ZutaCore, a developer of two-phase, direct-to-chip liquid cooling technology, has raised a $100 million Series C round to expand its cooling platform for AI data centers, according to a report published by Pulse 2.0 on June 6, 2026. The reported purpose of the raise is to scale the company’s platform as AI workloads push rack power densities beyond what air cooling can handle.
Executive Summary
The headline fact is simple: ZutaCore has secured $100 million in Series C funding to expand its AI data center cooling platform. At that size, the round places ZutaCore among the better-capitalized independent players in liquid cooling, a segment that has moved from niche engineering concern to strategic infrastructure category in roughly three years.
Why it matters: modern AI accelerators draw hundreds of watts per chip, and racks packed with them can reach power densities that air-based cooling physically cannot dissipate economically. That has turned the cooling layer — cold plates, coolant distribution units, dielectric fluids, and the engineering services around them — into one of the most actively funded niches in data center infrastructure. A $100 million commitment to a two-phase cooling specialist signals that investors believe the transition to liquid cooling is durable, and that there is room in the market beyond the largest incumbent thermal vendors.
Capital Keeps Flooding the Cooling Layer
Cooling used to be a line item buyers negotiated down. In the AI build-out it has become a gating constraint: if you cannot remove the heat, you cannot deploy the chips, no matter how much power or floor space you have. That inversion explains why investors have poured money into thermal specialists across every approach — single-phase cold plates, immersion tanks, rear-door heat exchangers, and two-phase systems like ZutaCore’s. A $100 million Series C for a company focused specifically on the AI cooling problem fits squarely into that pattern and suggests the funding window for the category remained open as of mid-2026.
The strategic logic for investors is that cooling vendors sit at a chokepoint. Every generation of AI accelerator raises thermal design power — the amount of heat a chip is engineered to shed — and each increase expands the addressable market for liquid cooling retrofits and new builds alike. The risk, equally, is that a crowded field of well-funded competitors compresses margins before any single vendor achieves scale.
What Two-Phase Cooling Actually Is — and Why It Is Contested Ground
Most liquid cooling deployed today is single-phase direct-to-chip: water or a water-glycol mix flows through a cold plate bolted to the processor, absorbs heat, and carries it away without changing state. Two-phase cooling instead uses an engineered dielectric fluid — a liquid that does not conduct electricity — that boils on contact with the hot chip. The phase change from liquid to vapor absorbs far more energy per unit of fluid than simple warming does, which is the core efficiency argument for the approach. ZutaCore has long positioned its platform around this waterless, two-phase principle, marketing it as eliminating the risk of water leaks onto expensive electronics.
The counterarguments are practical rather than theoretical. Two-phase systems are mechanically more complex, the specialty fluids cost more than water, and the fluorinated chemistries commonly used in the category face growing regulatory scrutiny in several jurisdictions. Meanwhile single-phase cold plates have become the default choice for the current generation of AI racks because hyperscalers understand water. ZutaCore’s raise is, implicitly, a bet that as chip power keeps climbing, the physics advantage of phase change wins share back from the simpler incumbent approach. The release, as reported, does not detail how the company plans to argue that case to buyers.
Winners, Losers, and the Consolidation Question
If the round accelerates ZutaCore’s manufacturing and deployment capacity, the immediate beneficiaries are data center operators seeking alternatives to water-based cooling — particularly in facilities where water usage or leak risk is a board-level concern. Chipmakers benefit from any credible expansion of thermal headroom, since cooling capability directly constrains how they can specify future products.
The open competitive question is whether independent cooling specialists remain independent. The thermal management sector has seen sustained acquisition interest from large industrial and infrastructure players, and a well-funded specialist with differentiated technology is a natural target. A Series C of this size can be read two ways: as fuel for a run at standalone scale, or as valuation-building ahead of eventual consolidation. The reporting available does not indicate which trajectory ZutaCore’s investors have in mind.
Background
ZutaCore is a specialist in waterless, two-phase, direct-to-chip liquid cooling, an approach it has promoted for years as a safer and denser alternative to water-based cold plates. The company sells the hardware and supporting infrastructure that let standard servers shed heat through a dielectric fluid that vaporizes on the processor, and it has positioned that platform squarely at the AI data center market as accelerator power consumption has climbed.
The broader context is a rapid industry transition: liquid cooling moved from a high-performance-computing niche to mainstream AI infrastructure in the mid-2020s, drawing venture capital, private equity, and acquisition interest across cold plate, immersion, and two-phase vendors alike. ZutaCore’s Series C lands in the middle of that capital wave.
PJM Interconnection’s independent market monitor has concluded that AI-driven data center growth is reshaping the power markets it oversees, according to a June 2026 report from Data Center Knowledge. PJM operates the largest wholesale electricity market in the United States, coordinating the grid across 13 states and the District of Columbia for roughly 65 million people.
The finding matters because it comes from the market’s designated referee rather than from a vendor or developer: the monitor exists precisely to assess, without commercial interest, whether the market is functioning competitively — and it is now attributing a fundamental shift in that market to data center load.
Executive Summary
The headline is short but consequential: PJM’s market monitor — the independent body charged with policing competition in the nation’s largest electricity market — has identified AI data center growth as a force actively reshaping that market. For two decades, US grid planners worked in a world of essentially flat electricity demand, where efficiency gains offset economic growth. That assumption has broken, and PJM, whose footprint includes Northern Virginia’s Data Center Alley, is where it broke first and hardest.
When the market monitor says demand growth is ‘reshaping’ the market, it is signaling that data center load is no longer a forecasting footnote but a structural driver of prices, planning, and investment decisions. PJM’s recent capacity auctions — the mechanism that pays generators to be available years in advance — have produced record-setting results widely attributed in part to surging demand forecasts, and those costs flow through utility bills to every customer class.
For the industry, an independent confirmation of this shift cuts both ways. It validates the scale of the AI infrastructure build-out that developers have been describing. It also raises the stakes for how that growth is managed: who pays for new transmission and generation, how speculative interconnection requests are filtered from real ones, and whether supply can be added fast enough to keep reliability and affordability intact.
