Most analysis of quantum computing and data center power ends in the same place. Quantum machines might use more electricity than classical computers, because they need extreme cooling and heavy error correction to keep working. They might use less, because some problems that take classical machines millions of operations could take quantum machines a handful. Nobody knows, so plan for both.

That is a fair conclusion for an explainer. It is not enough for anyone making capital decisions now.

I published a piece this week that takes the question one step further. The short version: quantum does not escape the power constraint facing data centers. It inherits it.

The constraint is more specific than a shortage of electricity. The technologies that can supply clean power around the clock largely exist or are close. What slows them down is validation time. A data center needs 99.999 percent availability, known as five nines, which allows about five minutes of unplanned downtime a year. Before an operator will stake a facility on a new power system, that system has to be permitted, connected to the grid, brought into service, and run long enough to build a track record. That takes years. None of it gets faster because the computing inside the building changes.

The strongest counterargument is substitution. If quantum takes over tasks where classical computers are very inefficient, such as simulating molecules, total computing energy could fall. I grant that case in full. It still arrives in the 2030s at the earliest, after the buildout that is straining grids today. A cheaper calculation does nothing to shorten a grid connection.

This is one instance of what I call the Temporal Trilemma. The energy trilemma is the long-standing tradeoff between security, affordability, and sustainability. What has changed is the rate. Demand now arrives on a cadence of quarters. The system that has to supply it plans in multi-year cycles. When the two do not match, operators install gas turbines on site to meet their dates, and decarbonization goes backward.

The full piece covers the demand projections, the quantum timeline, and where the Structured Transition Model fits. [Read the full article]

Two developments would change my view. The first is fault-tolerant quantum machines, meaning systems that run reliably despite constant hardware errors, reaching large-scale deployment well ahead of current roadmaps. The second is a real reduction in how long it takes to validate firm clean power. The second is the one to watch, because it is the one the industry can influence.

Fewer than one in ten data centers in Texas had complied with a state law requiring them to report how much power and water they use.

Subscribe now

That fact, more than any poll or protest, is what made this month’s intervention possible.

On August 3 the Governor of Texas wrote to the state’s utility regulator and to the operator of its grid, instructing them to audit every data center moving through the connection process before any further project advances. Projects that fail are to be refused connection. Eleven days later, officials put a number on the scope: roughly 250 to 300 projects, representing about 200 gigawatts of future demand. For scale, that is more than twice the largest amount of electricity the Texas grid has ever delivered at one time.

Most of the coverage has read this as a response to public anger. I think that misreads the mechanism, and the difference matters to anyone underwriting these projects.

Public anger is real and it is increasing. But anger is not a legal instrument. It cannot pause a connection queue. What paused the queue was a reporting obligation the sector had largely ignored, and a table showing who had ignored it. The state did not need to win an argument about whether data centers are good for Texas. It needed a compliance rate.

That is a far more durable basis for action than public mood, because compliance data does not decay with the news cycle.

The part that should worry developers more

Texas is building three ultra-high-voltage import corridors into the Permian Basin. They are driven substantially by the electrification of oil and gas operations, with regional demand projected toward roughly 26 gigawatts by 2038. Grid officials have testified that without them the region faces rolling blackouts as early as next summer.

Data centers did not cause that plan. They are now, in part, blocking it.

After a fifteen-hour hearing in late July dominated by landowner testimony, the Lieutenant Governor called on the regulator to refuse the certificates those lines require until the legislature has revised the approval process. Opposition generated by one class of infrastructure has attached itself to another that the state needs regardless.

That is the thing worth noticing. Not that the sector is unpopular. That its unpopularity has become transferable.

What I think this changes

A connection date used to be an engineering output. Queue position, study cycles, transformer lead times. All of it modellable with reasonable confidence.

In Texas today, a connection date is a function of an audit with no completion deadline, a rulemaking due to conclude in December, and a legislature that convenes in January. None of those are engineering variables, and none of them respond to paying more.

Anyone still building a Texas schedule from queue mechanics is pricing the wrong distribution.

There is a reasonable case against reading it this way, and it deserves a proper answer rather than a dismissal. I have set out both the argument and the objection in full on my own site.

Read the full piece: Texas Turned Public Opposition Into Procedure

Five Nines and Fast Power - Making Better Decisions in AI Energy Investments - Read the Book

There is a serious case circulating that the data center industry does not need to build its own power. It deserves a serious answer. I published one this week. This is the short version.

