiency improvements. A separate pair of bills, House Bill 6140 and Senate Bill 1049, would bar elected officials and operators from signing nondisclosure agreements about construction when a project receives a tax incentive. Intellectual property could still be redacted. Together, these bills would make the terms of the local negotiation largely public for subsidized projects.

    Two Chambers, Two Ways to Price Power

    The energy bills share one principle: data centers should pay for their own power, so that other ratepayers do not absorb the cost. Senate Bill 1047 would set firm terms. Contracts would run at least 20 years, customers would be billed for at least 90% of their contracted demand whether they use it or not, and applications would cost $100,000. That billing floor is aimed at the risk of a utility building for a large load that never fully arrives. The Senate bill would also require operators to procure clean energy equal to 90% of their annual usage and to join a demand response program, which pays or credits large users for cutting consumption when the grid is strained.

    House Bill 6135 would create a large-load commercial rate class with contracts of at least 15 years. Its financial terms are less strict, and it has no clean energy requirement. On the House side, House Bill 6142 would require every Public Service Commission-approved data center contract to include a decommissioning plan. The plan would need financial assurance, meaning money set aside, sufficient to cover dismantling the facility, and it would be reviewed every three years. Developers comparing states would treat the choice between these two frameworks as a real cost difference, not a technicality.

    Labor Standards, and No Moratorium

    House Bill 6141 and Senate Bill 1048 would bar the Public Service Commission from approving data center contracts unless construction and maintenance work is done under a labor agreement or collective bargaining agreement. The International Brotherhood of Electrical Workers has urged members to oppose moratorium legislation. The union argues that data centers bring well-paid construction jobs and long-term union employment. That position helps explain the package’s shape: it regulates how data centers are built, not whether they can be built.

    The package contains no moratorium and no tax incentive changes. Those questions are being handled in separate Michigan bills. They include proposals to repeal incentives, to exclude data centers from certain property tax exemptions, and to extend a brownfield-zone exemption. The package’s premise is that data centers will keep coming to Michigan and the state should set the terms, which is a middle position between welcoming them unconditionally and blocking them.

    Background

    Data centers house the servers behind cloud computing, streaming and artificial intelligence. The largest require power on the scale of a small city, and some draw millions of gallons of water a day for cooling. As construction has spread into new regions, state legislatures have moved from offering incentives to setting conditions. States considered hundreds of data center bills last year. This year, six states other than Michigan have weighed community benefit agreement mandates, twelve have weighed closed-loop cooling requirements, and twelve have considered decommissioning rules.

    Michigan’s legislature has been active on the issue throughout 2026. Separate bills address data center tax incentives: some would repeal them, one would narrow property tax exemptions, and others would extend a brownfield-zone exemption. The July package is the state’s broadest attempt yet to set operating conditions for the industry. At the federal level, a proposal from Representative Robert Bresnahan would tie federal data center tax credits to community benefit agreements.

    Sources

    Source: Michigan Data Center Legislation Targets Energy Use and Community Benefits (MultiState), an overview of Michigan’s proposed data center bills on energy, water, community benefits, noise, nondisclosure agreements and labor standards.

  • Japan’s 27,500-GPU National AI Factory Is Really a 140-Megawatt Power Project

    Japan’s 27,500-GPU National AI Factory Is Really a 140-Megawatt Power Project

    TL;DR · 30-second read

    The Short Version

    Japan’s government is backing a giant computer center, built with chipmaker NVIDIA and a Japanese company called Noetra, to train artificial intelligence for robots, factories and hospitals.

    It will hold 27,500 of NVIDIA’s newest high-end chips and be able to draw 140 megawatts of electricity, roughly what a small city uses.

    The chips get the headlines, but lining up that much power and cooling is the harder job. The announcement does not yet say where it will be built, when, or who is paying.

    NVIDIA announced on July 16, 2026 that it is working with Noetra Corp. to build an NVIDIA Vera Rubin “AI factory” in Japan with 13,750 Vera CPUs and 27,500 Rubin GPUs, delivering 140 megawatts of data center capacity on NVIDIA’s DSX reference platform and Spectrum-X Ethernet networking. The facility will provide the computing foundation for the FRONTia Project, a program launched by Japan’s Ministry of Economy, Trade and Industry (METI) to develop multimodal foundation models for robotics and “physical AI.”

    NVIDIA calls it the world’s first national AI infrastructure for physical AI. Noetra will share the pretrained weights of its models with Japanese developers and enterprises, alongside NVIDIA’s open model families and software libraries.