From Forecasting Footnote to Structural Force
The most important word in this story is ‘reshaping.’ Grid operators revise load forecasts constantly; what they rarely do is declare that the character of the market itself has changed. PJM’s service territory covers all or part of 13 states and DC, and it includes the densest concentration of data centers on the planet in Northern Virginia. When demand there grows, it does not simply add megawatts — it changes which power plants run, where transmission congestion appears, and how much capacity the market must procure years ahead.
An assessment from the independent market monitor carries different weight than one from PJM itself or from data center developers. The monitor’s role — in PJM’s case performed by an outside firm — is to evaluate market competitiveness and flag structural problems without a commercial stake in the outcome. Its reports are read closely by federal and state regulators. Framing AI data center growth as market-reshaping effectively puts the issue on the regulatory agenda, not just the industry conference circuit.
Capacity Markets, and Who Ends Up Paying
PJM runs a capacity market: generators are paid not only for the electricity they produce but for committing to be available during future peak periods. When demand forecasts rise sharply — as data center growth has caused them to — the market must procure more capacity against a supply base that has been shrinking as older coal and gas plants retire. Basic economics follows: tighter supply against higher demand means higher clearing prices, and PJM’s recent auctions have set records that state officials and consumer advocates have publicly protested.
Capacity costs are socialized across ratepayers, which is where the political friction originates. Households and small businesses in PJM states are seeing bill increases driven partly by demand they did not create. Expect the policy debate to center on cost allocation: large-load tariffs that require data centers to underwrite the infrastructure they trigger, minimum take-or-pay commitments, and rules for co-located or behind-the-meter arrangements where a data center pairs directly with a power plant. How those rules land will materially affect data center project economics in the region.
Winners, Losers, and the Speculation Problem
The near-term winners are clear: owners of existing generation in PJM, whose assets have been revalued by scarcity, and transmission developers with projects in flight. Data center operators with secured power — signed interconnection agreements and energized substations — hold an asset that is increasingly the scarcest input in the industry. The squeezed parties are late-arriving developers facing multi-year waits for grid connection, and energy-intensive industries competing for the same electrons.
The unresolved analytical problem is demand-forecast quality. It is widely acknowledged in the industry that developers file interconnection requests with multiple utilities for the same prospective project, meaning some portion of announced demand is duplicative or speculative. If markets procure capacity against inflated forecasts, ratepayers overpay; if forecasts are discounted too aggressively and the load shows up, reliability suffers. Distinguishing real load from phantom load is arguably the central technical challenge the monitor’s finding implies — and one the industry itself has an interest in helping solve, since credibility with regulators depends on it.
The Supply Response Is the Whole Game
High prices are a symptom; the cure is new supply, and here timelines diverge badly. A hyperscale data center can be built in roughly two to three years. New gas turbines face multi-year equipment backlogs, nuclear operates on decade scales, and renewables plus storage — often the fastest option — face their own interconnection queues and siting fights. Transmission, the connective tissue, is slower still.
That mismatch, more than any single auction result, is what ‘reshaping the market’ means in practice. It pushes data center operators toward creative structures: siting near existing generation, contracting directly for new-build power, investing in on-site generation, and accepting flexibility obligations — curtailing or shifting load during grid stress — in exchange for faster connection. For infrastructure providers, grid access has moved from a line item in site selection to the decisive variable.
Background
PJM traces its roots to a 1927 power pool between Pennsylvania and New Jersey utilities and has grown into the largest regional transmission organization in the US, dispatching power across 13 states and DC. An independent market monitor oversees its wholesale markets and publishes regular assessments of their competitiveness and health. For most of the 2000s and 2010s, PJM — like the rest of the US grid — planned around flat demand, as efficiency gains offset economic growth.
That era ended as cloud and then AI data center construction accelerated, concentrated in PJM territory around Northern Virginia. The region’s recent capacity auctions have produced record-setting prices that drew objections from state officials and consumer advocates, putting data center load growth at the center of an escalating debate over grid reliability, cost allocation, and how fast new generation and transmission can be built.
Utah’s governor has tightened the rules that apply to a giant AI data center project backed by investor Kevin O’Leary, according to a Business Insider report published May 30, 2026. The action places state-level conditions on one of the highest-profile celebrity-backed entries into the AI infrastructure race.
Details of the specific requirements were not spelled out in the available source material, but the reported move fits a broader pattern: states courting AI data center investment are simultaneously attaching guardrails around the resources those campuses consume — chiefly water and electric power.
Executive Summary
According to Business Insider, Utah’s governor moved to tighten the rules governing Kevin O’Leary’s planned large-scale AI data center in the state. O’Leary, the investor best known from Shark Tank, has spent the past two years positioning O’Leary Ventures as a developer of very large AI computing campuses, most prominently the multibillion-dollar ‘Wonder Valley’ concept announced in Alberta, Canada, in late 2024. A Utah project extends that ambition into one of the fastest-growing — and driest — states in the American West.
Why it matters: AI data centers are among the most resource-intensive facilities ever built at commercial scale. A single hyperscale campus can demand hundreds of megawatts of electricity — comparable to a small city — and, depending on cooling design, substantial water. Utah is an arid state where water politics are already charged, notably around the shrinking Great Salt Lake. When a governor personally intervenes to condition a marquee project, it tells the industry that resource guardrails are moving from county zoning boards up to the statehouse.
For developers, the message is that incentives and permits increasingly come bundled with obligations. For AI tenants and investors, it means project timelines and economics now carry a regulatory variable that did not meaningfully exist three years ago.
Guardrails Are Becoming the Price of Admission
Through 2023 and 2024, states competed for data centers almost purely with carrots: tax abatements, fast-track permitting, cheap land. The reported Utah action reflects the next phase. Legislatures and governors in Georgia, Virginia, Texas, and elsewhere have begun asking who pays for the grid upgrades a gigawatt-class campus requires, and whether existing ratepayers end up subsidizing a private tenant’s load. Utah itself passed legislation in 2024 creating a framework for ‘large load’ customers to be served under separate terms, precisely so that massive new consumers do not shift costs onto households. Tightening rules on a flagship AI project is consistent with that trajectory: welcome the investment, but ring-fence its externalities.
