An Essay on AI, Compute, and the Long Game

America doesn't get to opt out of the AI race. Stratos is the part where we show up.

The case for powerful AI systems, the cybersecurity stakes that deserve plain talk, and a closer look at what the Stratos project in Box Elder County actually is.

Mark Hilgenberg Salt Lake County, Utah May 2026 ~12 min read · expand for sources

In May 2026, the UK's AI Security Institute published a finding worth pausing on. The length of complex cybersecurity tasks that frontier AI models can complete on their own has been doubling roughly every four to five months. see the evidence

The newest models — Claude Mythos Preview and GPT-5.5 — blew past that already-accelerated trend. Both completed multi-hour expert-level offensive cyber tasks at near-100% success rates. One of them solved a 32-step simulated corporate network attack that no prior model had ever finished.

This is the world we now live in. The question is not whether powerful AI exists. It does. The question is who builds the infrastructure that runs it, who sets the standards, and who has the compute capacity to defend against the same capabilities being developed by adversaries who feel no obligation to slow down.

Frontier AI is now strong enough to be a meaningful cybersecurity actor. The only real decision left is whether American defenders have parity with American adversaries.

01 / The Upside Is Not HypotheticalThe case for building these systems

Critics often talk about AI as a future risk to be feared. The benefits are already here and worth being specific about.

Medicine

AI imaging models are running in radiology departments at major hospitals today, catching early-stage cancers that the human eye misses. Drug discovery pipelines that historically took a decade are being compressed to a few years. The first AI-designed drugs are now in human clinical trials. None of this is speculative. the receipts

Science and the economy

Protein folding was a fifty-year unsolved problem. AlphaFold solved enough of it that researchers can now design proteins from scratch — the foundation for new vaccines, materials, and enzymes. Climate models, weather forecasting, battery chemistry, fusion containment — AI is now the leading research instrument in fields that used to take entire careers to advance. expand

Cybersecurity, in particular

This one deserves attention. Modern enterprise security cannot run at human speed anymore. The volume of telemetry, the speed of attacks, and the sophistication of the adversary all require AI on the defending side. The defenders who have the best models win. The ones who don't, lose. the defender case

4.7 mo
AI cyber capability doubling time
~50%
drug discovery timeline reduction
200M+
protein structures predicted by AlphaFold

02 / Why This Requires Compute at ScaleYou cannot do this on a laptop

Training a frontier AI model is one of the most computationally demanding tasks humans have ever attempted. The current generation of models requires tens of thousands of specialized chips running for months. The next generation will require an order of magnitude more.

This is not a software problem. It is a steel, concrete, electricity, and cooling problem. Whichever country builds the physical infrastructure to support the next decade of training runs gets to define the safety standards, the deployment policies, and the commercial terms for the rest of the world. why scale matters

The number that gets misquoted

You will see people online say America has 5,400 data centers to China's 400, and use that to argue we're already ahead. That comparison is misleading because it counts every facility — corporate storage, social media servers, video streaming caches, cloud apps. The number that actually matters for the AI race is the count of frontier AI training campuses. By that measure (per the AI Data Center Index, May 2026): the United States has 84 tracked AI facilities with 70.6 GW of capacity. China has 7 facilities with 2.5 GW. The US lead is enormous. It is also fragile — China is building gigawatt-scale AI campuses on aggressive timelines while domestic permitting in the US can take years. read the comparison

03 / What Stratos Actually IsThe project, on the record

Stratos is the proposed AI compute campus in Hansel Valley, Box Elder County, Utah. It is being developed by O'Leary Digital — yes, that Kevin O'Leary — as part of a larger initiative called Wonder Valley that also includes a sister site in Alberta.

The Utah project area covers about 40,000 acres of mostly unincorporated land. The data center buildings themselves will only sit on a small fraction of that — Phase 1 is capped at a 2,000-acre footprint, and even at full buildout the campus structures occupy a sliver of the total acreage. The rest is buffer, utility corridors, on-site power generation, and undeveloped land. full specs and history

How the phases work

Governor Cox attached an important condition to the project: it cannot be built all at once. The phases are gated, with reviews and clear metrics required after each stage before the next gets approved.

The phases will not be running the same technology

This is worth saying directly because it tends to get lost in the debate. Phase 1 will be built with 2026-2027 generation chips and cooling. Phase 2, if it gets approved, will use whatever exists at that time — probably more efficient silicon, better cooling, possibly novel power sources. Phase 3 is far enough out that some of the technology may not exist yet. Orbital compute, advanced nuclear, and architectures we cannot yet describe are all in the realm of plausible. The phased gating is a feature, not a bug — it forces the project to justify itself against the technology of the moment, not the technology of 2026.

Phased buildout also means Phase 1 approval is not approval of the full 9 GW. State air quality review, water rights adjudication, and the Design Review Committee still have to weigh in on every stage. The original water rights application was withdrawn and the new application is now under review. The county received a $16.2 million upfront commitment from the developer to offset initial impacts.

This is what's actually on the table. A lot of the public debate has been about a project that does not match the one being built.

04 / The Real Concerns, Taken SeriouslyWhere the opposition has a point — and where it doesn't

Water

This is the strongest concern and deserves the most honest engagement. Northern Utah is in drought. The Great Salt Lake is in crisis. Any large industrial water user in this basin has to be scrutinized.

The Stratos design uses hybrid closed-loop cooling, which recirculates the same coolant continuously instead of evaporating it. The difference matters enormously. Existing Utah data centers tell the story: NSA Bluffdale uses about 126 million gallons of water per year on mostly evaporative cooling. Novva's closed-loop campus in West Jordan uses about 3 million gallons for similar capacity — over 40 times less. Stratos is being designed in the Novva tradition, with treated water returned to the local aquifer that feeds the Great Salt Lake.

Oracle has published detailed designs for closed-loop cooling in its AI data centers. Microsoft is piloting zero-water-cooling in Arizona and Wisconsin. NVIDIA's Vera Rubin platform, entering production in late 2026, runs on 113°F warm water — hot enough to eliminate chillers entirely. By the time Stratos Phase 2 is being built, that is likely the standard. The original Stratos water rights application was withdrawn precisely because the developer agreed the first version did not meet the bar. That is the system working. the full water analysis

Heat and the "23 atom bombs a day" claim

The same Utah State physicist quoted on the 16 GW figure has argued that the campus would release waste heat equivalent to 23 atomic bombs per day, raise local nighttime temperatures by up to 28°F, and turn Hansel Valley into a Sahara-like desert. That number has now traveled in Futurism, Rolling Stone, and across social media. It deserves a careful response, because the math behind it is a rhetorical sleight of hand and the substantive answer is more interesting than the framing.

On the math: 23 "Hiroshima-sized bombs" of waste heat per day works out to about 14 kilotons of TNT each — the smallest nuclear weapon ever built, and roughly 1/3,500th the size of the largest bomb ever tested. The largest hydrogen bomb in history released, in one second, about 150 times more energy than the entire daily output Davies is describing. The Sun delivers more energy to Earth every 90 seconds than Stratos would release over its entire decade-plus operational life. The "atomic bomb" unit is designed to feel alarming, not to inform.

On the substance: waste heat from data centers is recoverable and increasingly required to be recovered. Stockholm's Data Parks program uses waste heat from over 20 data centers to warm 30,000 apartments. Denmark's Orbis data center exports 165,000 MWh of heat per year, supplying nearly 11,000 homes. Microsoft is building the world's largest data center heat recovery scheme outside Helsinki to heat the city of Espoo. Germany now requires waste heat reuse. California's Title 24 is the first US state mandate. NVIDIA's Vera Rubin platform produces outlet water at temperatures specifically suited for greenhouse, agricultural, and district-heating uses downstream. The same warm-water cooling that eliminates chillers also makes the waste heat economically useful for the first time. the full heat analysis

Air quality and emissions

The site sits outside the Wasatch Front non-attainment zone, but emissions do not respect zoning lines. Phase 1 natural gas turbines will produce nitrogen oxides and CO2. The Utah Division of Air Quality has not yet completed its review, and it should. This is a legitimate area for public input.

What rarely gets mentioned is what's coming behind Phase 1. The Allam-Fetvedt Cycle, developed by 8 Rivers and commercialized by NET Power, burns natural gas with pure oxygen and uses supercritical CO2 as the working fluid. Byproducts: liquid water and a stream of pipeline-ready CO2. Atmospheric emissions: essentially zero. A 50 MW demonstration plant has been delivering power to the ERCOT grid since 2021. Commercial 280-300 MW plants are under development now. Carbon-capture-ready turbine design at Stratos means capture equipment can be added when the economics permit. Small modular reactors from NuScale, X-energy, Kairos, and others are in NRC review, with first commercial operation expected 2028-2030. The Phase 2 and Phase 3 generation mix will not look like the Phase 1 mix. expand

The "actually it will use 16 gigawatts" claim

A Utah State physics professor was quoted in Rolling Stone arguing the campus will actually consume closer to 16 GW because natural gas generation is thermodynamically inefficient. That number has been widely repeated. Whether the professor's full analysis was captured fairly by the publication is its own question — Rolling Stone has an editorial slant and selective quoting is not uncommon — so what follows responds to the framing as reported.