    Executive Summary

    The announcement pairs a state industrial program with a single, very large training cluster. METI’s FRONTia Project sets the mission: build reliable multimodal foundation models (large AI models that learn from text, images, video and sensor data together) that can underpin robots, digital twins and industrial automation. Noetra will build and run the hardware, and NVIDIA supplies the full stack from chips and racks to networking and software.

    It matters for two reasons. First, it is an explicit bet that Japan’s manufacturing base, and the real-world industrial data it generates, can be turned into a competitive advantage in AI for machines that act in the physical world. Japan’s AI Robotics Strategy targets more than 30% of a global AI robotics market it estimates at $133 billion by 2040. Second, the size of the build means this is a utility-scale infrastructure project as much as a computing one: 140 megawatts is the scale of a large hyperscale campus, and the announcement leaves open where that power and the cooling to go with it will come from.

    The Unit That Matters Is the Megawatt

    The headline numbers are chips, but the figure that governs how this project gets built is 140 megawatts. Divide it by 27,500 GPUs and the facility is provisioned at roughly 5 kilowatts per GPU once cooling, networking, storage and power-conversion losses are counted. The release describes the build as Vera Rubin NVL72 racks, which pair 72 GPUs with 36 CPUs, the same 2-to-1 ratio as the 27,500 and 13,750 figures announced. That works out to roughly 380 racks sharing 140 megawatts, or about 365 kilowatts of facility capacity per rack. The release does not separate IT load from cooling and overhead, so the per-rack draw at the chip level will be lower. Even so, the order of magnitude is far beyond the 10 to 20 kilowatts a conventional enterprise rack uses.

    Density at that level has direct operational consequences. Air cannot remove that much heat, so racks of this class are liquid-cooled, which means pumps, heat exchangers and a heat-rejection plant sized for a small power station’s output. A 140-megawatt load also needs a high-voltage grid connection and on-site substation capacity, which in most markets are the longest-lead items in any data center schedule. NVIDIA itself frames the platform this way: its DSX reference design is pitched on raising “token throughput per megawatt,” a metric that treats power, not silicon, as the scarce input.

    For NVIDIA as a named partner, securing the GPUs is the part most within the project’s control. Site selection, grid interconnection, permitting and cooling construction are the parts that determine when the factory actually trains its first model. The parties most affected are the utility that serves the site, the electrical and mechanical contractors, and cooling-equipment suppliers. The release names none of them.

    What “National” Means in This Model

    The novel element is less the hardware than the arrangement around it. METI supplies the policy mandate through FRONTia, Noetra builds and operates the cluster, and the output is intended as a shared national asset. Pretrained model weights will be “broadly available” to domestic developers and enterprises. That is closer to how governments treat roads or research reactors than to a commercial GPU cloud, where capacity is rented by the hour to whoever pays.

    The “world’s first national AI infrastructure” claim deserves a careful reading. Japan already operates government-backed AI computing, such as the ABCI supercomputer run by the national research institute AIST, and METI has previously subsidized domestic cloud providers to expand GPU capacity. The release qualifies its claim in the body text as the first national AI infrastructure “for physical AI.” That narrower framing, a national-scale system dedicated to robotics and industrial foundation models, is more defensible than the headline version.

    Physical AI and the 2040 Target

    “Physical AI” refers to models that perceive and act in the real world: robots on a factory line, autonomous logistics systems, digital twins that simulate a plant before it is changed. Training them is data-hungry in a different way from chatbots, because they need video, sensor and motion data from real industrial settings. METI’s pitch is that Japan’s manufacturers hold exactly that data and that pooling it around a common model gives domestic firms a head start they could not buy individually. Noetra’s CEO, Hironobu Tamba, makes the same argument: these are “challenges no single company can solve alone.”

    The link from one cluster to a 30% global market share by 2040 is a long one, and the release does not attempt to draw it. The $133 billion figure is an estimate attached to a policy goal, not a forecast tied to this facility. The release says the factory will support trillion-parameter-scale models “as the AI factory expands,” which suggests a phased build. The claims of “lower token costs” and “breakthrough AI performance” are likewise unquantified.

    Who Gains From the Full-Stack Design

    The project is NVIDIA end to end: Vera CPUs, Rubin GPUs, BlueField data processing units, Spectrum-X Ethernet, and the Nemotron, Cosmos, Isaac GR00T and NeMo software that will ship alongside Noetra’s models. For Japanese developers, that means a ready toolchain and a common starting point, which lowers the cost of building robotics applications. It also means the national physical-AI ecosystem will be built on one vendor’s architecture from the silicon upward, a trade-off between speed and supplier diversity that sovereign AI programs worldwide are weighing. Choosing Ethernet rather than InfiniBand for the cluster fabric fits a wider industry move toward standards-based networking at scale, which may make it easier to find operators and engineers who can run it.