For laypeople, the key concept is that electricity and water are shared systems. A data center does not simply buy power the way a household does; at hundreds of megawatts it reshapes the utility’s entire planning horizon — what plants get built, what transmission lines get strung, and who bears the cost if the promised load never materializes.
Water Is the West’s Hard Constraint
Power can, eventually, be built. Water in the Great Basin largely cannot. Utah is one of the driest states in the country, and the decline of the Great Salt Lake has made every large new water commitment politically visible. Data centers vary enormously here: evaporative cooling designs can consume millions of gallons a day, while closed-loop and air-cooled designs use a small fraction of that — at the cost of higher electricity draw. Any state-imposed water condition effectively forces a design decision, pushing developers toward dry cooling and shifting the burden back onto the power system. That trade-off — water versus watts — is now a central engineering and political negotiation in every arid-state siting, and Utah’s reported action puts it on the record at the gubernatorial level.
The Celebrity-Capital Model Meets Institutional Reality
Kevin O’Leary’s data center ventures have been announced with characteristic showmanship — Wonder Valley in Alberta was unveiled with a headline figure of roughly $70 billion over its life. Announcements at that scale invite fair scrutiny: mega-campuses require anchor tenants, firm power agreements, water rights, transmission interconnection, and tens of billions in project finance, most of which is rarely secured at announcement time. A governor tightening the rules is, in one reading, simply the institutional system doing its job — converting a promotional vision into enforceable commitments. That is not necessarily adversarial. Projects that survive rigorous conditioning tend to be more bankable, because lenders and hyperscale tenants prefer sites where the regulatory ground has already been tested.
Winners, Losers, and the Signal to the Market
If the guardrails are well designed, the winners are Utah ratepayers, competing water users, and — perhaps counterintuitively — disciplined developers, who gain a clearer rulebook than rivals face in states still improvising. The risk side: conditions that are vague or shifting can chill investment, and Utah competes with Texas, Wyoming, and the Midwest for AI capital. AI tenants watching this will price in regulatory friction when choosing between states. The market signal is unmistakable either way: the era of announcing a gigawatt campus first and settling the resource questions later is closing.
Background
The AI boom that followed ChatGPT’s 2022 debut triggered a global race to build computing campuses of unprecedented scale, drawing in hyperscalers, private equity, sovereign funds — and celebrity investors. Kevin O’Leary entered the field through O’Leary Ventures, announcing the ‘Wonder Valley’ mega-campus in Alberta in December 2024 with a stated long-term vision of roughly $70 billion, and subsequently pursuing sites in the United States, including Utah.
Utah, meanwhile, has courted technology infrastructure — Meta and others operate large facilities there — while wrestling with the American West’s defining constraint: water. In 2024 the state established a legal framework for serving very large new electricity loads without shifting costs to ordinary ratepayers. The reported tightening of rules on the O’Leary project sits at the intersection of those two currents: aggressive AI-infrastructure recruitment and hardening resource guardrails.
Data Center Knowledge reported on May 30, 2026 that water and wastewater capacity have joined — and in some markets now rival — electrical power as the decisive factors in where AI data centers can be built. The report’s framing marks a shift in an industry that has spent the past several years describing its siting problem almost entirely in megawatts.
Executive Summary
The report argues that the availability of water for cooling, and just as importantly the capacity of municipal systems to accept the water a facility discharges, now determine whether an AI data center project is viable at a given site. That is a meaningful reframing: since the AI buildout accelerated, the industry conversation has centered on grid interconnection queues and power procurement, with water treated as a secondary sustainability metric rather than a gating constraint.
Why it matters: if water and wastewater capacity are genuine go/no-go criteria, the map of viable AI data center locations changes. Sites with abundant power but strained water or sewer systems lose ground, while regions with underused water and treatment infrastructure gain a new selling point. It also pulls a different set of actors — water utilities, sewer authorities, and municipal planners — into negotiations that were previously dominated by electric utilities.
From Megawatts to Gallons: A New Siting Calculus
For most of the AI infrastructure boom, the binding constraint has been electricity: how many megawatts a utility can deliver, and how fast. Water has been discussed mostly in sustainability reports. The shift Data Center Knowledge describes — water as a siting decision, not a disclosure line item — reflects how AI-scale facilities actually work. High-density computing throws off enormous heat, and many cooling designs, particularly evaporative systems, consume large volumes of water to reject that heat to the atmosphere. A campus that can secure power but not water is still an unbuildable campus.
Wastewater is the less obvious half of the equation, and arguably the more interesting one. Water that runs through cooling systems and is not evaporated must go somewhere, often into municipal sewer systems as industrial discharge. Treatment plants are sized for the communities they serve; a single large industrial user can consume capacity a municipality planned to allocate over decades of residential growth. Discharge from cooling systems can also be warmer and more mineral-concentrated than household wastewater, which treatment plants must be equipped to handle. A town can have a river next door and still lack the permits, pipes, and treatment headroom to host an AI campus.
Winners, Losers, and the New Bargaining Table
If this framing holds, the winners are jurisdictions that can offer both power and water headroom — including regions with cooler climates that reduce cooling demand, or with industrial water infrastructure left over from manufacturing that has since departed. Water utilities and engineering firms that design treatment and reuse systems gain leverage and business. The relative losers are water-stressed markets that have competed for data centers on power and tax incentives alone, and developers holding land banks in places where the sewer authority, not the electric utility, turns out to be the limiting party.
For operators, the economics push toward designs that trade water for electricity or capital: closed-loop liquid cooling, dry coolers, and water recycling all reduce consumption but raise power draw or upfront cost. That trade-off means water scarcity does not just move projects — it changes their engineering and their operating cost profile. Expect water-use effectiveness (WUE), the industry’s ratio of water consumed per unit of computing energy, to get the same contractual and public scrutiny that power-use effectiveness (PUE) received a decade ago.
What the Framing Does and Does Not Establish
A note of even-handedness: the source available to us is a report headline and premise, not a dataset. The claim that water now “decides” siting is directionally consistent with well-documented industry trends — public disputes over data center water use in drought-affected regions, and the growth of water-positive pledges from major cloud providers — but the strength of the claim varies by market. In cool, wet regions with modern treatment plants, water may barely register as a constraint; in arid, fast-growing metros it can be decisive. Readers should treat “water decides siting” as an increasingly common condition, not a universal law, and ask for market-specific evidence — permit denials, moratoria, or utility capacity studies — before generalizing.