The thermodynamics is real. Every gas turbine loses energy as heat in fuel-to-electricity conversion. What the framing misses: the same loss applies to every gas plant in the grid. Modern hyperscale facilities run at PUE ratios of 1.09 to 1.20 versus 1.5 to 1.8 for traditional enterprise data centers. AI chip efficiency improves roughly 30% per generation. NVIDIA's 800-volt DC architecture, shipping now, cuts power-chain losses by up to 60% by eliminating multiple conversion steps. Direct-to-chip liquid cooling cuts cooling overhead dramatically. Dry cooling eliminates evaporative water use entirely in dry climates. Each phase of Stratos will be built with whatever is most efficient at that moment. the full efficiency breakdown

The "this is just for surveilling Americans" claim

This is one of the loudest opposition framings, picked up by Books Behind Borders, NaturalNews, SHTF Plan, and a layer of social media accounts. The framing: Stratos is a giant surveillance facility being snuck through MIDA authority to spy on US citizens. It is worth taking the claim seriously and looking at whether the infrastructure profile actually fits the accusation.

It doesn't, and not by a small margin. The NSA's Utah Data Center at Bluffdale — the country's actual flagship signals-intelligence facility, the one a Greenpeace blimp famously flew over in 2014 with "Illegal Spying Below" painted on it — runs at roughly 100 to 150 megawatts of total power. Stratos at full buildout is 7,500 to 9,000 megawatts. That is 50 to 90 times larger than the facility that actually does the work critics are accusing Stratos of being built to do. No surveillance program in any country needs that much compute. Bulk telephone metadata, internet traffic interception, and pattern-of-life analysis are large workloads, but they're not gigawatt-scale-AI-training workloads. The shape is wrong.

And surveillance does not require US soil to begin with. The Five Eyes intelligence partnership already conducts signals collection from facilities in the United Kingdom, Canada, Australia, and New Zealand, plus overseas SIGINT stations around the world. The legal architecture for collecting on Americans through allied partners has existed for decades. Building a 9 GW campus in northern Utah specifically to surveil US citizens would be the most expensive and least efficient way to do something that is already being done elsewhere on much smaller infrastructure.

The actual DoD connection at Stratos is something different and worth being honest about: a senior Air Force official requested that MIDA identify locations to expand military computing capacity, and Stratos was structured as a response to that request. That is AI training and inference for defense applications — pattern recognition in satellite imagery, logistics optimization, cyber defense, autonomous systems research — which is the kind of workload that genuinely benefits from gigawatt-scale GPU compute. Most of the campus, in practice, will serve commercial AI customers (hyperscalers, labs, enterprise), with revenue partially flowing back to military infrastructure under the MIDA arrangement. That is a legitimate policy question. It is not a domestic spying program. the full breakdown

05 / Why This Is PersonalMy heart, my kids, and what AI is actually doing

People ask how AI actually helps people's health. Let me tell you from my own life.

At 40 years old, with my wife seven months pregnant with our fourth child, I had emergency heart bypass surgery. The cause wasn't diet. It wasn't sedentary living. It wasn't anything the standard cardiac risk score would have flagged. The cause was lipoprotein(a) — Lp(a) — a genetically inherited cholesterol-like particle that drives aggressive atherosclerosis and that statins and most standard cholesterol medications cannot meaningfully lower.

About one in five people worldwide carry elevated Lp(a). That is roughly 1.4 billion people. Most of them have no idea. The American Heart Association now recommends a one-time test for every adult, and almost nobody is getting tested. I found out the hard way, with my chest open on an operating table.

For most of medical history, there was nothing to do about it. You knew you had it, you took aggressive measures on everything else (LDL, blood pressure, lifestyle), and you hoped.

That is changing because of AI.

Five drugs are now in late-stage clinical trials specifically to lower Lp(a) — pelacarsen, olpasiran, lepodisiran, zerlasiran, and an oral small-molecule called muvalaplin. In Phase 2 trials, these drugs reduced Lp(a) by 80 to 98 percent. Most of them are antisense oligonucleotides or small interfering RNAs that silence the LPA gene at the source, before the dangerous particle even gets produced. The Phase 3 cardiovascular outcomes data from the Lp(a) HORIZON trial, the first cardiovascular outcomes trial ever conducted for an Lp(a)-lowering therapy, is reading out this year.

AI is woven into how these drugs were designed. Machine learning platforms helped researchers identify which RNA sequences would bind most precisely to the LPA gene. Predictive models accelerated the screening of candidate compounds. Computational biology helped optimize delivery — figuring out how to get the drug specifically into liver cells where Lp(a) is produced. Generative protein and nucleic acid design tools, the same kind of work that earned a Nobel Prize for AlphaFold in 2024, are now being applied to the next generation of cardiovascular therapies.

I am over 20 years out from my bypass. I am still here. The daughter I almost didn't see is now 20 years old. I want to walk her down the aisle. I want to be there for the grandkids I haven't met yet. AI-driven cardiovascular research is one of the reasons I think those things are possible.

Beyond the trials, AI is becoming the most useful tool I have for managing my own health — and it has stopped looking like a search engine and started looking like a research assistant that works while I sleep. I have agents scanning ClinicalTrials.gov on a schedule, watching for new Lp(a) studies I might qualify for and alerting me when enrollment opens. When I get new lab work, I hand it over and the system cross-references my values against my baseline trends, current cardiology guidelines, and recent journal literature — flagging anything that has drifted or any interaction with what I am currently taking. I have it monitoring the pipeline of cardiovascular research for me, surfacing relevant preprints and trial readouts as they happen rather than waiting for me to search. It builds exercise programs that respect my cardiac history and adjusts them as my data changes. It cross-checks supplements and medications for interactions. It is not a glorified search engine — it is an autonomous layer working continuously in the background on my behalf. None of this replaces my cardiologist. It makes me the most informed patient who walks into his office, and it catches questions I would not have known to ask.

This is what AI infrastructure is actually for. The compute that runs at Stratos and the other AI campuses being built across America is the same compute that trains the models that design these drugs, that reads the medical literature, that compares my labs against millions of similar patients, that helps researchers find the next class of therapy. Every additional gigawatt of American AI capacity is, downstream, more lives saved from diseases that used to be untreatable.

AI is not the enemy. It is turning genetic risks that were once death sentences into manageable conditions. For people like me with Lp(a), for the other 1.4 billion of us out there, this is not abstract — it is the difference between watching our kids grow up and not.

That is why I support building. That is why I support Stratos. And it is why I think the people opposing it deserve a clearer picture than the one they have been given — including a clearer picture of who is giving it to them.

06 / Who Is Organizing the OppositionFrom a federal data-center moratorium bill down to your county commission meeting

Most of the people showing up to Box Elder County Commission meetings are sincere. They have wells, ranches, kids, and a justified skepticism of large-scale industrial projects. They deserve honest answers — and they have not been getting them. The framing they are hearing about Stratos has not bubbled up from Hansel Valley. It has been built top-down by a national political project, then pushed down through state, regional, and local layers until it reaches the county fairgrounds. Here is what that project actually looks like.

The federal layer: Sanders, AOC, and a national data-center moratorium bill

On March 25, 2026, Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced the Artificial Intelligence Data Center Moratorium Act of 2026. The bill would impose a nationwide halt on construction or upgrading of AI data centers until Congress passes federal legislation regulating AI. It would also ban exports of advanced AI chips to countries lacking equivalent regulations. In Sanders's words at the introduction press conference: "we cannot sit back and allow a handful of billionaire Big Tech oligarchs to make decisions that will reshape our economy, our democracy and the future of humanity."

This is not a fringe gesture. The bill names AI infrastructure as the political target. It would, if passed, freeze Stratos and every other proposed AI campus in the country. Senator John Fetterman called the bill a "China First" policy. The bill is unlikely to pass this Congress. The political project around it is not designed to pass legislation — it is designed to shift the Overton window so that local approvals like Box Elder County's become politically toxic. The bill is the strategic anchor; opposing Stratos is the tactical execution. expand

The state layer: at least 11 states with data-center pause proposals

Lawmakers in at least 11 states have proposed moratoriums or pauses on data-center construction, according to the National Conference of State Legislatures. Senators Markey and Warren are separately leading an effort to block any federal preemption of state-level AI regulation — they sent a Dear Colleague letter to Senate Democrats in late 2025 urging unified opposition to a Big Tech-backed AI regulation moratorium in the NDAA. Their framing is that states must retain authority to regulate AI on their own. In practice, that means each state Democratic Party can run its own version of the moratorium fight independent of federal politics, and the state-level apparatus exists to do exactly that.

The funded national-organization layer

The opposition's national messaging has documented funding sources worth knowing about. This is not the same as saying the people repeating the messaging are paid — they aren't. The sincere local resident who shares a Peoples Dispatch article about Stratos is not being funded by anyone. They are repeating a frame that originated upstream, where the funding actually lives.

