I sit on an HOA board.

Eighty-six units on a golf course in Westchester. Twice a year we review applications to install whole-home gas generators. We check the decibel ratings. We schedule the weekly test runs for when the lawn crews are out so nobody complains.

It is a small vote on a small thing.

But it is a vote. The applicant doesn’t install the generator if we say no.

Now scale it.

A one-gigawatt campus needs a rezoning, a conditional-use permit, a water agreement, an air permit, road access, and a local tax deal.

Six approvals, in six different forums.

The Thing Nobody Models

The AI buildout is usually discussed as a resource problem.

Enough chips. Enough capital. Enough land. Enough electricity.

Every serious analysis of the sector asks some version of that question, and the answers are getting better.

There is another question that doesn’t fit so neatly into a capacity forecast.

Where the resources do exist, will the people who live there permit them to be used this way?

A data center can have the land, the power, committed capital, and political support and still fail before construction starts.

There is no single approval that gets it built.

And the institution with the power to stop it changes as the project moves.

So who can still say no?

Read more

There are roughly six million EVs on U.S. roads. About one million can already move electricity in both directions.

At roughly 10 kW per vehicle, that puts the theoretical export capacity close to 10 GW.

That sounds enormous. It is also a little misleading if you stop there.

A million cars do not behave like a power plant. Some won’t be home when the grid peaks. During a long outage, the owner may need to drive. Most aren’t connected to a utility program that can call on the battery anyway.

Still, 10 GW is hard to ignore. It’s roughly the output of nine Vogtle-sized reactor units, except the batteries have already been purchased and are sitting in garages and driveways.

That’s where I started digging.

Why I Started Looking at This

We live in a condo complex in Westchester, New York. The power lines inside the development are underground, but they are fed by overhead lines outside it, and we lose power more often than we would like.

I sit on the HOA board, so I see the applications for whole-home gas generators. They work if you have an end unit. We don’t. We looked at home batteries too, but with a single-car garage, there isn’t much room to give up.

What we do have is a Tesla Model Y.

Our Model Y can’t currently power the house. Tesla hasn’t enabled vehicle-to-home charging on it, although there are signs the underlying hardware may be capable of it. Other EVs can already do this.

So we started looking for a big battery that could keep the house running when the power goes out.

Ideally, one we could drive and still like as a vehicle.

The Promise: Cars as Grid Assets

That introduced me to a whole collection of acronyms.

V2L, or vehicle-to-load, lets you plug appliances or equipment into the car. Useful, but fairly limited.

V2H, vehicle-to-home, is what interests us. With the right equipment, the car can supply the house during an outage.

Then there is V2G, vehicle-to-grid. In that case, the battery can send power beyond the house and back to the grid.

That was the part I hadn’t really thought about before.

Utilities are looking for new capacity for data centers, AI and other load growth. Meanwhile, millions of fairly large batteries are being bought by consumers and spending most of their time parked.

I wanted to know how much of that capacity could really be useful.

Only V2G turns the car into a grid resource.

The 10 GW estimate is the attention-grabbing number. The useful number is smaller.

To get from one to the other, you have to look at when the cars are actually available, what it costs to connect them, how repeated cycling affects the battery, and whether utilities can call on that power and pay for it.

I also wanted to see where V2G is already working. There are a few real programs with actual vehicles and actual compensation, and they tell us quite a bit about what it would take to scale this.

That’s what I looked at next.

Subscribe to keep reading.

Read more

On Wednesday morning, July 22, 2026, more than three gigawatts of electricity demand disappeared from the largest U.S. power grid in a matter of seconds.

A transmission line in Northern Virginia went out of service, prompting numerous data centers to disconnect from the grid and transfer to backup power. The sudden shift removed roughly 3% of PJM’s total demand at the time and created a voltage disturbance that was detected from Washington, D.C., to Chicago.

PJM said the event didn’t affect overall grid reliability, and Dominion Energy restored normal operating conditions within minutes. But the incident offered a glimpse of what happens when concentrated computing demand becomes large enough to behave like grid infrastructure.

Most software companies could grow for years without asking a utility whether enough power would be available for the next product release.