Background
Since the generative AI boom began in late 2022, data center development has grown at a pace that strained electric grids, making interconnection queues and power procurement the industry’s defining bottleneck. Water surfaced periodically as a flashpoint — community disputes over data center water consumption in drought-affected regions drew attention, and major cloud providers responded with public water-stewardship and replenishment pledges — but it was generally treated as a reputational issue rather than a siting gate.
Data Center Knowledge, the trade publication behind the report, has covered the industry’s infrastructure constraints throughout the buildout. Its framing of water and wastewater as decisive siting factors reflects the arrival of AI-scale campuses whose cooling demands, and whose discharge volumes, exceed what many municipal systems were designed to accommodate.
Politico reported on May 30, 2026 that the North American Electric Reliability Corporation (NERC) — the body that writes and enforces mandatory reliability rules for the continent’s bulk power grid — is pushing back on AI companies demanding rapid grid connections for their data centers. The message from the grid’s gatekeeper, per the report’s framing: the newest and hungriest class of electricity customers needs to learn the rules that everyone else on the grid already plays by.
Executive Summary
The AI buildout has turned electric power into the binding constraint on data center construction, and companies that once measured competition in chips now measure it in megawatts and interconnection dates. Politico’s report captures the resulting collision: AI developers want grid connections on startup timelines, while NERC — an organization most people outside the utility industry have never heard of — insists that speed cannot come at the expense of the engineering discipline that keeps the lights on.
It matters because NERC is not a lobbying group or a trade association. It is the FERC-certified reliability regulator for the bulk power system, and its standards carry legal force for the utilities and grid operators who would actually plug these data centers in. When NERC signals that giant new loads deserve closer scrutiny, that posture propagates into utility study processes, interconnection agreements, and ultimately into how fast — and under what conditions — AI capacity gets energized.
The Grid’s Gatekeeper Steps Into the AI Boom
NERC occupies an unusual position in American infrastructure: a not-for-profit corporation whose reliability standards are mandatory and enforceable, with penalty authority, under oversight from the Federal Energy Regulatory Commission. Its job is narrow but existential — keep the bulk power system from failing — and it has historically focused on the supply side: generators, transmission owners, and grid operators. The AI era is dragging it toward the demand side, because individual data center campuses are now being proposed at scales that used to describe power plants or small cities.
That shift explains the tone Politico’s headline captures. For decades, new load arrived gradually and predictably, and reliability planning could treat demand as a smooth curve. A single AI campus that wants hundreds of megawatts on an aggressive schedule breaks that model. From NERC’s vantage point, the question is not whether AI is worth powering — it is whether loads this large, connecting this fast, behave in ways the grid’s protection schemes, planning studies, and operating procedures were built to handle.
Why Giant Loads Make Reliability Engineers Nervous
An ‘interconnection’ is the formal process of studying and approving a new connection to the grid, so that a new customer or generator does not destabilize the network around it. Reliability engineers worry about large data centers for reasons that have little to do with total energy consumption. These facilities can change their draw very quickly, and their internal protection systems can disconnect them from the grid in a fraction of a second during a routine voltage disturbance. When a load the size of a small city vanishes instantaneously, the surplus power has to go somewhere, and the grid must absorb the swing without cascading into a wider failure. NERC has been studying exactly this class of large-load behavior in its recent reliability work.
This is why ‘learn the rules’ is more than institutional gatekeeping. The rules — ride-through expectations, modeling requirements, coordination of protection settings — exist because the bulk power system is a single interconnected machine, and every large participant’s behavior affects everyone else on it. AI developers accustomed to moving at software speed are encountering a domain where the failure modes are physical, shared, and measured in blackouts rather than bugs.
Speed Versus Stability: The Economics of the Standoff
Time-to-power is now arguably the scarcest commodity in AI infrastructure. A data center that energizes a year earlier than a rival’s can capture training contracts and cloud commitments worth far more than the cost of the facility’s electricity. That asymmetry pushes AI companies to treat interconnection queues and study timelines as bureaucratic friction to be compressed — and pushes them toward workarounds like on-site generation and co-location with existing power plants, arrangements that are themselves generating regulatory disputes.
The likely equilibrium is not that either side simply wins. Grid operators and utilities want this load — it is the largest organic demand growth the industry has seen in a generation, and it spreads fixed costs over more sales. But reliability institutions cannot underwrite shortcuts, because they absorb the blame when the system fails. Expect the practical outcome to favor developers who invest early in grid engineering competence: those who show up with credible load models, flexible operating commitments, and patience for the study process will connect faster than those who treat the grid as a vendor to be pressured. In infrastructure, sophistication about the rules is itself a competitive advantage.
Background
NERC traces its origins to the aftermath of the 1965 Northeast blackout, and its standards became mandatory and enforceable after the 2003 blackout prompted Congress to create a certified Electric Reliability Organization in the Energy Policy Act of 2005. For most of its history, its work centered on generators, transmission owners, and grid operators — the supply side of the system.
That focus is shifting because U.S. electricity demand, roughly flat for two decades, is now growing again, with AI data centers among the largest drivers. Individual campuses are being proposed at scales once associated with power plants, and NERC’s recent reliability assessments have increasingly flagged large loads — their size, speed of arrival, and electrical behavior — as an emerging risk category the grid’s rules were not originally designed around.
Data Center Dynamics has published a comparison of the two competing approaches to direct-to-chip liquid cooling — single-phase, where a liquid coolant absorbs heat and stays liquid, and two-phase, where the coolant boils at the chip and carries heat away as vapor — framed around a single question: which is right for AI data centers in 2026?
That the trade press is treating this as a live, unsettled debate is itself the news. As AI accelerators push per-chip power beyond what air can remove, direct-to-chip liquid cooling has moved from exotic to expected, and the industry has not yet converged on which of the two variants will define the next generation of facilities.