Subscribe now

The setup first, because the argument is good.

In February 2026, EPRI announced a collaboration with Nvidia, Prologis and InfraPartners to site small data centers, 5 to 20MW each, at electrical substations that already have unused capacity. There are roughly 55,000 substations in the United States. Spare capacity of 5 or 10 or 20MW at each of them adds up quickly. Duke University’s Nicholas Institute found that the largest 22 balancing authorities could absorb around 76 gigawatts of new load if that load accepted curtailment averaging 0.25 percent of its annual consumption. At 1.0 percent, the figure reaches 126 gigawatts.

The reason is structural. The US power system was built to serve a peak it rarely touches. Generation capacity runs at around 53 percent utilization on average. The highest-demand hours happen for fewer than 200 hours a year.

So there is a large quantity of underused capacity sitting inside the existing distribution network. That is correct. The industry has spent three years talking about new generation while under-using the flexible assets already installed. Anyone dismissing this as a distraction is wrong.

My problem is not the pilot. It is what the pilot is being used to prove.

The percentage and the hours are different numbers

EPRI’s micro data center lead has said workloads in the pilot would need to move to a different substation only about 0.1 percent of the time. That sounds trivial. It is roughly nine hours a year.

The same reporting notes that peak demand periods run to just under 200 hours a year. Those two figures are more than twenty times apart and they describe the same grid.

The Duke work explains why. The headline curtailment rate is measured in energy, not in hours. A site asked to shed 20 percent of its load for four hours registers as a very small energy percentage. It is still a four-hour operational event. At the 0.25 percent energy figure, Duke found some degree of curtailment occurring across about 85 hours a year. At 0.5 percent, around 177 hours.

That is the number an operator needs before signing a flexible interconnection agreement. It is not the number in circulation.

Three things nobody has published

Headroom is a stock, not a flow. Each megawatt of spare distribution capacity gets allocated once. Every assumption behind the 0.1 percent figure is a first-mover assumption, calculated against a network where nobody else is doing this yet. If the model works, it gets copied, and the copying consumes the conditions that made it work. What do the terms look like for the second and third tenant at the same substation?

Peak non-coincidence is the load-bearing assumption. Distributing 25 sites across five utilities delivers resilience only if those five utilities experience stress at different times. Uri in 2021 and Elliott in 2022 both produced simultaneous stress across multiple balancing authorities. Extreme heat behaves the same way. The events that actually threaten a data center’s power supply are the events most likely to be correlated across wide geography. The correlation analysis has not been published.

It solves for the load that was never the constraint. Inference is one of very few workloads that can be dynamically routed. The pilot is designed for inference, and that is sound engineering. But the interconnection queues and the multi-year utility timelines that started this entire debate are driven by training campuses drawing hundreds of megawatts at a single location. The substation model does not touch them.

What the argument quietly concedes

This is the part that has gone unremarked.

Duke names three mechanisms by which a large load delivers flexibility: shift the workload, reduce operations, or run onsite generation. One of the report’s authors has said directly that the growth of onsite generation and storage is part of why this approach is viable now, because a curtailed facility does not have to go dark.

So the substation model does not remove onsite generation from the system. It changes what onsite generation is for.

In the conventional framing, generation on site exists because the grid cannot deliver enough power fast enough. In the flexible interconnection framing, generation on site exists because it is the instrument that makes the flexibility commitment credible. Without it, curtailment means lost revenue and broken SLAs. With it, curtailment is a fuel switch.

The asset moves from being the primary supply to being the thing that makes the contract signable. That makes it more strategically important, not less.

The technical debate looks like it is about capacity. It is actually about which side of the meter the dispatch decision sits on. Taking spare substation capacity means accepting that somebody else decides when your load comes down, in exchange for connecting years earlier. That is a rational trade. It is also a real transfer of control, and it should be priced as one.

My expectation is that these sites get built, that they work, that they are genuinely useful for inference at the edge, and that a meaningful proportion end up with generation or storage behind them anyway. Not as a fallback. As the reason the interconnection agreement could be signed at all.

Spare grid capacity is real. Firmness is a separate product, and somebody still has to supply it.

The full article sets out the four disclosures that would change my mind, and why the first three tell you whether this is a durable model or a first-mover arbitrage.

Read the full analysis: What the Substation Headroom Argument Leaves Out

This piece first appeared at alexmarshallenergy.com.