Neville Roy Singham. Singham is an American tech entrepreneur who sold his consulting firm ThoughtWorks for $785 million in 2017 and now resides in Shanghai. A New York Times investigation in August 2023 tracked hundreds of millions of dollars from Singham-linked entities to a global network of advocacy organizations and media outlets that, in the Times's words, "mix progressive advocacy with Chinese government talking points." Singham has personally attended a Chinese Communist Party workshop on promoting the party internationally, according to NYT reporting, and previously worked as a consultant for Huawei from 2001 to 2008. He denies acting at the direction of the Chinese government.

One of the most prominent organizations in his funded network is Peoples Dispatch — the same publication whose framing of Stratos as a fight over "who controls the infrastructure behind the AI economy" has been reproduced across opposition coverage. The People's Forum, a Manhattan-based organization funded by Singham and his wife Jodie Evans (co-founder of CODEPINK), publicly acknowledged receiving over $20 million from them between 2017 and 2022 through shell companies and donor-advised funds, according to a September 2025 House Ways and Means investigation. BreakThrough News, No Cold War, and several Party for Socialism and Liberation-aligned media projects sit in the same network. These are some of the outlets producing the framings that get picked up and shared by sincere local opponents who have no idea where the talking points came from.

Hansjörg Wyss. Separately, an April 2026 American Energy Institute report identified twelve organizations actively opposing US data-center development that have collectively received more than $39 million in funding from foreign donors based in Switzerland, the United Kingdom, and Denmark. The most prominent named donor: Swiss billionaire Hansjörg Wyss, whose Berger Action Fund channels his personal fortune (Wyss, as a foreign national, is barred by US law from donating directly to American candidates or political committees) through dark-money vehicles including the Sixteen Thirty Fund. Top recipient organizations include Indivisible, 350.org, Oil Change International, GAIA, the Sierra Club, and Americans for Financial Reform — mainstream environmental and advocacy groups whose underlying missions are entirely legitimate, and most of whose individual members are sincere.

The point of naming this funding is not to claim that the rancher at the commission meeting is on someone's payroll. She isn't. The point is that the messaging she is sharing — the Peoples Dispatch frame, the moratorium-bill language, the "billionaires versus working people" framing — was authored upstream by organizations that are funded, in part, by foreign-resident billionaires routing money through American dark-money vehicles. Sincere people parroting top-down-funded messaging is a normal feature of every political fight, on the left and the right. It is worth knowing when it is happening. the money trail

The local layer: state parties, DSA, PSL, Peoples Dispatch

At the bottom of the funnel, the framing reaching residents in Hansel Valley comes through a specific set of channels. The coverage describing Stratos as a fight over "who controls the infrastructure behind the AI economy" comes from Peoples Dispatch, a publication tied to the international socialist movement. The Democratic Socialists of America have active Utah chapters involved in organizing. The Party for Socialism and Liberation has published organizing material framing American AI buildout as capitalist consolidation that working people should resist. State Democratic parties in several swing states have adopted anti-AI infrastructure planks. These organizations publish their goals openly — this is not a conspiracy theory.

It is worth knowing that their objection to Stratos is not really about water. It is about private capital allocating large-scale infrastructure at all. A version of Stratos that used 80% less water and emitted nothing would still be opposed by these groups on those grounds. who's saying what at the local level

If America's AI infrastructure gets built in Wyoming, Texas, Alberta, or Saudi Arabia instead of Utah, the project still gets built. The only thing Utah loses is the jobs and the tax base.

The Box Elder rancher worried about her well deserves a hearing on the substance. She is not on anyone's payroll, and the point of naming the funding sources above is not to suggest she is. The point is that the framing she is sharing — that Stratos is fundamentally about billionaires versus working people, that the moratorium is the responsible position, that the project should be stopped rather than carefully reviewed — was authored upstream by organizations that are funded, in part, by foreign-resident billionaires. Sincere local people parroting top-down-funded messaging is a normal feature of political fights, and the people doing it are usually doing so in good faith. They still deserve to know whose framing they have picked up.

Engage her concern honestly. Water rights. Air quality. Phased approval. Real review. Treat the substance as substance. Just do not mistake the messaging she is repeating for a spontaneous local consensus when, in fact, two of the largest funders behind the national framing she is echoing live in Shanghai and Geneva, while the infrastructure they want to stop would otherwise be built in Hansel Valley.

Build. Carefully. Honestly. But build.

Want the full picture?

Each section below expands the argument with sources, data, and the stories behind the numbers. Click any block to dig in.

Read the deep dive Back to top
— Below: Sources, Stories, and Receipts —
01.AThe AISI cyber capability evidence, in full+

The UK's AI Security Institute is a government research body inside the Department for Science, Innovation and Technology. They publish empirical work on what frontier AI models can actually do, with human baselines for comparison.

In May 2026, they reported that on their "narrow cyber suite" — a battery of cybersecurity exploitation tasks with timed human-expert baselines — the length of task that frontier models can complete autonomously had been doubling every 4.7 months since reasoning models emerged in late 2024. That doubling time is itself faster than their previous estimate of 8 months from November 2025.

Two newer models, Claude Mythos Preview and GPT-5.5, exceeded even the accelerated trend line. They completed multi-hour expert tasks at success rates approaching 100% under the test's token cap. On a longer-running benchmark — a 32-step simulated corporate network attack called "The Last Ones" — Mythos Preview solved it 6 times out of 10. Mythos also solved a second range called "Cooling Tower" 3 times out of 10. No model had ever completed "Cooling Tower" before.

What this means in plain terms

AI models can now perform several hours of expert-level offensive cybersecurity work without human intervention. The capability is improving faster than the safety institutes can publish about it. The METR organization, which runs a similar benchmark on broader software engineering tasks, has measured a consistent 4.2-month doubling time on their own benchmark.

Why this argues for building, not slowing down

This capability exists. It will exist regardless of whether Utah builds a data center. The only question is whether the people defending American networks have access to the same class of model running on the same class of infrastructure. If they don't, every American bank, hospital, power utility, and water system is operating with an asymmetric disadvantage against state-backed adversaries who have no such constraint.

02.AAI in medicine — what's deployed vs. what's promised+

Honest framing matters here because medical AI marketing tends to overstate. The actual state of play in 2026:

Deployed today

  • FDA-cleared diabetic retinopathy screening (IDx-DR) running in primary care offices without a specialist on site
  • FDA-cleared mammography assistance tools at hundreds of US imaging centers
  • Stroke triage models (Viz.ai, Aidoc) used in over a thousand US hospitals for rapid CT review
  • Radiology workflow models that prioritize urgent scans inside the reading queue

In late-stage trials

  • The first fully AI-designed small-molecule drugs have entered human Phase II trials (Insilico Medicine, INS018_055 for idiopathic pulmonary fibrosis)
  • Generative protein design from companies like Isomorphic Labs, Generate Biomedicines, and Cradle is producing candidates entering Phase I

Honest caveats

No fully AI-designed drug has received final FDA approval as of this writing. The clinical accuracy figures in early model studies often look better in published papers than in deployed reality, because real-world data is messier than benchmark data. The right framing is: these tools are starting to be useful and the trajectory is steep, not that they are already replacing physicians.

02.BScience: AlphaFold, materials, weather, fusion+

AlphaFold is the cleanest example. Protein structure prediction was a fifty-year open problem in biology. In 2021 DeepMind released structures for nearly all known proteins (200+ million), making them freely available to researchers worldwide. The downstream effect has been measurable: new antibiotics, new enzymes for plastic degradation, new vaccine candidates, and a generation of researchers who can now ask questions that were previously unanswerable.

Beyond biology, AI has produced material breakthroughs in:

  • Battery chemistry — DeepMind's GNoME identified roughly 380,000 stable inorganic crystal candidates, many with potential energy storage applications
  • Weather forecasting — Google's GraphCast and similar models now outperform traditional numerical weather prediction at a fraction of the compute cost, allowing communities to prepare for storms with better lead time
  • Fusion containment — Reinforcement learning has been used to control plasma in tokamak reactors (Swiss Plasma Center, DeepMind collaboration)
  • Mathematics — AI systems have produced novel proofs in combinatorics and conjectures in pure math
02.CThe cybersecurity defender case+

If you only read the AISI report, you might conclude that the right response to AI cyber capability is to slow AI development. That reading is wrong, and here's why.

Cybersecurity is fundamentally asymmetric. Attackers need to find one vulnerability. Defenders need to find all of them. Until recently this asymmetry favored the attacker because human analysts could not review code or telemetry at the scale modern systems generate.

AI changes this calculation in the defender's favor — if the defender has access to comparable capability. Modern AI systems can:

  • Audit large codebases for vulnerability classes in minutes rather than weeks
  • Triage security alerts at machine speed, freeing human analysts to focus on the genuinely novel
  • Detect anomalous behavior across telemetry that would overwhelm any human team
  • Generate patches for discovered vulnerabilities with reasoning that can be human-reviewed

Adversaries — including state-backed groups in China, Russia, Iran, and North Korea — are not going to stop developing offensive AI. They have built parallel chip supply chains, training infrastructure, and red-team operations. If American defenders do not have access to comparable compute and comparable models, the asymmetry of cyber warfare returns to favoring the attacker, with potentially catastrophic consequences for the banking system, the power grid, and hospitals.