Executive Summary
Direct-to-chip liquid cooling puts a cold plate in contact with the processor and runs coolant through it, removing heat far more efficiently than blowing air across a heatsink. Within that category, two architectures are competing. Single-phase systems circulate a liquid — typically treated water or a water-glycol mix — that warms up as it passes over the chip and is cooled elsewhere. Two-phase systems use an engineered dielectric fluid that boils directly on the cold plate; the phase change from liquid to vapor absorbs a large amount of heat at a nearly constant temperature, and the vapor is condensed back to liquid to repeat the cycle.
The choice matters because it is not easily reversible. Coolant chemistry, pressure ratings, manifolds, coolant distribution units, and facility water loops are all designed around one approach or the other. An operator committing today to a multi-hundred-megawatt AI campus is effectively placing a bet on which architecture will best handle the chips of 2028 and beyond — and on which supply chain, service model, and regulatory environment will mature fastest.
The DCD piece lands at the moment this bet has become unavoidable. Air cooling handled decades of servers; single-phase liquid is handling today’s AI racks; the open question is whether tomorrow’s thermal densities force the industry through a second transition to two-phase — or whether single-phase engineering keeps stretching to meet the need.
Why the Question Exists at All
For most of computing history, this debate would have been academic. Air cooling was cheap, well understood, and sufficient. AI training hardware broke that equilibrium: modern accelerators concentrate so much power in so little silicon that the limiting factor is no longer the data center’s chillers but the last few millimeters between the chip surface and the coolant. Direct-to-chip designs attack exactly that bottleneck, which is why they have become the default assumption for new AI builds.
Single-phase direct-to-chip won the first round largely on familiarity. Water-based cooling loops are a known quantity — data center engineers, plumbers, and component suppliers have decades of experience with pumps, valves, and leak management for liquid water. Two-phase systems promise something physically compelling in exchange for novelty: boiling a fluid absorbs latent heat, meaning the coolant can soak up substantially more energy without a large temperature rise, and it does so uniformly across the hottest parts of the chip.
The Engineering Trade-Offs, Plainly Stated
Single-phase’s strengths are operational. The fluids are inexpensive and benign, the components are commodity, leaks are messy but manageable, and the industry’s existing skills transfer directly. Its weakness is headroom: as chips run hotter, single-phase designs must push more liquid, faster, through smaller channels, and must manage the temperature gradient across the cold plate — the chip’s inlet edge runs cooler than its outlet edge, which complicates thermal design as power climbs.
Two-phase inverts that profile. Boiling heat transfer offers high performance and near-isothermal operation — the whole cold plate sits close to the fluid’s boiling point — which is attractive precisely where single-phase strains. But the costs are real: engineered dielectric fluids are far more expensive than water, systems must manage vapor and pressure rather than simple liquid flow, servicing a sealed two-phase loop is a different discipline, and several candidate fluids belong to chemical families (such as PFAS-related compounds) facing regulatory scrutiny in major markets. A technically superior heat-transfer mechanism does not automatically win if its fluid supply or compliance picture is uncertain.
Who Wins and Loses on Each Path
If single-phase continues to stretch, the winners are incumbents: established cooling vendors, existing supply chains, and operators who have already deployed water-based loops and want continuity. Chip designers absorb more of the burden, engineering packages and cold plates to live within single-phase limits. If two-phase becomes necessary, the advantage shifts toward specialist fluid and systems companies, and toward operators willing to build new competencies early — with the corresponding risk of backing immature technology.
There is also a middle path worth naming: hybrid facilities, where single-phase handles the bulk of the load and two-phase (or other advanced techniques) is reserved for the hottest components or highest-density halls. Many operators will likely hedge this way rather than commit wholesale, which suggests the 2026 answer to “which is right?” may genuinely be “both, in different places” — an unsatisfying but rational outcome for an industry making thirty-year infrastructure bets on three-year chip roadmaps.
What This Means for the Broader Market
The cooling decision cascades outward. Coolant choice affects how much heat a facility can reject to the outside world and at what temperature, which shapes heat-reuse opportunities and water consumption. It affects colocation providers, who must decide which architecture to offer tenants whose hardware they do not control. And it affects the retrofit market: the vast installed base of air-cooled data centers faces different conversion economics depending on which liquid architecture prevails. Standardization efforts — common connectors, fluid specifications, and safety practices — will matter as much as raw thermal performance in determining which camp scales fastest.
Background
Data centers spent decades cooled almost entirely by air: chilled air pushed through raised floors and hot aisles, with per-rack power low enough that fans and heatsinks sufficed. The AI buildout broke that model. Training clusters pack accelerators drawing unprecedented power into dense racks, pushing the industry through its biggest thermal transition since the mainframe era — first to rear-door heat exchangers and now to liquid brought directly to the chip.
Data Center Dynamics, the publication behind this comparison, is a long-running trade outlet covering data center design and operations. That its editorial attention has moved from whether to liquid-cool to which liquid architecture to choose reflects how quickly direct-to-chip cooling has become the baseline assumption for AI infrastructure — and how much unresolved engineering debate still sits beneath that baseline.
Vattenfall, the Swedish state-owned energy company and one of Europe’s largest power producers, announced on 27 May 2026 a partnership with Nscale, an AI infrastructure provider with operations in Norway, to support the growth of AI infrastructure in the country. The arrangement pairs Vattenfall’s position in the Nordic power market with Nscale’s GPU-based data center capacity.
The announcement, published through Vattenfall’s newsroom, frames the deal around enabling AI compute expansion in Norway with clean Nordic energy. Specific capacity figures, financial terms, and timelines were not detailed in the source material available to us.
Executive Summary
The partnership joins two sides of the equation that now defines AI infrastructure: electricity and compute. Vattenfall brings decades of experience generating and trading power in the Nordic region, where abundant hydropower keeps both electricity prices and carbon intensity among the lowest in Europe. Nscale brings the other half — data centers built to house GPUs (graphics processing units, the specialized chips that train and run AI models) — including an existing Norwegian footprint.
Why it matters: access to power has replaced access to chips as the binding constraint on AI buildout in much of the world. Grid connection queues in major markets stretch years, and hyperscalers increasingly sign deals directly with energy companies rather than waiting in line. A named partnership between a major European utility and a GPU infrastructure specialist is a signal of how the market is reorganizing — with power producers moving up the value chain toward compute, and compute providers moving upstream toward generation.