The servers existed. The data centers existed. The cloud kept most of that machinery out of view. Product teams could add users, enter new markets, and ship new features without waiting for a substation or transmission study.

The more I worked through the reporting for this edition, the more one idea kept resurfacing.

AI isn’t just changing software anymore. It’s changing what software depends on.

A proposed data-center campus requiring 500 megawatts to 1 gigawatt is not just a larger software deployment. At that scale, compute depends on generation, transmission, substations, cooling, land, equipment, utility agreements, and permits.

Models can improve in a matter of months. Building the infrastructure underneath them doesn’t.

That’s the part I don’t think the market has fully absorbed yet. The timelines are completely different, and those timelines are starting to shape where AI can actually be built.

None of this makes chips or models less important. But they now depend on a physical system that moves at a completely different speed.

That changed the question for me.

The debate is no longer whether AI needs more electricity. Everyone accepts that.

The harder question is who secured the right power, in the right market, before everyone else realized how scarce it was becoming.

In the full edition, I walk through why the software playbook no longer explains what’s happening.

We look at how grid timelines are becoming part of AI strategy, why the biggest technology companies are moving upstream into energy and infrastructure, and why announced capacity often has very little to do with capacity that can actually be delivered.

We also look at how the largest technology companies are moving upstream into energy and infrastructure, how power is redrawing the map of AI development, and why announced capacity is not the same as capacity that can be delivered.

If AI really is becoming infrastructure, then investors may have to start evaluating it very differently.

Continue with the full analysis as a paid subscriber.

Subscribe now

Read more

I opened Moonshot’s pricing page on July 16 expecting the story everyone was already writing. I went in expecting another article about cheaper AI. By the time I closed the pricing page, that wasn’t the article I wanted to write anymore. China ships another frontier model, the price of AI drops another notch, and somewhere in Virginia a data center developer quietly recalculates a spreadsheet. I’ve read that story enough times this year. But when I looked at the numbers, they didn’t match the headline.

Moonshot AI announced Kimi K3 that day — a 2.8-trillion-parameter system, the first open model it says belongs in the three-trillion-parameter class. It was live immediately through Moonshot’s products and API. The full model weights were released on July 27 under Moonshot’s Kimi K3 License.

Within hours, the narrative was familiar: another Chinese lab forcing prices down, more pressure on the American labs trying to earn back the cost of their buildout.

It’s an understandable conclusion. The problem is that we’ve seen this same conclusion after almost every major Chinese release over the past year. Moonshot’s own pricing points somewhere else.

Last week I listed four developments that would make me reconsider the AI infrastructure overbuild. K3 forced me to revisit one of them.

Compared with Moonshot’s previous model, K3 costs about three times more for new input and nearly four times more for output. Even the discounted rate for text the model has already processed doubled. K3 may need fewer tokens to finish a given task — but customers are still paying more to use it. That doesn’t tell us what K3 costs Moonshot to run. It tells us what Moonshot thinks the market will bear.

K3 forced me to separate two questions I’d been treating as one. Is AI getting cheaper to use? Or is it becoming more efficient?

K3 made me rethink the assumption that cheaper AI and more efficient AI are the same thing. The system is far larger than Moonshot’s last release. The technical report shows the model activates only a small fraction of its parameters for each token.

More capability may be arriving without a matching rise in the compute it takes to answer a question.

Greater efficiency doesn’t necessarily lower the price customers pay. It may also change where the machines run, how the electricity demand is distributed, and whether that demand appears where the grid planners were expecting it.

The market is watching the price. I think the infrastructure risk is hiding in the design.

In the full edition, I look at why pricing, efficiency, and openness have started to diverge, what that means for infrastructure planning, and why policy decisions may end up shaping AI deployment as much as engineering.

Continue with the full analysis as a paid subscriber.

Subscribe now

Read more

There’s a number I can’t stop turning over. ERCOT’s large-load queue — the list of companies asking Texas for power — is sitting at roughly 438 gigawatts. About 9 of those gigawatts have actually been approved to energize. Less than 4 are running.

Think about that for a second. That’s a claim staked on a physical system, made by people who have no idea yet which of their projects will get financed, built, or ever switched on.