This is the reason a thoughtful person who is worried about AI risk should still support American AI infrastructure. The alternative is unilateral disarmament.

03.AWhy training frontier models requires gigawatt-scale compute+

Training a frontier AI model is the most computationally demanding task humans have ever attempted at scale, and the scale is accelerating.

  • GPT-4 training is estimated to have required ~25,000 high-end GPUs running for around three months
  • Frontier 2025-2026 training runs are running on clusters of 100,000+ chips
  • Next-generation training runs anticipated for 2027 and beyond will require clusters approaching 1 million chips and gigawatts of dedicated power

The bottleneck has shifted. It used to be chip supply. Now it is increasingly power, cooling, and the physical real estate to host the infrastructure. A 7.5 GW campus like Stratos can support several frontier training runs simultaneously and a much larger volume of inference workload on top of that.

This is why the AI build-out is increasingly a real-estate, energy, and construction story rather than a pure software story. Whichever country wins the next decade of AI will have built the steel and concrete to host it.

03.BThe US vs. China compute picture — and why the data center counts you see online are misleading+

The bad comparison that's everywhere

You will see versions of this on social media: "The US has over 5,400 data centers and China only has 368, so America is way ahead." That comparison is technically true and almost completely useless for the AI race. Here's why.

A "data center" can mean almost anything — a single rack in a closet that serves a regional bank, a Facebook server farm hosting photos, a cloud storage facility, a streaming video cache, or a frontier AI training cluster with hundreds of thousands of GPUs. The total US figure of roughly 5,400 includes all of these. The vast majority are not training AI models. Most are running enterprise email, e-commerce, social media, and storage workloads that have nothing to do with frontier compute.

The metric that actually matters

The AI Data Center Index, a source-backed registry that tracks facilities specifically designed for AI workloads, gives a much clearer picture. As of May 2026:

  • United States: 84 tracked AI data center facilities, 70.6 GW of known capacity, 46 operators
  • China: 7 tracked AI data center facilities, 2.5 GW of known capacity
  • Canada: 8 facilities, 13.0 GW
  • United Arab Emirates: 8 facilities, 11.5 GW

Two important caveats on this comparison. First, the index tracks publicly disclosed facilities, and China's sovereign AI infrastructure is opaque by design — the actual count is almost certainly higher than 7. Second, even with that caveat, the disclosed gap between US and Chinese AI compute is enormous, on the order of 30x by capacity. The US lead in frontier AI infrastructure is real, not narrative.

The US AI buildout is still mostly ahead, not finished

Of the 84 tracked US facilities, only 46 are operational. The other 38 are in pipeline:

  • Under construction: 16 facilities, 23.0 GW
  • Planned: 13 facilities, 13.6 GW
  • Announced: 9 facilities, 9.6 GW

Stratos is part of this pipeline. So is Meta's 5 GW Hyperion campus in Louisiana, the 5 GW Stargate AI Supercomputer, Microsoft's 3.3 GW Project Fairwater in Wisconsin, AWS Project Rainier in Indiana (2.4 GW, where Anthropic trains models), and dozens of others. The US lead is being maintained by an active build, not coasting on operational infrastructure.

The "AI is all dirty fossil fuel" claim is false

Critics frequently argue that AI data centers represent a fossil-fuel boom. The actual energy mix of the 84 US AI facilities tells a different story:

  • Renewable: 44 facilities (52%)
  • Nuclear: 15 facilities (18%)
  • Mixed: 11 facilities (13%)
  • Grid / Unknown: 10 facilities (12%)
  • Fossil / Gas: 3 facilities (4%)
  • Hydroelectric: 1 facility (1%)

Pure gas-powered AI campuses are 4% of the US AI buildout. Nuclear-powered facilities — including Meta's 5 GW Louisiana site, the Three Mile Island reactivation for Microsoft, the Diablo Canyon AI integration, AWS's nuclear acquisition, and several Meta-Vistra and Meta-Oklo nuclear partnerships — outnumber pure-gas sites by 5 to 1.

Why the lead is fragile despite the numbers

  • America's "Big Four" hyperscalers (Alphabet, Amazon, Meta, Microsoft) plan combined AI spending of $650 billion in 2026, with total US AI compute infrastructure spend projected to exceed $2.8 trillion by 2029
  • China responded with a $100 billion "New Infrastructure" initiative and accelerated permitting
  • Chinese state policy treats AI infrastructure as a strategic priority — permits in weeks rather than years
  • US export controls on advanced chips have meaningfully slowed China's access to the most capable training silicon, but China is building parallel domestic supply chains and producing competitive (if not leading-edge) alternatives
  • The American bottleneck has shifted from chip supply to power, transformers, switchgear, and the buildout of physical infrastructure on multi-year permitting timelines

The takeaway: the US is genuinely ahead, the lead is meaningful, and the gap is not preserved by accident. Projects like Stratos — large, integrated AI campuses with on-site power — are how the lead gets maintained. The alternative is watching China close the gap with faster construction and looser permitting while American projects sit in environmental review.

03.CStratos: full specs, timeline, and structure+

The basics

  • Location: Hansel Valley, Box Elder County, Utah — about 62 square miles of mostly privately held land across three separate sites
  • Footprint: ~40,000 acres total project area; Phase 1 data center buildings confined to 2,000 acres; full buildout still uses only a fraction of total acreage
  • Power capacity: 7.5 to 9 gigawatts at full buildout (Gensler architectural materials show 7.5 GW; MIDA executive director Paul Morris cites 9 GW)
  • Primary power source: Natural gas via the Ruby Pipeline (TallGrass Energy supply agreement); designed to operate off-grid as an energy island
  • Developer: O'Leary Digital (Kevin O'Leary, CEO Paul Palandjian), partnered with West GenCo on energy
  • Architect: Gensler (master plan)
  • Cooling: Hybrid air and liquid, closed-loop
  • Reliability target: 99.999% availability with N+1 redundancy across critical systems

The phased buildout — this is the part that matters

Governor Spencer Cox imposed a critical condition on the project: development is approved only in stages, with required reviews and metric-based gates between phases. The county vote on May 4, 2026 was not approval of the full 9 GW campus.

  • Phase 1: Capped at 1.5 GW capacity, confined to a 2,000-acre footprint. Roughly $1 billion in site preparation, then more than $4 billion for the Phase 1 build (~500 MW initial buildings).
  • Early phase scaling: Approaching 3 GW at a cost approaching $20 billion, contingent on a Phase 1 review showing metrics are met.
  • Phase 2 and beyond: Subject to additional state and county approvals. The technology stack will not be the 2026 stack — by the time Phase 2 is being designed, chip generations, cooling architectures, and possibly power sources will have advanced. Orbital compute, advanced small modular nuclear, and other approaches are real possibilities for later phases.
  • Full buildout: 7.5 to 9 GW spread across a decade or more.

The MIDA structure

The Stratos Project Area is governed by Utah's Military Installation Development Authority. MIDA is a state authority created to support facilities tied to military readiness. Its enabling statute allows it to designate project areas, negotiate development agreements, and capture revenue that flows back to military infrastructure (in this case, Hill Air Force Base and the Utah National Guard).

The Stratos designation was requested by a senior Air Force official seeking to identify locations to expand military computing capacity. O'Leary Digital's response was framed as a direct answer to that solicitation. Stratos sits near Hill Air Force Base, Northrop Grumman's Promontory facility, and the Utah Test and Training Range.

Timeline and approvals

  • April 24, 2026: MIDA approved the development agreement
  • May 4, 2026: Box Elder County Commission voted to authorize the project area
  • May 2026: Utah Department of Environmental Quality confirmed no environmental application yet received — full permitting could take more than two years
  • Pending: Master plan submission to the Design Review Committee
  • Pending: New water rights application (the original was withdrawn) is under State Engineer review
  • Pending: Utah Division of Air Quality review
  • Construction: Phased over a decade-plus; each phase gated on review of the previous

Economics

  • $16.2 million upfront commitment from developer to offset initial county budget impacts
  • Energy use tax negotiated from 6% down to 0.5%, still expected to generate ~$30 million annually for Box Elder County in Phase 1
  • Estimated 2,000 permanent jobs; thousands more during construction
  • Phase 1 capital: $1 billion site prep + $4 billion+ build. Early phase total approaching $20 billion. Sister Alberta campus described as a $70 billion development.
04.AThe water question, honestly+

Northern Utah is in drought. The Great Salt Lake is at historic lows and continues to shrink. Any large industrial water user in this basin deserves scrutiny. That is the starting point.