For Norway specifically, the deal reinforces the country’s bid to convert its renewable surplus into digital exports rather than only raw electricity — though it also lands amid an active Norwegian debate about which industries deserve scarce grid capacity.
Why AI Compute Keeps Moving North
The Nordics offer a combination few regions can match: hydropower-dominated grids with low, relatively stable wholesale prices; a cold climate that slashes cooling costs (cooling can be a significant share of a data center’s energy bill in warmer markets); political stability; and strong fiber connectivity to continental Europe. Norway in particular generates the overwhelming majority of its electricity from hydropower, which is both renewable and — unlike wind and solar — dispatchable, meaning it can run around the clock the way AI training clusters demand.
That is why Norway has attracted a steady stream of data center investment over the past decade, and why AI-focused operators like Nscale planted their flags there. Training large AI models is less latency-sensitive than serving consumer applications, so remote-but-cheap-and-green locations are a rational fit for training workloads even when end users are far away.
What a Utility Brings to the GPU Race
The scarce resource in AI infrastructure is no longer just GPUs — it is firm, sizable grid connections and the energy to feed them. Utilities control exactly that. A partnership with Vattenfall potentially gives an AI infrastructure operator earlier visibility into available capacity, structured long-term power purchase agreements (PPAs — contracts that lock in electricity supply and price for years), and credibility with grid operators and regulators. For Vattenfall, AI data centers represent something European utilities have lacked for years: large, creditworthy, growing demand in a region where industrial electricity consumption had been flat.
This mirrors a broader industry pattern of energy companies and compute companies converging — through PPAs, co-located campuses, and equity partnerships. The strategic logic is sound on both sides, but the value of any specific deal depends entirely on terms the parties disclose: how much power, at what price, for how long, and with what firmness. None of that is specified in the material available here.
A Thin Release, and the Questions Norway Is Already Asking
Based on the source available, this reads as a directional announcement rather than a detailed commercial agreement — no megawatts, sites, investment figures, or delivery dates are cited. That does not make it empty: named partnerships between a state-owned utility and an AI infrastructure firm typically precede concrete projects, and both parties accept reputational cost if nothing follows. But readers should distinguish between an announced intent to cooperate and a contracted buildout.
The deal also lands in a live Norwegian policy debate. Norway’s grid operators have faced more connection requests than the system can serve, and policymakers have discussed prioritizing which loads get capacity — weighing data centers against electrifying industry and transport. A fair reading is that partnerships like this one are partly designed to navigate that environment: aligning with an established utility is a way to demonstrate seriousness and secure standing in the queue. Whether Norwegian regulators and communities view AI data centers as valuable industry or as competition for their renewable advantage remains an open, legitimate question on all sides.
Background
Vattenfall, founded in 1909 and wholly owned by the Swedish state, is one of Europe’s largest electricity producers, with a generation fleet spanning Nordic hydropower, wind, and nuclear, and a stated strategy of enabling fossil-free energy across its markets. Nscale is a newer entrant that emerged in the mid-2020s wave of AI infrastructure specialists, building GPU data centers for AI training and inference and anchoring its early operations in Norway to take advantage of hydropower and a cool climate.
The partnership fits a broader industry realignment: as AI compute demand collided with constrained power grids across Europe and North America, energy companies and compute providers began pairing up through power purchase agreements, co-located campuses, and strategic alliances. The Nordics — with cheap renewable power and cold air — have been among the biggest beneficiaries of that shift, attracting hyperscalers and specialist operators alike over the past decade.
A $3.6 billion artificial-intelligence data center campus is planned for Rapides Parish in central Louisiana, according to a May 25, 2026 report by the Louisiana Illuminator. The project would rank among the largest private capital investments in the parish’s history and, per the reporting, involves a power arrangement with Cleco, the regulated utility serving the region.
Executive Summary
The reported plan places a multibillion-dollar AI campus in Rapides Parish, whose seat is Alexandria — a part of Louisiana that has not historically competed for hyperscale data center projects. At $3.6 billion, the investment is on the scale that typically implies hundreds of megawatts of computing load, purpose-built substations, and years of construction, though the report available to us does not specify capacity, acreage, or a construction timeline.
Why it matters: the announcement is another data point in a clear pattern. AI training and inference facilities are landing in the South — Louisiana, Mississippi, Texas, Georgia — where land is available, power can be contracted at scale, and state incentives are aggressive. For a mid-sized regulated utility like Cleco, a single customer of this size can reshape its entire resource plan. That dynamic, more than the campus itself, is the story worth watching.
Louisiana’s Second Act in the AI Land Rush
Louisiana entered the hyperscale conversation in late 2024, when Meta announced a roughly $10 billion AI data center campus in Richland Parish in the state’s northeast — at the time the largest such announcement in Meta’s fleet. That project demonstrated that Louisiana could deliver what hyperscalers need: large contiguous sites, a cooperative regulatory environment, and a utility (there, Entergy Louisiana) willing to build generation for a single anchor customer. A $3.6 billion campus in Rapides Parish suggests that playbook is now being run in Cleco territory as well.
For central Louisiana, the economic-development logic is straightforward. Data centers bring outsized capital investment and property-tax base relative to their headcount — construction employs thousands for several years, but steady-state operations typically employ dozens to a few hundred. Communities weighing these projects should therefore evaluate them primarily as tax-base and infrastructure plays rather than as mass employers, a distinction that matters when incentives are negotiated.
Why the Utility Is the Real Story
Cleco serves roughly the central third of Louisiana and is small compared with national investor-owned utilities. A data center campus at this investment level would likely represent a load addition measured in hundreds of megawatts — material against a system of Cleco’s size. In regulated markets, serving that load means new generation, transmission upgrades, or long-term power purchases, all of which flow through integrated resource plans and rate proceedings before the Louisiana Public Service Commission.