The easy explanation is speculation — developers padding requests because queue position is worth something, hyperscalers keeping every door open, utilities planning for demand that might be a mirage. It’s all true; but I just don’t think it’s the whole story.

Because you can see the same thing earlier than the utility filing. Long before anyone calls ERCOT.

When an AI system hits a wall — a memory bottleneck, a latency problem, whatever — the industry’s reflex is rarely to slow down and solve it. The reflex is to add hardware: brute force. More chips. More racks. More power. The campus grows to fit the workaround, not the problem. And when nobody’s sure what the future looks like, you build for every version of it at once — which is a very reasonable thing for any one team to do, and a strange thing to watch happen a hundred times in parallel.

I’ve started calling this unearned scale. Not because big is automatically bad — some workloads genuinely need everything they’re asking for. Unearned scale is the capacity you add because adding capacity is easier than fixing the problem that made it necessary in the first place.

And here’s the part that should bother you more than the ERCOT number itself: none of this stays contained inside the data center. A compute decision quietly becomes a rack requirement. The rack requirement becomes a power request. The power request becomes a substation, a transmission line, a fight over water rights, and — eventually — an argument about whose electric bill absorbs the cost. By the time it reaches the grid, the original decision that caused all of it is nearly invisible. Nobody in that fight even remembers where it started.

We’ve spent a few years treating scale itself as proof of progress — bigger model, bigger campus, bigger check written. I keep asking a different question: what happens the moment the systems underneath can’t absorb the cost of all that bloat?

That’s the overbuild story. The queue is just the place where it finally becomes visible.

The paid edition traces where this begins, how it adds up, and which projects are getting the most exposed as the grid, the public, and capital markets start moving the goalposts. Subscribe to keep reading.

Subscribe now

Read more

Earlier this week I was joined by for a conversation about AI infrastructure.

Andy publishes on AI & the Environment, where he examines claims about AI’s environmental impact, particularly around energy, water, and data centers.

I publish AI Grid Report and spend much of my time working with energy developers, including inside a live ERCOT AI data-center project.

We were approaching the same issue from different directions: Andy studies the environmental claims surrounding AI, while I work inside the physical infrastructure required to power it.

After an hour together, three ideas stayed with me.

1. AI has become a physical infrastructure story.

For the past few years we’ve talked about models, chips and software.

I’m becoming convinced the harder questions now are physical.

Can we generate enough electricity?

Can we build transmission fast enough?

Can we connect projects before demand overtakes the grid?

Those aren’t software problems.

They’re infrastructure problems.

2. We spend a lot of time debating the costs of data centers.

We spend much less time debating what happens if we don’t build them.

That’s not an argument for approving every project.

Some deserve to be rejected.

But every decision has trade-offs.

The conversation should include both sides.

3. I’m starting to think we’ve been asking the wrong question.

Instead of asking:

Are data centers good?

Maybe we should be asking:

Compared to what?

Compared to not upgrading the grid.

Compared to slowing AI adoption.

Compared to delaying investment in infrastructure we’ll probably need anyway.

The full discussion covered a lot more ground, including AI infrastructure, electricity demand, environmental trade-offs, grid expansion, and where we think the conversation is heading over the next decade.

If you’d like to watch the complete conversation, you can find it here:

▶️ Watch the full conversation: AI, Data Centers & the Future of the Grid

The AI Grid Report

Live with Neil Winward & Andy Masley - Everything You Hate About Data Centers - But Shouldn't

Listen now

a month ago · 4 likes · Neil Winward and Andy Masley

I’d be interested to hear which idea you agree or disagree with most or anything you think we missed.

The AI Grid Report

Most conversations about AI focus on models.

Most conversations about energy focus on generation.

The AI Grid Report sits where those two worlds collide.

Each week, we explore the physical infrastructure powering AI—from electricity and transmission to data centers, permitting, capital, and the policy decisions that will determine what gets built next.

If you want to understand the forces shaping AI before they become headlines, subscribe below.

Subscribe now

Yesterday, I spent another day inside a Batch Zero project.

Nobody was arguing about transformers.

Nobody was arguing about transmission lines.

We were arguing over documents.