    Background

    Japan has pursued AI computing as industrial policy for several years, building government research systems and subsidizing domestic cloud providers to expand GPU capacity. METI’s AI Robotics Strategy, released in March 2026, sharpened that focus on robotics and physical AI, where Japan’s manufacturing base gives it a large pool of real-world industrial data. FRONTia is METI’s vehicle for turning that strategy into shared foundation models.

    NVIDIA has been promoting “sovereign AI,” national AI capacity built within a country’s borders, as a major market alongside its hyperscale cloud customers. Vera Rubin is its next-generation platform after Blackwell, pairing Vera CPUs with Rubin GPUs in dense, liquid-cooled NVL72 racks, and DSX is its reference design for building complete AI data centers around that hardware.

    Sources

    Source: Japan Government, Industrial Leaders and NVIDIA Launch the World’s First National AI Infrastructure (NVIDIA Newsroom, July 16, 2026). NVIDIA and Noetra announce a 140-megawatt Vera Rubin AI factory for METI’s FRONTia physical AI project.

  • Amazon’s 280 Gas Generators Show AI Backup Power Now Runs on Air-Permit Limits

    Amazon’s 280 Gas Generators Show AI Backup Power Now Runs on Air-Permit Limits

    TL;DR · 30-second read

    The Short Version

    Amazon is building a huge complex of computer buildings, the kind that run online services and artificial intelligence, on an old steel mill site outside Philadelphia.

    Pennsylvania plans to let Amazon add 280 natural gas engines that switch on if the power goes out. That is on top of 76 diesel engines already there.

    The catch is that Amazon agreed to a strict yearly pollution limit, so those engines can only run so much. For buildings like these, keeping the lights on is now decided partly by pollution permits and packed town meetings, not just the power company.

    WHYY reported on July 15, 2026 that Pennsylvania’s Department of Environmental Protection (DEP) plans to approve 280 natural gas-fired generators and three diesel-fired generators to supply backup power at Amazon’s data center complex under construction in Falls Township, Bucks County. Amazon has already installed 76 diesel generators at the 250-acre site, the former U.S. Steel mill now called the Keystone Trade Center. The generators would run when the site’s electricity supply from the utility PECO is interrupted.

    DEP officials presented the plan at a tense public meeting on July 14 at Pennsbury East High School. Under the pending application, Amazon agreed to keep annual emissions of nitrogen oxides and volatile organic compounds under 25 tons. The air quality plan approval opens for 30 days of public comment once it appears in the Pennsylvania Bulletin, which DEP expected on July 18.

    Executive Summary

    Amazon’s Falls Township campus, with up to 10 new buildings plus 1 million square feet of retrofitted warehouse space and more than half built, is set to carry 359 backup generators in total if DEP approves the new units. Amazon says the complex will host cloud computing infrastructure and “advance AI innovation.” It is part of a $20 billion investment Amazon has pledged in Pennsylvania.

    The figure that matters most is the 25-ton emissions ceiling, not the generator count. By fitting emissions controls on its larger generators and limiting fuel use, Amazon qualifies the site as a “synthetic minor” source. That means it accepts enforceable limits to stay below the threshold for heavier regulatory treatment. In effect, the campus’s backup power becomes an annual emissions budget written into an air permit.

    This moves the reliability question for large AI-oriented campuses out of the utility’s engineering studies and into state air permitting, public comment periods and community meetings. It also sits beside a separate, unresolved debate over who pays as data center demand grows on the regional grid.

    A 25-Ton Cap Turns Backup Power Into a Budget

    The headline number is 280 gas generators, but the number that governs how they can be used is 25. Under its pending application, Amazon agreed to keep combined annual emissions of nitrogen oxides and volatile organic compounds under 25 tons per year. Both pollutants can form ground-level ozone and worsen asthma. To stay under the limit, Amazon will install emissions control systems on its larger generators and limit fuel usage. Staying under 25 tons qualifies the campus as a “synthetic minor” source: a facility whose equipment could in principle emit more, but which accepts binding limits to remain below the level that triggers the more demanding permitting that applies to major sources.