The central question in every such deal is cost allocation: does the data center customer pay the full incremental cost of the capacity built to serve it, or do some costs socialize across residential and small-business ratepayers? Utilities and regulators across the South are actively developing large-load tariffs — special rate classes with long contract terms, minimum-take provisions, and exit fees — precisely to answer that question. The report available to us does not disclose the structure of the Cleco arrangement, so the fairest reading is that this is the item most deserving of public scrutiny as the project moves through regulatory review.
The Economics of Gigawatt-Scale Siting
The South’s dominance in recent AI-infrastructure siting comes down to arithmetic. Training-class AI facilities are constrained less by fiber or labor than by time-to-power: how quickly a utility can deliver hundreds of megawatts of firm capacity. States with vertically integrated utilities can compress that timeline by building dedicated generation, something fragmented or capacity-constrained markets struggle to match. Add comparatively cheap land, natural-gas proximity, and sales-tax exemptions on data center equipment, and the region’s pipeline of announcements becomes easy to explain.
The risk side deserves equal weight. Multibillion-dollar campus announcements are commitments of intent, not completed buildings; across the industry, some announced projects have been resized, phased, or delayed as AI demand forecasts and chip supply evolve. A parish and utility that invest in infrastructure ahead of a project that later shrinks can be left carrying costs. Well-structured agreements put that risk on the developer through take-or-pay terms — which is why the unpublished details matter more than the headline number.
Background
Louisiana emerged as an AI-infrastructure destination in late 2024, when Meta selected Richland Parish for a roughly $10 billion data center campus backed by dedicated generation from Entergy Louisiana — at announcement, one of the largest data center commitments in the United States. The state offers hyperscalers large rural sites, abundant natural gas, sales-tax relief on data center equipment, and vertically integrated utilities that can build power for anchor customers.
Cleco, headquartered in Pineville in Rapides Parish itself, is central Louisiana’s regulated utility. For a utility of its size, a single hyperscale customer represents a step-change in load — the kind of demand shock that utilities across the South are now addressing through integrated resource plans and new large-load rate structures overseen by state regulators.
Data Center Frontier profiled TeraWulf’s Lake Mariner campus in Barker, New York, in a May 25, 2026 feature framing the site as a prototype for the “AI factory” — a large-scale data center purpose-built for artificial-intelligence computing. The campus occupies the site of the retired Somerset coal-fired power plant on the shore of Lake Ontario, and the piece traces how TeraWulf, a company that began as a bitcoin miner, has been converting that inherited industrial infrastructure into high-performance computing capacity.
Executive Summary
The core story is one of conversion twice over: a coal plant site converted to digital infrastructure, and a cryptocurrency-mining operator converting itself into an AI-infrastructure landlord. Lake Mariner’s appeal rests on assets that are nearly impossible to recreate quickly — an existing high-capacity grid interconnection built for a power station, access to abundant water for cooling, zoned industrial land, and a regional grid in upstate New York that draws heavily on zero-carbon hydroelectric generation.
Why it matters: the binding constraint on AI data center construction has shifted from chips to power. Utilities in major markets are quoting multi-year waits for large new grid connections, so sites that already have them — like retired thermal power plants — jump the queue. If Lake Mariner works as a template, the industry gains a playbook for turning stranded fossil-fuel assets into AI campuses, with meaningful implications for former coal communities, grid planners, and the competitive map of the data center industry.
The Interconnection Is the Asset
A modern AI campus can require as much electricity as a small city, and the slowest step in delivering it is usually not construction but the grid interconnection — the physical and contractual link that lets a facility draw power from the transmission system. New requests in constrained markets can sit in utility study queues for years. A retired power plant inverts that problem: the wires, switchyard, and transmission rights were built to push hundreds of megawatts out, and much of that capacity can be repurposed to pull power in.
That is the essence of the Lake Mariner thesis. TeraWulf did not have to win a greenfield site fight; it inherited the Somerset plant’s industrial footprint and grid position. The same logic explains a broader industry pattern — operators across the market have been scouting retired or retiring thermal plants precisely because the interconnection, land, and water rights are already in place. In that sense the “prototype” label is apt: the question the site tests is whether coal-to-compute conversion can be repeated at scale, not whether it can be done once.
From Bitcoin Mine to AI Landlord
TeraWulf built Lake Mariner as a bitcoin mining facility, and that history matters more than it might appear. Bitcoin mining taught the company to energize large amounts of power-dense compute quickly and cheaply — but mining revenue is volatile, tied to cryptocurrency prices and periodic “halving” events that cut miner rewards. High-performance computing (HPC) hosting for AI customers offers something mining never could: multi-year contracted revenue from creditworthy counterparties, which is the kind of cash flow lenders and infrastructure investors will finance.
The catch is that the two businesses are less similar than the shared electrical infrastructure suggests. AI training clusters demand far higher reliability, denser cooling — increasingly liquid cooling delivered directly to the chips — and enterprise-grade operations that mining sheds never needed. The conversion is therefore a genuine re-engineering exercise, not a tenant swap, and execution on that transition is the fair test by which TeraWulf and its bitcoin-miner peers should be judged.
The Zero-Carbon Power Angle
Upstate New York’s grid is unusually clean by U.S. standards, anchored by large-scale hydroelectric generation. For AI customers under pressure to report the carbon footprint of their computing, siting workloads on a predominantly zero-carbon grid is a marketable advantage — and there is a certain narrative symmetry in AI compute replacing coal combustion on the same acreage.
The claim deserves precision, though. A clean regional grid is not the same as dedicated clean power, and every large new load consumes headroom that grid planners had earmarked for other purposes. The substantive questions for any site making a sustainability case are how the incremental demand is matched with generation, and what the facility’s water and community impacts look like — questions that apply to Lake Mariner exactly as they apply to every competing campus.
Winners, Losers, and the Watchlist Question
If the coal-to-AI conversion model scales, the winners include former plant communities that regain a tax base and jobs, utilities that get to reuse stranded transmission assets, and early movers holding converted sites when capacity is scarce. The pressure lands on operators pursuing greenfield builds in queue-constrained markets, who must wait for infrastructure that conversion players already own.