What the Stratos design actually proposes

  • Hybrid closed-loop cooling (air-and-liquid), not the open evaporative cooling that older data centers used
  • Use of existing ranch water rights — cleaning naturally salty groundwater, running it through the closed loop, and returning most of the treated water back to the local aquifer that feeds the Great Salt Lake
  • Stated commitment to use less than 0.2% of annual river flow at full buildout
  • Closed-loop systems do reuse the same water repeatedly — but not indefinitely. Mineral and salt buildup requires periodic small purges and fresh-water additions to prevent corrosion. The alternative (draining and refilling the entire system at once) would shut the campus down for days, so designers do small ongoing purges instead. This is normal engineering, not waste.
  • On-site water utility infrastructure to be built as part of the campus

How Stratos compares to existing Utah data centers

This is the part most opposition coverage skips. Stratos is not the first major data center in Utah. Several have been operating for years, and their water use varies dramatically depending on cooling technology. Here is the comparison:

Facility Capacity Annual Water Cooling
NSA Bluffdale100-150 MW126M gallonsMostly evaporative
Aligned West Valley (SLC-04)~100+ MW80M gallonsEvaporative
Aligned West Jordan (SLC-02/03)~80-100 MW47M gallonsEvaporative
Meta Eagle Mountain100+ MW (expanding)35M+ gallonsMixed
eBay South Jordan~30-50 MW19.5M gallonsEvaporative
DataBank Granite Point~50+ MW7.7M gallonsMostly closed-loop
Novva West Jordan120 MW (expanding to 200+)~3M gallonsClosed-loop / waterless chillers

Look at the spread. Novva's closed-loop, waterless-chiller campus in West Jordan uses about 3 million gallons annually for 120+ MW of capacity. Compare that to evaporative-cooled NSA Bluffdale at 100-150 MW using 126 million gallons — over 40 times more water for similar capacity. This is the difference cooling technology makes. Stratos is being designed in the closed-loop tradition, not the evaporative one. And by Phase 2, Vera Rubin's 113°F warm-water cooling may eliminate the need for chillers entirely on portions of the load.

What's still unresolved

  • The original water rights application was withdrawn. A new application has now been filed and is under State Engineer review
  • The 0.2% figure is the developer's stated target, not an adjudicated limit
  • Water rights in Utah are governed by prior appropriation, and any new allocation must be reviewed by the State Engineer

Comparative context that often gets left out

Agriculture is by far the largest water user in the Great Salt Lake basin — alfalfa farming alone uses more water than every Utah city combined. A 7.5-9 GW data center with modern closed-loop cooling, even at the high end of reasonable estimates, would use a fraction of what existing agricultural users consume. The conversation about Utah water is legitimate, but it has to be honest about scale.

The cooling technology trajectory

Phase 1 cooling will be modern hyperscale closed-loop. By the time Phase 2 is being built — likely 2028 onward — the technology will have moved meaningfully further.

  • NVIDIA Vera Rubin warm-water cooling. NVIDIA's next-generation Vera Rubin platform, entering production in late 2026, runs on 45°C / 113°F warm-water direct liquid cooling. At that inlet temperature, no chillers are needed — the system can be cooled with ambient air across most of the year, especially in Utah's climate. Jensen Huang put it bluntly at CES 2026: "We are basically cooling this supercomputer with hot water." When the announcement hit, HVAC company stocks (Modine, Johnson Controls, Trane, Carrier) dropped 5-7% in a single day. The industry understands this is a structural shift.
  • Oracle's closed-loop AI data center design. Oracle has published detailed engineering for closed-loop cooling specifically for AI workloads — recirculating coolant continuously instead of evaporating it. Microsoft, Meta, and Google are all deploying similar designs in new builds.
  • Dry cooling. Air-to-fluid heat exchangers (essentially industrial-scale radiators) can run with near-zero water use in dry climates. Northern Utah's climate fits the profile well. SPG Dry Cooling and similar suppliers are now standard in new data center designs where water access is constrained.
  • Direct-to-chip and immersion cooling. NVIDIA's Blackwell platform is purpose-designed for direct-to-chip liquid cooling. Two-phase immersion cooling submerges entire server boards in engineered dielectric fluid — no fans, no water. Production deployments exist at LiquidStack, Iceotope, Submer, and others. These approaches can cut cooling energy more than 90% versus air cooling.
  • Heat reuse. Liquid cooling produces outlet temperatures up to 60°C, hot enough to be useful elsewhere — district heating, greenhouse warming, industrial processes. In Scandinavia, heat reuse is now a regulatory requirement for urban data centers.

The same campus that uses millions of gallons in Phase 1 may use a small fraction of that by Phase 3, on the same footprint, supporting more compute. The static-snapshot framing misses this entirely.

04.BHeat, the "23 bombs" claim, and why waste heat is recoverable+

What's actually being claimed

Robert Davies, a physics professor at Utah State University, gave The Salt Lake Tribune a calculation he summarized this way: Stratos at full buildout would be "the equivalent of about 23 atom bombs worth of energy dumped into this local environment every single day." Davies and BYU ecology professor Ben Abbott, who reviewed the calculations, projected the campus could raise local daytime temperatures by 5°F and nighttime temperatures by up to 28°F — what Abbott described as "the difference between Utah's semi-arid climate and the Sahara Desert." The 23-bombs framing has now appeared in The Salt Lake Tribune, Futurism, Rolling Stone, MSN, and across social media.

The unit-conversion problem

Andy Masley, who has written widely on data center energy accounting, did the back-math on the claim. Davies starts from a 16 GW thermal load (the contested figure, but let's accept it for the sake of argument). 16 GW × 24 hours = 384 GWh per day. Divide by 23 bombs and you get 16.7 GWh per "bomb," which works out to about 14.3 kilotons of TNT equivalent.

That is roughly the size of Little Boy, the bomb dropped on Hiroshima — and among the smallest nuclear weapons ever built. Most modern nuclear weapons are far larger. The largest bomb ever tested was about 50,000 kilotons. Masley notes that the largest hydrogen bomb in history released, in one second, about 150 times more thermal energy than this entire data center would emit over an entire day.

The Sun delivers about 174,000 terawatts of power to Earth continuously. The proposed 16 GW thermal load is roughly one one-millionth of one percent of that. To put it the other way: the Sun delivers more energy to Earth's surface in 90 seconds than Stratos would release over a decade of operation. The atomic-bomb framing is designed to feel alarming, not to inform.

This is a real pattern in environmental reporting on industrial facilities. Almost any large industrial process can be described in atomic-bomb-equivalents. A typical natural gas power plant releases similar waste heat as Stratos. A modern aluminum smelter, a steel mill, a refinery, a major airport's combined thermal output — all could be described in bomb-equivalents and all would yield similarly scary-sounding numbers, while none actually behaves anything like a nuclear weapon. Atomic bombs are dangerous because of blast pressure, prompt radiation, and radioactive fallout, none of which a data center produces.

The substantive heat question

Setting aside the rhetorical framing, large data centers do produce large amounts of waste heat. That heat has to go somewhere. The good news: the industry has spent the last decade figuring out how to capture and reuse it.

Production-scale heat recovery already exists

  • Stockholm Data Parks: 20+ data centers feed the city's district heating network through "Open District Heating." Today, recovered data center heat warms approximately 30,000 modern apartments in Stockholm and reduces grid emissions by ~50g CO2 per kilowatt-hour. The city's target is to meet 10% of total heating demand from data center waste heat by 2035.
  • Microsoft Espoo, Finland: Microsoft is building what it describes as the world's largest data center heat recovery scheme to heat the city of Espoo (Finland's second-largest city) plus two neighboring municipalities.
  • Orbis Data Center, Odense, Denmark: In partnership with Munters, exports 165,000 MWh of heat per year, supplying nearly 11,000 homes.
  • EcoDataCenter Falun, Sweden: Heat goes to fish farms and vegetable greenhouses, making industrial agriculture carbon-neutral or positive when co-located with the data center.
  • Mäntsälä, Finland: Nebius recovers about 20,000 MWh of energy annually, enough to heat 2,500 Finnish homes.
  • Eskilstuna, Sweden: Scandinavian Data Centers operates an integrated ecosystem of data center + battery storage + district heating with reuse built into the design from day one.

The mandates are starting

  • Germany: Now requires data center waste heat reuse
  • European Union: Energy Efficiency Directive includes heat reuse requirements for new builds
  • California: Title 24 has begun mandating waste heat recovery; other US states expected to follow

Why the Vera Rubin announcement matters for heat reuse

NVIDIA's Vera Rubin platform, entering production in late 2026, runs on 45°C / 113°F outlet water. That is exactly the temperature range that makes waste heat economically useful for industrial and agricultural reuse. The previous generation of data centers ran cooler — outlet water around 25-35°C — too cold to be useful for most downstream applications without further upgrading. The shift to warm-water cooling is, simultaneously, the shift that makes heat reuse practical at scale.

What's possible at Stratos specifically

Hansel Valley is rural agricultural land. The opportunities for productive heat reuse are real and economically credible:

  • Greenhouse agriculture — Utah has a short growing season; greenhouses powered by data center heat could meaningfully extend it
  • Aquaculture — fish farming is energy-intensive and matches data center heat output profiles
  • Industrial process heat for any future co-located manufacturing
  • District heating for any nearby development (Box Elder County has multiple small towns within reach)
  • Snow removal and de-icing on roads and rail in winter

Whether any of these get built at Stratos depends on the operators and economics of Phase 2 and beyond. The point is that "waste heat" is a 2015 framing. In 2026 and beyond, that heat is increasingly the next industrial input.