On paper, it looked like legal language.

Around the table, everyone understood something different.

One clause changed who committed capital first.

Another determined who could walk away.

A third decided who carried the risk if market conditions changed.

That’s when I realized something.

Most investors spend their time studying the wrong part of the project.

Every few weeks another AI infrastructure project makes headlines.

A company announces billions in investment. Politicians praise the jobs. Investors debate the megawatts. The market moves on.

It creates the impression that once the announcement is made, the difficult part is over.

Working inside one of these projects has shown me the opposite.

The announcement is where the real work begins.

Since April, I’ve been involved in the development of a large AI data-center project in ERCOT. While I can’t discuss confidential details, I can explain the process itself. And that process looks very different from the story most people imagine.

Long before a shovel reaches the ground, developers are negotiating land, financing, transmission, environmental reviews, utility requirements, and a regulatory process designed to determine which projects are actually ready to build.

Most of those decisions never make the headlines.

Yet they determine whether a project moves forward at all.

Inside this issue

You’ll learn:

• Why ERCOT abandoned its first-come, first-served interconnection queue

• What Batch Zero is designed to accomplish

• The seven gatekeepers standing between a developer and the grid

• Why today’s bottleneck isn’t where most investors are looking

• The lesson Batch Zero taught me about evaluating AI infrastructure investments

AI Grid Report examines the engineering, regulation, financing, and infrastructure decisions that shape the AI buildout long before they become headlines.

If you want to understand what happens after the press conference ends, you’re in the right place.

Subscribe to continue reading.

Subscribe now

Read more

Someone wrote an anti-data-center op-ed last week.

They researched it with an AI assistant.

They drafted it with an AI assistant.

The assistant answered from a rack in a building exactly like the one the op-ed condemns.

Every query drew power off a grid the piece argues the industry is straining.

None of that proves the op-ed is wrong.

It does suggest we’re asking the wrong question.

The dominant data center story is an outrage story: the developer is greedy; the neighbor is a NIMBY; the booster is a shill for the tech bros; the opponent is virtue-signaling.

Each of these is a claim about the believer—about motive, sincerity, and status.

None of them can tell you whether a specific facility is a fair deal for the people who live next to it.

A hypocrite can be right.

A saint can be wrong.

The character of the person holding the opinion isn’t what matters.

Instead of asking whether someone is sincere—or whether they’re reading the right newspapers (or Substacks)—ask a different question:

Is the arrangement itself fair?

Who bears the costs?

Who captures the benefits?

Does the price paid match the burden imposed?

That question needs a lens.

If someone tells you data centers are either saving the economy or destroying the grid, don’t argue. Ask them to score the deal.

This issue lays one out, scores a real facility with it, and shows why much of today’s data-center debate is arguing about the one thing the scorecard tells you to ignore.

Once you start separating the arrangement from the outrage, you’ll notice the same mistake in almost every infrastructure debate.

If you invest in, site, finance, or regulate this infrastructure, the character debate is a distraction you can’t afford.

For paid subscribers

This is where the article stops describing the debate and starts scoring it.

Below, I apply the framework to a real data-center tariff, compare two nearly identical projects that receive opposite fairness scores, and show why Ohio and Texas are quietly moving toward the same underlying model.

You’ll also get:

• the three fairness axes,

• the but-for gate,

• the seven distributive tests,

• a side-by-side comparison of two seemingly identical projects that receive opposite fairness scores, and

• two real regulatory case studies showing how PJM and ERCOT are already converging on this approach.

The goal isn’t another opinion.

It’s a scorecard you can apply to the next project that lands on your desk.

Once you’ve scored a real project, it’s surprisingly difficult to go back to arguing about motives.

Subscribe now

Read more

The campus was finished.

The financing had closed.

The customer was ready.

Power had been contracted.

The building still couldn’t turn on.

Not because the money disappeared.

Not because permits were delayed.

Not because demand weakened.

One machine hadn’t arrived.

That surprised me.

For the past two years, most of the conversation around AI infrastructure has focused on capital spending, hyperscaler demand, and billion-dollar campus announcements.

Those numbers are easy to measure.

They’re also becoming less useful for answering the question investors really care about:

When does a project begin producing value?