Explore the Book - Five Nines and Fast Power - Making Better Decisions in AI Data Center Investments. Available on Amazon.

A fair challenge keeps coming back to me about the Structured Transition Model (STM): the individual pieces, onsite gas generation, combined heat and power (CHP), microgrids, grid import, storage, aren’t new. Industrial and utility power has been sequencing these technologies for decades. So where’s the actual contribution? It’s in the order, not the parts.

Subscribe now

A power asset built today runs for decades. Over that life it will see multiple changes of federal and state government, and each one can shift the economics: incentives appear and expire, carbon rules tighten and loosen. The sequence a developer commits to on day one decides how well the asset absorbs those swings, or whether it gets stranded by them. Most buyers are shown a configuration for today. They’re rarely shown how that configuration ages.

This gap exists because two markets haven’t met yet. Utility and industrial power has treated this sequencing as routine work for years. Data center developers are now arriving at the same decisions from a different direction, at a different speed, and largely in a different room. Today, data center development optimizes for two things: reliability (the “five nines” standard of 99.999 percent uptime) and speed-to-power (how fast a site can be energized). Decarbonization gets treated as a later problem, which means the sequence chosen at the start can turn grid constraints and policy shifts into a trap rather than a plan, usually after the site is already built.

I’ve been in this specific conversation, data center power as the binding constraint, since 2016, well before it became a consensus industry topic. That tenure is what STM is built on: not a claim to have invented the underlying tools, but a framework for applying a known order to a market that hasn’t been handed one yet.

So the honest answer to “what’s new” is: the components aren’t. The order, applied to data center power, for a buyer facing decisions that will outlast several policy cycles, is where the value sits.

The full piece, including the case for why this hasn’t already been done, is on [alexmarshallenergy.com].

Subscribe now

For most of the past twenty years, power planning across the developed world rested on a single assumption: the grid would be there. Backup generation was insurance against a rare failure. Distributed generation, meaning power produced at or near the site where it is consumed, was treated as a specialist answer for hospitals, islands, mines, and countries where the central network had not been built out.

Subscribe now

That assumption is now under strain in precisely the markets that held it most confidently.

Interconnection queues in parts of the United States now run for years. Transmission is congested. Reserve margins are tightening. Weather-driven disruption is more frequent. The result is that the conversation in digital infrastructure has moved from how to back up power to how to secure power at all.

None of this is new. It is just new here.

What weak-grid markets already solved

I have spent a lot of my career around power systems in places where the grid was never the default. Remote mining operations in Australia. Industrial operators across Africa. Emerging systems in places like Papua New Guinea. In those environments, resilience was never a design preference. It was the only way the asset ran.

The engineering priorities that fall out of that constraint are consistent:

• Uptime treated as a commercial input, not a technical target

• Modular capacity that can be added in stages rather than in one build

• Fuel logistics and maintainability weighted as heavily as efficiency

• Hybrid configurations, because heavy continuous loads and intermittent renewable output do not match without firm generation in the system

• Infrastructure designed to evolve, because nothing gets replaced wholesale

Read that list again with a 500 megawatt AI campus in mind. It is the same list.

Reliability is decentralizing

The more consequential shift is not technical. It is about who owns reliability.

Utilities used to own it. Businesses consumed electricity and held backup systems for the rare occasion the utility failed. That boundary is dissolving. Hyperscale developers and large industrial operators are now building their own reliability layers out of onsite generation, storage, microgrid controls, thermal integration, and fuel flexibility.

This is not a rejection of the grid. It is a recognition that the grid is becoming one component of operational certainty rather than the whole of it.

That recognition changes the strategic question. It stops being what fuel a facility uses and becomes which parts of its reliability the operator is willing to leave outside its own control. Transition, in practice, is sequenced infrastructure evolution under reliability and speed constraints, not fuel substitution on a schedule.

Why this matters for the next three years

The data center sector is not inventing distributed energy. It is importing a set of assumptions that industrial operators in constrained markets have worked under for decades, and it is importing them at a scale and speed that industry has never had to handle.

The operators who move fastest will be the ones who stop treating onsite generation as a contingency line item and start treating it as part of the asset.

Full piece on the site:

What Remote Infrastructure Teaches Us About Resilience

If you are working on power procurement for a large digital infrastructure project right now, I would be interested to know which of these constraints you are hitting first: interconnection timing, transmission capacity, or equipment lead times.