On the heat island and "Sahara" claim

Heat islands form primarily where large hard surfaces (asphalt, concrete, dense urban building masses) absorb daytime solar radiation and reradiate it at night, in still air with little vegetation to evapotranspirate. The classic urban heat island measures 1-7°F warmer than surrounding rural land.

Davies and Abbott's projection of 28°F nighttime warming and "Sahara-like" conditions extrapolates a maximum possible heat-island effect onto an open agricultural valley with extensive empty land, closed-loop cooling (which exhausts much less heat to the surrounding air than evaporative cooling), and no urban density. The number has not been published in any peer-reviewed climate model. Independent modeling has not been conducted. Anyone applying the standard heat-island literature to Hansel Valley's geography would arrive at substantially smaller numbers. The "Sahara" comparison is rhetorical, not modeled.

04.CAir quality, emissions, and what's coming behind Phase 1+

Phase 1, honestly

The Stratos site sits north of the Wasatch Front non-attainment zone that the EPA has designated for poor winter air quality. The site itself is in a geographically open valley with prevailing winds that generally carry away from the Wasatch population centers, but emissions do not respect lines drawn on maps.

Phase 1 will run on conventional natural gas generation, designed as "carbon-capture-ready." That phrase means the turbines are built with the physical and electrical infrastructure to add carbon capture equipment if and when economics permit, but they are not currently capturing carbon. Without capture equipment in operation, the Phase 1 campus will emit CO2 and NOx in proportion to its power production.

The Utah Division of Air Quality review is the venue for this concern and should be allowed to proceed rigorously. Anyone genuinely worried about air quality from the project should be filing public comments in that review.

What's coming behind Phase 1

The phased buildout means the technology mix at full operation will be substantially cleaner than what breaks ground in Phase 1. Several specific technologies matter.

The Allam-Fetvedt Cycle (zero-emission natural gas). Developed by 8 Rivers Capital and commercialized by NET Power (owned by Constellation Energy, Occidental, Baker Hughes, and 8 Rivers), the Allam-Fetvedt Cycle burns natural gas with pure oxygen instead of air and uses supercritical CO2 as the working fluid. The only byproducts are liquid water and a stream of high-purity, pipeline-ready CO2 — no NOx, no SOx, no particulates, no atmospheric CO2 release. The 50 MWth demonstration plant in La Porte, Texas, delivered first power to the ERCOT grid in 2021 and has been operational since. Commercial-scale 280-300 MW plants are now under development in Texas, Colorado, and Illinois. Projected efficiency: up to 59% at lower heating value. Projected capital costs: $900-$1,200/kW — competitive with conventional gas. This is the technology that lets a campus like Stratos retain the speed and reliability advantages of gas generation while eliminating atmospheric emissions entirely.

Small Modular Reactors (SMRs). NuScale, X-energy, Kairos, TerraPower, and Oklo all have designs in active NRC review. Factory-built nuclear modules of 50-300 MW, 60-80 year lifespans, zero emissions, ~50 acres of footprint, 95%+ capacity factor. First commercial operation expected 2028-2030 for first-of-a-kind units. Major hyperscalers have already signed PPAs and partnership agreements: Amazon with X-energy (5+ GW by 2039 at the Cascade Advanced Energy Facility), Google with Kairos (first reactor by 2030), Meta soliciting 4 GW of new nuclear, AWS-Talen Energy 1.92 GW PPA at the Susquehanna nuclear plant.

Enhanced Geothermal Systems (EGS). Fervo Energy's enhanced geothermal hit 3.5 MW of grid supply in Nevada in 2025, powering Google facilities through NV Energy. EGS is location-dependent but Utah has favorable geology in several parts of the state.

Carbon capture retrofits. The carbon-capture-ready turbine design at Stratos means capture equipment can be added without rebuilding the plant. Allam-Fetvedt is the more elegant solution because capture is intrinsic to the cycle, but capture-ready conventional turbines are a meaningful intermediate step.

The Phase 2 and Phase 3 generation mix at Stratos will not look like the Phase 1 mix. A campus that is 100% conventional gas at Phase 1 may be 40% Allam-Fetvedt zero-emission gas / 30% SMR / 20% geothermal / 10% conventional by Phase 3 — exact mix dependent on which technologies clear regulatory review on what timeline. Opposing the early phases means losing the platform that enables the cleaner later phases.

04.DThe "16 gigawatts" claim and what it leaves out+

Dr. Robert Davies, a physics professor at Utah State University, was quoted in Rolling Stone giving a preliminary estimate that Stratos would consume closer to 16 GW at full capacity due to natural gas inefficiency. He was reportedly quoted framing this as "the energy footprint of 40,000 Walmart supercenters." That number has now been repeated in coverage across the country.

One caveat up front

We do not know what Dr. Davies said in full. Rolling Stone is a publication with a clear editorial slant, and selective quoting on technical subjects is common when reporters are trying to make a point. He may well have included context that did not make it into the published piece. What follows is a response to the framing as it has been reported and amplified, not a critique of his complete analysis.

What the framing has right

Natural gas turbines do not convert fuel energy to electricity at 100% efficiency. Modern combined-cycle gas turbines achieve 50-60% thermal efficiency. Simple-cycle (peaker) turbines are closer to 35-40%. So if Stratos produces 9 GW of electricity at full buildout, the input fuel energy is indeed higher — roughly 14-18 GW of thermal input, depending on which turbine technology is used. The thermodynamics is real.

What the framing leaves out — four things

One: this same inefficiency applies to every gas plant on the grid. If Stratos drew its 9 GW from Utah's existing gas-and-coal grid instead of generating on-site, the upstream fuel consumption would be roughly the same. The thermodynamics is a property of fossil generation generally, not of Stratos specifically. A version of the same critique could be applied to every existing power user in the state. Framing it as if Stratos uniquely "wastes" energy is misleading.

Two: hyperscale facilities are far more efficient than older data centers. The metric here is Power Usage Effectiveness, or PUE. A PUE of 1.0 would mean every watt of incoming electricity went directly to compute. A PUE of 2.0 would mean half the incoming electricity was lost to cooling, lighting, and overhead. The state of the art in 2026:

  • Industry average (mixed enterprise data centers): ~1.56 PUE
  • Traditional enterprise data centers: 1.5 to 1.8 PUE
  • Modern hyperscale: 1.2 PUE or better
  • Best-in-class hyperscale (Google fleet-wide, Meta): 1.09 PUE
  • Direct liquid cooling / immersion (current frontier): 1.04 to 1.10 PUE

Stratos is being built with hybrid air-and-liquid cooling — a hyperscale design profile. A facility built to 2026 hyperscale standards is roughly twice as efficient per useful watt of compute as an enterprise data center from a decade ago.

Three: NVIDIA is actively rolling out 800-volt DC power architecture, and it changes the math significantly. This is the part of the story almost no one outside the industry has heard, and it is worth understanding.

Today's data centers run on AC (alternating current) electricity from the grid. That AC gets converted to DC (direct current) and stepped down through multiple voltage stages before it reaches the chip. Each conversion loses energy as heat. End-to-end efficiency of this traditional power chain is typically less than 90%. NVIDIA, working with Texas Instruments, STMicroelectronics, Schneider Electric, Flex, and others, is moving the industry to a new architecture: convert AC to 800 VDC once at the data center perimeter, then distribute 800 VDC directly to the racks. The result:

  • Multiple inefficient AC-to-DC and DC-to-DC conversion stages eliminated
  • Flex's GB300 NVL72 power shelf, designed for this architecture, hits 97.5% peak efficiency at half load
  • Conversion losses cut by 60% compared to conventional 54V designs
  • Roughly 200 kg of copper per cabinet eliminated, reducing material costs and waste heat
  • NVIDIA estimates ~5% end-to-end efficiency gain at the system level — which at gigawatt scale is enormous

NVIDIA published the 800 VDC reference architecture at the OCP Global Summit in October 2025 and showcased the full ecosystem at NVIDIA GTC in March 2026. The upcoming Kyber rack architecture is designed around 800 VDC from day one. This is not theoretical — it is shipping product being deployed in new builds.

For a phased project like Stratos, this matters enormously. Phase 1 will likely deploy with a mix of current and emerging architectures. Phase 2 and beyond will be native 800 VDC, meaning the same gas turbines will produce the same electricity but the useful compute extracted per watt at the chip will be meaningfully higher.

Four: chip efficiency keeps improving, and each phase of Stratos uses newer hardware. Per-GPU performance per watt has improved dramatically with each chip generation. NVIDIA's Blackwell architecture (current frontier) delivers roughly 25 times the inference performance of the previous Hopper generation at similar power. Amazon's Trainium2, which the company highlighted in Q1 2026 earnings, delivers about 30% better price-performance than comparable GPUs. The next-generation Vera Rubin architecture is expected to push these gains further.

This matters specifically for a phased buildout. Phase 1 of Stratos (2026-2028) will deploy with the chips available in that window. Phase 2 (2028-2030) will use whatever exists then — likely 2 to 3 times more efficient per watt than Phase 1. Phase 3 and beyond may use chip architectures, cooling approaches, or even power sources that have not been invented yet. The same wattage at the end of the buildout will produce vastly more useful compute than the same wattage at the start.