    A fuel-use limit works in practice as a limit on running time. Data center backup generators typically run for routine testing and maintenance, and for real outages. The permit turns that into a fixed annual allowance shared across the whole fleet. In a year with a reliable grid, most of the allowance goes to testing. A long outage or repeated grid emergencies would draw it down faster. Reliability planning for this campus therefore has a regulatory ceiling written into the air permit, alongside the electrical design.

    That affects the people who plan, build and permit these sites. Developers sizing backup fleets for large campuses have to design against local emission thresholds as well as uptime targets. The approving authority for that design is a state air regulator operating in public: a packed auditorium meeting, a 30-day comment window, and a petition critics say has drawn more than 4,000 signatures. Pennsylvania’s Permit Fast Track Program sped up earlier permits for this project. The backup power decision still runs through public comment.

    Hundreds of Engines Behind a Grid-Connected Campus

    The campus is designed to run on grid power from PECO. The generators exist for the moments that supply fails. Even so, the fleet is large. Seventy-six diesel units are already installed, and 283 more are proposed, for 359 in total. That scale reflects what hyperscale operators expect from a site this size. A hyperscale campus is a very large, standardized facility built by a major cloud provider, and customers of cloud and AI services expect it to keep running through a utility outage.

    The new units are mostly gas-fired rather than diesel. Amazon has not publicly explained the choice. In general, gas units draw fuel from a pipeline rather than on-site tanks and produce less soot than diesel. They still emit nitrogen oxides, which is why the emissions controls and the tonnage cap matter. Amazon has not disclosed the combined generating capacity of the fleet, so it is not possible to say how much of the campus load the generators could carry, or for how long.

    The Grid Cost Question the Transmission Agreement Does Not Settle

    PECO’s Glen Murphy said the utility found the project will not raise the distribution rates PECO customers pay. He said the project’s second phase will require infrastructure upgrades, but a “take-or-pay” transmission security agreement with Amazon protects residents from those costs. Under a take-or-pay arrangement, the customer commits to pay whether or not it uses the capacity. Murphy called the project a likely net benefit to customers that could “bring down rates a bit.”

    Those agreements do not cover a different cost: the price of electricity supply itself, which rises with demand on the regional grid. PECO is part of PJM, the grid operator serving Pennsylvania, New Jersey and other states. New Jersey Assemblyman Balvir Singh argued that added demand will push up bidding and prices across that shared grid. The two claims are not contradictory. One concerns the wires that connect the campus; the other concerns the market for the power flowing over them. Neither side offered a figure for the supply-cost effect.

    Weighing the Claims on Both Sides

    Amazon’s Becky Ford said the company’s Pennsylvania data centers use water for cooling around 6% of the year on average, relying mostly on outside air. An engineer working for the Morrisville Municipal Authority said the project will draw Delaware River water rather than drinking water, in an amount requested to equal about 70 homes’ use. These are useful figures, but an annual average says little about hot-weather peaks, and a requested amount is not metered use. Ford also said the nearest homes are about 5,000 feet away and will hear no noise from the complex. That answers the residential question, not the one resident John Baldassano raised: whether generators running during an emergency would disrupt workers at industrial sites closer by.

    DEP’s Jillian Gallagher said projected emissions are similar to those of hospitals, universities and snack food manufacturers in the state. Resident Seema Kazmi countered that the area is already pollution-stressed. A per-facility threshold does not by itself address cumulative burden, which is her concern. Her point about emissions from grid power supplying the site also falls outside the scope of this air permit. Better data is coming. The state budget signed July 12 requires large data centers to report their energy and water use annually to DEP. Proposals to tie tax incentives to new clean generation, and to pause data center development, did not pass.

    Background

    Amazon’s Falls Township campus sits on the Keystone Trade Center, a former U.S. Steel mill site along the Delaware River in Bucks County, north of Philadelphia. It is one of two Pennsylvania data center projects in Amazon’s pledged $20 billion state investment. The other is in Salem Township beside the Susquehanna nuclear plant. The local utility, PECO, operates within PJM, the regional grid and wholesale power market that spans Pennsylvania, New Jersey and neighboring states.

    Large data centers keep fleets of on-site generators so servers stay running through grid outages. In Pennsylvania these generators require air quality approval from DEP. Pennsylvania has courted data center investment through fast-track permitting and tax abatements, while debate continues over grid costs, water use and local pollution.

    Sources

    Source: Amazon to install hundreds of gas-fired generators at Bucks County data center site (WHYY), reporting on Pennsylvania DEP’s planned air permit approval and the July 14, 2026 public meeting in Falls Township.