Subscribe now

Five Nines and Fast Power

Making Better Decisions in AI Power Investments

Explore the Book

A few years back, something I wrote on LinkedIn about fuel measurement got picked up by a trade publication. Not a funding round. Not a new engine. A distinction between two ways of measuring the energy content of natural gas, the kind of thing that normally lives three layers deep in an engineering appendix.

That it got attention at all is the interesting part.

Subscribe now

The definitions have not changed. Gas engines have always been rated one way, gas suppliers have always billed another way, and for twenty years that gap was a reconciliation problem handled quietly by specialists. Nobody outside the engineering team needed to think about it.

Onsite generation at data center scale changes that. A few points of apparent efficiency, multiplied across a multi-megawatt load running thousands of hours a year, turns a footnote into a line item. And the mistake that follows is invisible by design: when one technology gets measured on one basis and another gets measured on a different one, the comparison looks clean and is quietly wrong. Opex models get built on the wrong number. Efficiency claims drift from what the plant actually delivers. Capital moves toward the wrong answer, and nothing in the spreadsheet flags it.

I think this is the shape of a broader pattern, not a one-off. As more compute moves behind the meter, a whole layer of formerly invisible engineering convention, gas flow sizing, pressure regulation, redundancy math, is surfacing into commercial visibility. People who used to delegate these questions entirely are now the ones signing off on them.

The full piece walks through why the gap exists, what it does to a project’s economics when it gets missed, and why the next phase of competition in onsite power is going to reward whoever models it honestly.

Read it here: [link]

Subscribe now

The power conversation around AI data centers has collapsed into a procurement question. Which source is cheapest per megawatt-hour? Which is cleanest? Reasonable questions. Wrong starting point.

Subscribe now

The decision a developer actually faces is a capital allocation problem under a time constraint. The variable that determines the return on a multi-billion dollar build is not the unit cost of electricity. It is whether the facility can be energized when the revenue is there to be captured.

The numbers make the problem concrete. In major data center hubs, grid interconnection now runs four to seven years from request to energized site. Nearly 2,300 gigawatts of generation and storage sit in US queues, more than the entire installed capacity of the country. Meanwhile the IEA projects global data center electricity consumption to roughly double by 2030. Capital is ready. The grid is not. The gap between the two is where returns are made or destroyed.

In my latest piece, I work through two errors that both kill value. The first is obvious: a facility built and idle, waiting for grid power, while the interest clock runs and contracted demand goes to whoever energizes first. The second looks like prudence: waiting for the perfect energy stack, the cleaner source, the firmer connection. A flawless facility energized in 2031 may have missed the demand it was built to serve.

The piece also makes the case for two financial properties of onsite generation that never appear in a levelized cost calculation:

Time arbitrage. Onsite power does not compete with wholesale generation. It competes with the delivered cost of firm power at a specific site, in five years, and at a specific point in time. On project IRR, the premium is often the cheaper option, because the alternative is no revenue at all.

Optionality. An energy decision made today does not have to be the decision for the life of the asset. A site built to take gas now, storage as it becomes economic, and a grid connection when it arrives keeps the ability to adapt as prices, policy, and technology move. That flexibility has a financial value, and a least-cost-today decision that forecloses it is more expensive than it looks.

The logic has limits, and I state them. This is not an argument that onsite generation always wins. It is an argument that the decision belongs in a capital allocation frame, judged on time and optionality, not a procurement frame judged on unit cost alone.

Read the full piece here:

Speed to Power vs. Cost of Mistake: The New Capital Allocation Problem in AI Infrastructure

Subscribe now

For most of the past decade, efficiency in digital infrastructure has been argued on two grounds: operating cost and emissions. Both are real. Both now understate what efficiency does.

Subscribe now

The expansion of AI infrastructure has changed the terms. Speed-to-power, meaning how fast a site can be energized and brought into service, has become the constraint that governs everything else. Grid interconnection queues, transmission reinforcement, equipment lead times, and permitting all sit between a developer and a live facility. In that setting, efficiency stops being an operating metric and becomes something else. It becomes a way to need less of the infrastructure that is hardest to get.

That is the argument I want to put plainly. A facility that extracts more usable output from the same energy input requires less generation, less cooling, and less grid dependency. It draws less on the interconnection that is already the bottleneck. In a constrained market that shortens the deployment timeline directly. Efficiency compresses the amount of infrastructure that has to be built at all.