The "40,000 Walmarts" comparison

This phrase has done a lot of rhetorical work in opposition coverage. It is true that 16 GW is roughly the average electricity demand of 40,000 Walmart supercenters. It is also a comparison that tells you nothing about whether the electricity is producing something worth producing. By the same logic, every steel mill, every refinery, every aluminum smelter, every hospital complex, every airport could be described in Walmart-equivalents. The comparison is designed to feel large and wasteful. It is not designed to inform.

What an honest engagement with the energy concern looks like

  • The campus will use a lot of energy. This is true and inherent to the project.
  • It will be more efficient per unit of compute than older data center designs.
  • It will produce carbon and NOx emissions in proportion to the fuel it burns.
  • The phased buildout means later phases will be much more efficient than earlier ones — by chip generation, by power architecture, and by cooling design.
  • The 800 VDC architecture transition is actively rolling out across the industry and will cut conversion losses dramatically.
  • Carbon capture-ready turbine design means the option exists to add capture equipment when the economics permit.
  • The alternative is not zero energy use. It is that the same compute gets built somewhere else under weaker environmental rules.

The energy concern is legitimate. The "16 GW = 40,000 Walmarts" framing as it has been amplified is rhetorical packaging that obscures more than it reveals.

04.EThe military connection and the "domestic surveillance" claim+

What's actually being claimed

A particular branch of the opposition argues that Stratos is fundamentally a surveillance project — that the MIDA framework and the Air Force connection are cover for what is really mass surveillance of US citizens. The most explicit version appears in a Books Behind Borders piece titled "The Biggest Data Center In History Just Got Approved," which describes data centers like Stratos as places that "exist to collect data, process data, track behavior, monitor activity, train AI systems, and store an almost incomprehensible amount of information about human beings." NaturalNews and SHTF Plan have published similar framings, treating Stratos as a "betrayal of democracy." On social media, the language is more direct: this is the new NSA, this is mass spying, this is Bluffdale 2.0.

The claim deserves to be evaluated on its merits, not waved off. Bulk surveillance is a real thing that real governments do, including the US government. The NSA's Utah Data Center at Bluffdale exists. The Snowden disclosures happened. A Greenpeace blimp with "NSA Illegal Spying Below" painted on it flew over Bluffdale in 2014. So the question is fair: is Stratos that, just bigger?

The numbers don't support the claim

The NSA Utah Data Center at Bluffdale — the actual purpose-built signals-intelligence facility — runs at roughly 100 to 150 megawatts of total power on about 1.5 million square feet of building space. It has been operational since around 2014. Whatever bulk collection and analysis the US intelligence community does on US soil, this is the facility doing it.

Stratos at full buildout is 7,500 to 9,000 megawatts. That is 50 to 90 times larger than Bluffdale. For perspective:

  • The entire annual peak electricity demand of the state of Utah is approximately 4,000 megawatts. Stratos at full buildout is roughly twice that.
  • The total combined power of every existing NSA, FBI, and intelligence community data center in the United States is well under 1,000 MW.
  • One Phase 1 alone at Stratos (1.5 GW) is already 10x larger than Bluffdale.

If the actual goal were domestic surveillance, the appropriate facility would be another Bluffdale-sized site at 100-150 MW. Building 50-90x more capacity than the entire existing US signals-intelligence apparatus combined would be an absurd way to scale a surveillance program.

Surveillance doesn't need US soil

The other half of the rebuttal: even if the US government wanted to expand bulk surveillance of Americans, building a giant facility in northern Utah would be a strange way to do it. The Five Eyes intelligence partnership — US, UK, Canada, Australia, New Zealand — already conducts signals collection at allied facilities outside the United States. The legal architecture for sharing collection on US persons through allied partners has existed since at least the UKUSA Agreement in 1946. Additional SIGINT stations operate at overseas military installations and partner-country facilities. US-based collection (under FISA Section 702, Executive Order 12333, and related authorities) already happens through existing infrastructure that includes Bluffdale, fiber tap points at major internet exchanges, and cooperation from US telecommunications carriers.

None of that requires a new 9 GW campus in Utah. None of it benefits significantly from one. The compute profile is mismatched: bulk metadata and traffic analysis is a storage-heavy, networking-heavy, moderately-compute-heavy workload. Frontier AI model training is a GPU-heavy, power-heavy, cooling-intensive workload. The two have different infrastructure shapes. Stratos is being designed around the second profile, not the first.

What the actual DoD connection is

The record on the military connection is well-documented and worth being honest about:

  • A senior Air Force official formally requested that MIDA identify locations to expand military computing capacity
  • O'Leary Digital's Stratos proposal was structured as a response to that solicitation
  • Stratos is geographically near Hill Air Force Base, Northrop Grumman Promontory, and the Utah Test and Training Range — actual existing defense infrastructure
  • MIDA's enabling statute allows capture of revenue from project areas to flow back to military infrastructure
  • The bulk of compute at the campus, in practice, will serve commercial customers (hyperscalers, AI labs, enterprise) — not direct DoD workloads

What the DoD actually needs gigawatt-scale GPU compute for is the stuff that has been openly discussed in JADC2 (Joint All-Domain Command and Control) documentation, DoD AI strategy publications, and DARPA programs: training pattern-recognition models on satellite and reconnaissance imagery, logistics and supply chain optimization, cybersecurity defense, autonomous systems development, and modeling and simulation. These are AI training and inference workloads in the same shape as commercial AI work. They benefit from a commercial-grade hyperscale campus because it is the same architecture.

The MIDA structure

The Stratos Project Area is governed by Utah's Military Installation Development Authority. MIDA is a state authority created to support facilities tied to military readiness. Its enabling statute allows it to designate project areas, negotiate development agreements, and capture revenue that flows back to military infrastructure (in this case, Hill Air Force Base and the Utah National Guard). Whether the MIDA structure is the right framework for a project like Stratos is a legitimate policy question — and one Governor Cox responded to by announcing that the project will be approved one phase at a time, with public review at each phase. That is the system working.

The honest read

Critics arguing the DoD connection is "real" are right. Critics arguing this is fundamentally a domestic surveillance facility are not. The infrastructure is the wrong shape, the wrong size, in the wrong place, and points toward a totally different use case. People are entitled to oppose Stratos on grounds of water, air, energy use, MIDA scope, or general skepticism about industrial scale. Those are real conversations. The surveillance framing collapses on contact with the actual specifications of the facility.

06.AThe federal layer: Sanders, AOC, and the data-center moratorium bill+

The Artificial Intelligence Data Center Moratorium Act of 2026

Introduced March 25, 2026, by Senator Bernie Sanders (I-VT) and Representative Alexandria Ocasio-Cortez (D-NY14). The bill would impose a nationwide halt on the construction or upgrading of AI data centers until Congress passes federal legislation regulating AI. It would also ban exports of advanced AI chips to countries lacking equivalent AI regulations.

What the bill says, in the sponsors' own words

From Sanders's introductory statement: "we cannot sit back and allow a handful of billionaire Big Tech oligarchs to make decisions that will reshape our economy, our democracy and the future of humanity. We need serious public debate and democratic oversight over this enormously consequential issue. The time for action is now. We need a federal moratorium on AI data centers."

From Ocasio-Cortez's: "Congress has a moral obligation to stand with the American people and stop the expansion of these data centers until we have a framework to adequately address the existential harm AI poses to our society. We must choose humanity over profit."

What it would actually do

If passed, the bill would freeze Stratos and every other proposed AI data center in the United States. The Center for Data Innovation, a centrist policy think tank, has pointed out that the bill cites several rationales (AI existential risk, electricity costs, jobs) that do not actually require a data center moratorium to address — and that the moratorium framing was chosen first, with rationales assembled to support it. Senator John Fetterman called the bill a "China First" policy, arguing that halting US AI infrastructure construction while China continues its own buildout cedes the field to Beijing.

Why it matters for Stratos even though it won't pass

The bill is unlikely to clear either chamber of this Congress. That is not the point. The point of the bill is to establish "AI data center moratorium" as a coherent legislative position with the imprimatur of a sitting US Senator and a high-profile House member. Once that position exists in federal legislation, it can be cited by state legislators, county commissioners, opposition groups, and journalists as a legitimate, mainstream, federally-endorsed framework. The Sanders-AOC bill is the strategic anchor — local opposition like the Box Elder fight is the tactical execution.

06.BThe money trail: Singham, Wyss, and the funded opposition network+

Two foreign-resident funding sources keep appearing in the documented financing of US AI-infrastructure opposition. Both are openly disclosed in investigative reporting and government filings. Neither is a conspiracy theory.

Neville Roy Singham — Shanghai-based American

Singham is a 72-year-old American tech entrepreneur who sold his consulting company ThoughtWorks for approximately $785 million in 2017 and subsequently relocated to Shanghai. He describes himself as a socialist and has stated his admiration for Maoist political traditions. From 2001 to 2008 he worked as a consultant for Huawei, the Chinese telecommunications company at the center of the US-China tech-export-controls dispute.