That question sent me down a surprising rabbit-hole.

It led to a machine most investors never think about, because it has been around for over 100 years.

Relative to the campuses it serves, it’s inexpensive.

It receives almost no attention outside the utility industry.

Yet a missing transformer can leave a fully funded AI campus waiting years for electricity.

That isn’t a procurement detail.

It’s becoming an investment question.

Last week I argued that capital no longer guarantees a project will receive power. That edition was about regulation in PJM and the ability of a California utility to provide the electricity.

This week I wanted to understand some other pieces of the connectivity puzzle.

The answer turned out to be much smaller — and much more consequential — than I expected.

AI Grid Report exists for readers who want to understand the physical constraints behind the AI buildout: power markets, transmission, interconnection, utility planning, and the infrastructure bottlenecks that determine where capital becomes operating capacity.

If your work depends on knowing which AI infrastructure projects can actually get built, subscribe below to continue reading.

Subscribe now

For paid subscribers, the full report begins below.

Read more

Hundreds of millions of dollars were committed.

The site was selected.

The permits were approved.

Customers were ready.

The project still couldn’t turn on.

Not because financing disappeared.

Not because demand dried up.

Because the electricity never arrived.

If you invest in AI infrastructure, utilities, energy, or data centers, this isn’t an isolated case. It’s an early indication of how the economics of AI infrastructure are beginning to change.

For the past two years, investors have understandably focused on hyperscaler capital spending, GPU demand, and multi-billion-dollar campus announcements. Those numbers are easy to measure and easy to compare.

The harder question is whether those projects will receive power when they need it.

That question is increasingly important because every additional month waiting for electricity changes construction schedules, financing costs, projected returns, and competitive positioning. Building a data center and energizing one are no longer the same challenge.

This week’s research begins with a project in Santa Clara that appeared to have every advantage.

The land was secured.

The financing was complete.

The permits were approved.

A power agreement had already been signed.

Yet the project remained dark.

What happened there reveals something much larger than a single delayed development.

It exposes a structural constraint that is beginning to shape where AI infrastructure can be built.

After tracing the Santa Clara case, I found the same pattern appearing across multiple U.S. power markets. The bottleneck isn’t disappearing. It’s moving into parts of the system that receive very little attention but increasingly determine whether a project becomes an operating asset or an expensive stranded investment.

In this week’s report, we examine:

• Why more and more projects fail after the press release rather than before it.

• Why interconnection queues are only the first stage of a much longer timeline.

• Why even locating beside an existing nuclear plant doesn’t guarantee access to power.

• Why the same data center can become operational in one market while remaining stalled for years in another.

Most coverage of AI infrastructure follows the announcements.

AI Grid Report follows the physical system that determines whether those announcements ever become reality.

Every week I publish original research on power markets, transmission, interconnection, utility planning, and the infrastructure constraints that shape capital allocation across the U.S. energy system.

If your work depends on understanding where AI investment can actually become operating infrastructure—not simply where it is announced—I think you’ll find value in becoming a subscriber.

Subscribe now

For paid subscribers: the full report begins below, including the complete Santa Clara case study, supporting figures, and the analytical framework I now use to evaluate whether an AI project is likely to receive power on schedule.

Read more

Instead of writing more, I thought I’d share what was going through my mind after spending the week researching this issue.

One conclusion kept coming back to me.

The market may be asking the wrong question.

In this short video, I explain why I think the next AI boomtown won’t necessarily be the place a…

Read more

The most widely quoted number in China’s AI buildout may also be the least useful.

Over the past few weeks, one number has dominated the AI infrastructure conversation.

$295 billion.

Beijing’s proposed ¥2 trillion national AI infrastructure plan has been compared repeatedly with the roughly $725 billion that U.S. technology companies expect to invest this year.

The conclusion often sounds obvious.

America is spending more.

China is catching up.

The race can be measured in dollars.

That interpretation is appealing because it is simple.

It is also incomplete.

The comparison mixes one year of private U.S. investment with a five-year Chinese government program. It ignores purchasing-power differences that dramatically change what each dollar buys. Most importantly, it measures money committed rather than useful computing power delivered.