There is a second point that matters more than the first, and it is about time rather than cost. Infrastructure built quickly under deployment pressure tends to stay in place for decades. The generation type, the cooling architecture, and the fuel strategy chosen to energize a site fast will shape its operating flexibility and its emissions path for the life of the asset. The decision is made in weeks. The consequence runs for twenty years.

This is why I do not think the data center power question resolves to a single technology. It resolves to sequencing. Onsite generation, grid interaction, storage, thermal integration, and staged decarbonization, ordered so that the fast decision does not foreclose the adaptable one. Not as an ideological preference, but as the practical response to constraints that no single approach clears.

The full argument, including where combined heat and power fits and why early decisions carry such long penalties, is on the site.

Read the full piece → alexmarshallenergy.com/perspectives/efficiency-is-speed-to-power

Subscribe now

The forecasts have firmed up. The World Meteorological Organization now puts the odds of El Niño this summer at 80 percent, and NOAA gives better than three in five for a “super” event by year’s end. A strong one usually delivers the hottest year in the record books.

Subscribe now

The easy conclusion is that a record-hot year strengthens the climate case and moves low-carbon technology up the priority list. For the system I work in, that is the wrong call in the near term.

A strong El Niño suppresses rainfall across hydropower regions and raises cooling demand at the same time. The grid responds by leaning harder on whatever is firm and dispatchable, which today still means gas and, in many places, coal. El Niño is the warm phase of the Pacific temperature cycle that reorganizes weather across much of the planet, and its first-order effect on a stressed grid is to reprice resilience, not to accelerate clean power.

That is the surface of it. The full piece works through what a fast climate shock does to infrastructure that plans on a slow cadence, why AI data centers are unusually exposed when power, water, and heat converge on one site in the same season, and why the answer is sequence rather than ambition.

I argued in 2016 that power, not compute, would be the binding constraint on this industry. A super El Niño does not change that thesis. It stress-tests it in public.

Read the full argument here:

Alex Marshall

Subscribe now

The energy trilemma is the standard tool for weighing power decisions. You balance three goals that pull against each other: security, meaning reliable supply, then affordability, then sustainability. Improve one corner and you usually pay on another. It is drawn as a triangle, and the triangle held up for years because the balance point moved slowly.

It does not move slowly now.

Subscribe now

The IEA expects data center electricity demand to double by 2030, with AI demand tripling. That is not drift. It is a spike arriving faster than the planning system was built to handle, and it is bending the trilemma backward rather than just sideways. Facing multi-year waits for a grid connection, developers are buying speed and paying for it in carbon: on-site gas plants built for a single site, projects islanded off-grid for years. One procurement decision, made for speed, can give back years of emissions progress.

The triangle cannot show any of this, because it has no time axis.

So I named the missing dimension: the temporal trilemma. Time is not a fourth corner. You do not balance against it the way you weigh security against cost. It is the axis those three corners move along, and that axis now runs fast, and in both directions. Leave it out and you are deciding blind to the thing that has changed most.

If the problem is temporal, the response is sequencing. Decide what to deploy now, what to run as a bridge, and what to converge on, so a choice made for speed today does not foreclose your sustainability position tomorrow. That is the work the Structured Transition Model does.

I lay out the full argument, the data, and how the sequencing works in the new piece.

[Read the full article on alexmarshallenergy.com →]

— Alex Marshall

Subscribe now

The debate about AI right now is whether the spending is justified. The five largest US cloud and AI companies are guiding toward more than $635 billion in combined capital spending next year, most of it for AI infrastructure, and the skeptics have a real case on the economics.

That is not the question I care about. If you build or finance power generation, the one that matters is different. If AI investment corrects, what happens to the demand?

It holds. Most of it was never AI to begin with.

The IEA’s latest electricity outlook attributes only about half of projected US demand growth through 2030 to data centers. The other half is electrification, manufacturing moving back onshore, and the work of balancing a changing generation mix. None of that runs on AI capital cycles. It arrives whether or not the next training cluster gets built.

The supply side is already short. New projects are waiting years on transformers, switchgear, and generation. In a market that constrained, an AI correction does not leave a glut. It relieves pressure on a system that could not keep up with the non-AI demand on its own. That is the part the AI narrative misses. The exposure sits with the constraint, not the valuation cycle.

I worked through the full argument, the numbers behind it, and the two strongest objections, including why the dot-com fiber glut is the wrong comparison, in the new piece.

[Read the full article on alexmarshallenergy.com →]

— Alex Marshall