In August 2023, the New York Times published an investigation titled "How a U.S. Tech Mogul Used Nonprofits to Sow Chinese Propaganda." Reporters tracked hundreds of millions of dollars from Singham-linked entities to a global network of media outlets and advocacy organizations that, in the Times's documentation, "mix progressive advocacy with Chinese government talking points." Singham personally attended a Chinese Communist Party workshop on promoting the party internationally, according to NYT reporting. He has denied acting at the direction of the Chinese government.

Organizations directly tied to the Singham funding network include:

  • Peoples Dispatch — the international socialist publication whose framing of Stratos as a fight over "who controls the infrastructure behind the AI economy" has been reproduced across opposition coverage
  • The People's Forum — Manhattan-based, publicly acknowledged receipt of over $20 million from Singham and his wife Jodie Evans (co-founder of CODEPINK) between 2017 and 2022 through shell companies and donor-advised funds, per a September 2025 House Ways and Means Committee investigation
  • BreakThrough News — produces video and social-media content critical of US foreign policy and infrastructure projects
  • No Cold War — anti-US-foreign-policy organization the NYT linked directly to Singham's network
  • CODEPINK — co-founded by Jodie Evans, Singham's wife
  • Several Party for Socialism and Liberation (PSL)-aligned media projects

These are some of the outlets producing the framings that get picked up downstream. When a Hansel Valley resident shares a Peoples Dispatch article framing Stratos as capitalist consolidation, they are sharing content from a network that has documented funding ties to an American expatriate living in Shanghai with publicly reported connections to Chinese state media efforts. That is not the same as saying the resident is being paid, or that the resident is doing anything wrong — they are passing along a frame that sounded right to them. The frame's provenance is, however, worth knowing.

Hansjörg Wyss — Swiss billionaire

Wyss is a Swiss businessman whose American spending vehicle, the Berger Action Fund, was profiled by the Associated Press as "a nondescript name for a group with a rather specific purpose: steering the wealth of Hansjörg Wyss, a Swiss billionaire, into the world of American politics and policy." As a foreign national, Wyss is prohibited by US law from donating to candidates or political committees. He is permitted to fund 501(c)(4) "dark money" organizations not required to disclose donors.

The Berger Action Fund has donated tens of millions to the Sixteen Thirty Fund, which acts as a fiscal sponsor and donor-routing vehicle for many of the largest progressive advocacy groups in the United States. An April 2026 American Energy Institute report identified twelve organizations actively opposing US data-center development that have collectively received more than $39 million in funding from foreign donors based in Switzerland, the United Kingdom, and Denmark — with Wyss as the most prominent named donor. Top recipients include Indivisible, 350.org, Oil Change International, GAIA, the Sierra Club, and Americans for Financial Reform.

What this means for Box Elder County

The national organizations farm funding out to local affiliates — "Stop The Data Center Coming To Martindale Brightwood" in Indianapolis is one named example. The same playbook is run in every county where a new data center is proposed: national group provides staff, talking points, and resources to local affiliates; local affiliates appear at county hearings; coverage in Peoples Dispatch, BreakThrough News, and aligned outlets amplifies the story; the Sanders-AOC bill provides federal legitimacy.

To be very clear about what this is and is not: the residents of Hansel Valley showing up at county commission meetings are sincere. They are not on a payroll. They are not being paid by Singham or Wyss or anyone else, and nothing above should be read to suggest they are. Sierra Club has been engaged in environmental work for over a century, and individual Sierra Club members in Utah are mostly doing what they believe is right. The local Stratos opposition is, at the individual level, made up overwhelmingly of people acting in good faith on the basis of the information they have been given.

The reason this funding architecture matters is that the information they have been given was authored upstream — at the level of Peoples Dispatch articles, BreakThrough News videos, 350.org talking points, Sanders-AOC press releases, and the moratorium-bill language itself. That upstream content is what is funded. The sincere local resident shares a Peoples Dispatch piece because the argument sounds compelling, not because anyone paid them to share it. But the argument sounded compelling in part because it was workshopped, refined, and amplified through a well-resourced infrastructure that the sincere local resident has no visibility into.

This pattern — sincere people repeating top-down-funded messaging — is not unique to this fight or to one political direction. The same dynamic happens on the right with industry-funded skeptic groups, and on the left with foundation-funded advocacy. Naming it on either side is not an accusation against the people repeating the messaging. It is an attempt to make visible the part of the conversation that usually stays invisible.

If Stratos is a contested question about water and energy and the right way to allocate large infrastructure, that is a legitimate American argument that should happen on the substance. The substance review is happening — Governor Cox has committed to phased approval with public review at each phase. That is the right venue. The funding-source disclosure here is meant to add context, not to dismiss the people showing up.

06.CThe local layer — state parties, DSA, PSL, Peoples Dispatch+

At the bottom of the funnel — the layer that reaches Box Elder County residents directly — the framing comes through a specific set of channels. Most local Stratos opponents are sincere residents. This section is about the organized layer above them and the messaging being delivered into the local conversation.

The publications

  • Peoples Dispatch — published the foundational framing of Stratos as "a broader political struggle over who controls the infrastructure behind the AI economy." Peoples Dispatch is part of the Singham-funded international socialist media network (see 06.B). Its coverage of US infrastructure projects consistently follows the frame of capitalist consolidation versus working-class resistance.
  • Rebooting Democracy — published critiques of the Air Force AI land lease program, framing AI compute buildout as private capture of public resources
  • BreakThrough News, No Cold War — produce shorter-form social content amplifying the same frames into TikTok, Instagram, and X feeds
  • NaturalNews, SHTF Plan, Books Behind Borders — running a parallel surveillance-and-conspiracy frame (see 04.E on the surveillance claim) that overlaps with the socialist framing where both can attack the project

The organizations

  • Democratic Socialists of America (DSA) — Utah chapters have been active in the Stratos opposition. DSA's broader position is anti-private-capital infrastructure development
  • Party for Socialism and Liberation (PSL) — published organizing materials on AI infrastructure as a working-class issue. PSL is closer to the Singham-funded media network than DSA is, with shared messaging across publications
  • Indivisible, 350.org, Sierra Club Utah chapter — provide the environmental framing that fits more comfortably in mainstream local coverage. All three receive funding from the Wyss-network described in 06.B
  • State Democratic Party operations — several state parties have adopted anti-AI-infrastructure planks in 2026, providing electoral cover and infrastructure for local organizing

Why the substantive concerns will always be a moving target

The organized opposition's objection to Stratos is not really about water or air. A version of Stratos that solved every environmental concern would still be opposed by these groups, because their actual objection is to large-scale private infrastructure development as a category — and, for the Singham-network publications specifically, the additional strategic goal of slowing American AI capability development as a category. This is worth understanding because it means engaging the environmental substance — which is the right thing to do — will not resolve the political opposition. That opposition has a different goal.

06.DThe political map on AI infrastructure+

The politics of AI infrastructure do not map neatly onto traditional party lines, which makes the debate confusing for people who follow it through partisan media.

The pro-build coalition includes

  • National security hawks of both parties
  • Mainstream Republican governors (including Utah's Spencer Cox)
  • Most of the technology industry
  • Trade unions in construction and skilled trades
  • The defense establishment
  • A significant slice of mainstream Democrats (Senator Fetterman explicitly opposed the Sanders-AOC moratorium as "China First")

The anti-build coalition includes

  • The progressive left (Sanders, AOC, the Squad) at the federal level
  • Senator Warren and Markey on state-AI-regulation preemption specifically
  • Environmental NGOs concerned about water and emissions
  • Local NIMBY opposition (genuinely sincere, often justifiably so)
  • The democratic-socialist organizational infrastructure (DSA, PSL, allied publications)
  • The Singham-funded international socialist media network
  • Some libertarian voices concerned about MIDA-style public-private structures
  • A surveillance-and-conspiracy wing on the populist right (NaturalNews, SHTF Plan, parts of MAGA-adjacent commentary) opposed for unrelated reasons

What this means for Stratos

The coalition opposing Stratos is heterogeneous and includes participants with directly opposing worldviews — a Marxist organizer and a sovereign-citizen anti-surveillance blogger have nothing in common politically, but both end up at the same county commission hearing chanting against the same project. That heterogeneity is itself a sign that the opposition is being coordinated upstream: organic local opposition does not usually unite the populist right and the international socialist left.

For an individual Box Elder resident, the implication is straightforward. A rancher worried about her well is not in the same political position as a DSA organizer running anti-capitalist messaging, or a Books Behind Borders blogger running surveillance-state messaging. All three are visible in the opposition, but they have different goals and would settle for different outcomes. The rancher might be satisfied by water guarantees, real review, and phased approval. The organizers will not be.

This is the reason a pro-Stratos argument has to address the substantive concerns directly while also naming the political projects being run through them. Otherwise the substantive concerns become a perpetual moving target — and the rancher's legitimate concern gets used as cover for fights about something else entirely.

Share this with someone making up their mind.

The conversation gets better when more people show up with the full picture. Sources, data, and the stories behind the numbers are in the deep sections below.

Read the deep dive Back to top