For investors treating TeraWulf as a watchlist company, the prototype framing cuts both ways. It signals genuine strategic differentiation — but prototypes, by definition, have not yet proven repeatability. The durable questions are contract quality (who the tenants are and for how long), financing cost for the heavy capital expenditure AI-grade buildings require, and whether the company can operate to the uptime standards hyperscale customers demand. A compelling site thesis is necessary but not sufficient.
Background
TeraWulf was founded to mine bitcoin using predominantly zero-carbon energy and developed Lake Mariner on the grounds of the retired Somerset coal plant in Barker, New York, drawing on the region’s hydro-heavy grid. As demand for AI computing surged and power became the industry’s binding constraint, TeraWulf — like several other large miners — began redeveloping its energized sites for high-performance computing tenants, betting that its grid position would be worth more serving AI than mining cryptocurrency.
The broader market context is a structural shortage of grid-connected capacity: AI’s growth has pushed utilities in major data center markets to years-long interconnection queues, elevating any site with existing power infrastructure — especially former power plants — into strategic real estate.
GridCare, a Palo Alto-based grid-software startup, has raised $64 million to accelerate the connection of AI data centers to the electric grid, according to a SiliconANGLE report published May 16, 2026. The company’s pitch is to identify “stranded” capacity — grid headroom that already exists but sits unused most hours of the year — and match it with data-center developers who would otherwise wait years in utility interconnection queues.
Executive Summary
The announcement itself is brief: a $64 million raise aimed at speeding up AI data-center projects. The report, as circulated, does not name the investors, the round’s structure, or a valuation. What makes the round worth analyzing is the problem it targets. The single biggest constraint on AI infrastructure buildout in the United States is no longer chips or capital — it is access to electric power, and specifically the multi-year queue to connect large new loads to the grid.
GridCare’s approach inverts the usual answer to that problem. Rather than building new generation or new transmission — both slow — it uses software to find places where the existing grid has spare room, then structures deals in which the data center agrees to flex its consumption during the relatively few hours when the grid is genuinely constrained. If that model works at scale, it converts a five-year infrastructure problem into a months-long contracting problem. A $64 million round suggests investors believe the thesis has moved past the pilot stage, though the public announcement offers little evidence either way — a gap we detail below.
Why the Interconnection Queue Became AI’s Bottleneck
Every large new electricity user or generator must apply to connect to the grid, and the utility or regional grid operator must study whether the connection would overload wires and transformers. That process — the interconnection queue — has become the choke point of the AI era. Lawrence Berkeley National Laboratory has tracked roughly 2,600 gigawatts of generation and storage stuck in U.S. queues, and wait times for large projects commonly stretch five years or more. Data centers seeking hundreds of megawatts of new load face similar studies, similar upgrade bills, and similar timelines.
For AI developers, that timeline is intolerable. Model-training capacity is a competitive weapon measured in quarters, not decades, which is why hyperscalers have turned to behind-the-meter gas turbines, nuclear power-purchase agreements, and sites in far-flung jurisdictions. Any company that can credibly compress grid access from years to months is selling exactly what the market’s most aggressive buyers want most.
The Stranded-Capacity Thesis
The grid is engineered for its worst hour — the summer evening peak when air conditioning, industry, and households all draw at once. The rest of the year, much of that capacity idles. GridCare’s argument, made publicly since it emerged in 2025, is that this latent headroom is enormous; the company has claimed more than 100 gigawatts could be unlocked nationally. The catch is that a data center can only use that headroom if it gets out of the way during the constrained hours — by throttling workloads, drawing on batteries, or running on-site generation for a small fraction of the year.
The economics can be attractive for every party. The utility monetizes an asset it already built and spreads fixed costs over more sales, which can put downward pressure on rates for other customers. The data center trades a modest flexibility obligation for years of saved time — and in AI economics, time-to-power is often worth more than the cost of the flexibility. The hard part is trust and verification: utilities need enforceable guarantees that the load will actually curtail when called, because the consequence of a broken promise is overloaded equipment, not a missed earnings estimate. Software that can model, contract, and verify that behavior is the actual product being financed here.
A Crowded Race Around the Queue
GridCare is not alone in attacking time-to-power. Competitors come from several directions: developers building behind-the-meter gas or geothermal generation, battery vendors marketing “bridge power,” utilities themselves rolling out flexible-interconnection tariffs, and grid-analytics firms courting the same utility relationships. Regulators are moving too — federal and state proceedings on large-load interconnection are actively reshaping the rules GridCare must operate within, which is both a tailwind (flexibility is gaining formal recognition) and a risk (a tariff change can rewrite the business model overnight).
The structural advantage of the stranded-capacity approach is that it requires no new steel in the ground; its structural weakness is that it depends on utility cooperation, and utilities adopt new commercial models slowly. The likely winners of this period are parties on both sides of the deal: utilities that learn to monetize flexibility, and data-center developers that treat power procurement as a portfolio rather than betting on a single path. The losers are projects that queued up conventionally and now watch flexible newcomers connect first.
What $64M Signals — and What It Doesn’t
A round of this size, roughly a year after the company’s reported seed financing, implies investors saw commercial traction worth underwriting. But it is worth being precise about what the public announcement substantiates: a funding amount and a stated purpose. It does not, as circulated, disclose customers, megawatts under contract, utility partnerships, or revenue. Venture funding validates a thesis’s attractiveness to investors, not its operational success. The evidence that matters — signed interconnection agreements and data centers energized ahead of queue timelines — will come from utility filings and customer announcements, not from the size of the round.
Background
GridCare is a Palo Alto-based startup that emerged from stealth in 2025 with a reported $13.5 million seed round, led by founder and CEO Amit Narayan — who previously founded AutoGrid, a grid-flexibility software company acquired by Schneider Electric in 2022. GridCare’s founding claim is that over 100 gigawatts of usable capacity is hiding in plain sight on the U.S. grid, accessible to data centers willing to be flexible.
The company operates against the backdrop of a historic grid bottleneck: Lawrence Berkeley National Laboratory has tracked roughly 2,600 gigawatts of projects waiting in U.S. interconnection queues, and the surge in AI data-center demand since 2023 has made time-to-power the defining constraint of the buildout. Regulators, utilities, and hyperscalers are all converging on flexibility — the idea that new loads can connect faster if they yield during peak hours — as one of the few near-term answers.