Those are not the same thing.

For investors, that distinction matters.

Capital markets tend to reward headline spending long before they measure whether that spending produces usable infrastructure. AI is no different. A billion dollars invested in servers, substations, cooling systems, or transmission does not necessarily create the same amount of computing capacity—or create it on the same timeline.

The question investors should be asking is not:

Who is spending more?

It is:

Who is converting capital into useful compute more effectively?

That turns out to be a much harder question to answer.

It also leads to a very different view of the AI race.

In this research supplement, we break China’s AI buildout into the individual layers that determine outcomes—from construction costs and electrical infrastructure to chips, deployment speed, and the time required to bring new compute online.

Rather than comparing spending alone, we introduce a framework that separates two different competitive advantages:

• The cost of useful compute

• The speed at which that compute reaches production

Those two measures tell a very different story from the headlines.

If you’re already familiar with AI Grid Report, you’ll recognize the second concept immediately. Speed-to-power has become one of the defining themes of our research because deployment speed increasingly determines investment outcomes across AI infrastructure. This report extends that thinking beyond the U.S. grid to compare the broader economics of AI buildouts in the United States and China.

If you’re new here, welcome.

Every week, AI Grid Report examines the intersection of AI infrastructure, power markets, transmission, utilities, semiconductors, and capital allocation. Our goal is simple: identify the physical constraints shaping the AI economy before they become consensus.

Subscribers receive original research, proprietary frameworks, and in-depth analysis designed for investors, developers, operators, and energy professionals.

Subscribe now

Already a premium subscriber?

Your research supplement begins below.

Read more

Every investor is looking at Northern Virginia.

That may be exactly why it’s worth looking somewhere else.

For years, Northern Virginia has been the center of the data-center universe. Loudoun County’s Data Center Alley now contains one of the largest concentrations of digital infrastructure on the planet. Billions of dollars have poured into the region, and the AI buildout has only accelerated that trend.

Success, however, creates its own problems.

Power is harder to secure than it was a few years ago. Grid upgrades take time. New demand continues to arrive faster than infrastructure can be expanded. The market that attracted the first wave of AI investment is becoming a more difficult place to support the next one.

That observation led me to a simple question:

Where does the next boomtown emerge?

Not the next place to announce a project.

The next place capable of supporting one.

The distinction matters because the beneficiaries extend far beyond the companies building data centers. Utilities, generators, transmission owners, equipment suppliers, infrastructure developers, and local communities all participate when a region becomes a destination for large-scale investment.

Hundreds of billions of dollars are expected to flow into AI infrastructure over the coming decade. Some proposed campuses already require more electricity than many cities consume today.

The obvious assumption is that capital will continue flowing to the same places.

The grid data suggests the story may be more complicated.

Over the past several months, I’ve been tracking interconnection queues, transmission plans, utility filings, large-load requests, generation development pipelines, and regional infrastructure constraints across the major U.S. power markets.

A pattern began to emerge.

The regions attracting the most attention are not always the regions with the clearest path forward.

Some highly publicized markets are running into increasingly visible bottlenecks. Others rarely appear in AI headlines despite possessing many of the ingredients developers care about most: available power, development flexibility, existing infrastructure, and room to grow.

The next AI boomtown may already be taking shape.

Most investors simply are not paying attention to it yet.

This week’s research explores where that opportunity may emerge and what the underlying grid data is signaling.

Among the findings:

• PJM continues to dominate the conversation but faces mounting infrastructure challenges.

• MISO is attracting far more investment activity than many investors realize.

• Several regions possess stronger power-delivery characteristics than their reputations suggest.

• Some of the most interesting opportunities may emerge far from the markets receiving the most attention today.

For investors, the goal is not simply identifying where capital is flowing.

The goal is identifying where it may flow next.

If you’re new to AI Grid Report, this is exactly the type of research we publish every week.

Subscribers receive original analysis on AI infrastructure, power markets, grid constraints, transmission planning, utility strategy, and the capital-allocation implications shaping the future of the U.S. energy system.

Our goal is simple: identify the trends that matter before they become consensus.

Whether you’re an investor, developer, operator, utility executive, consultant, or infrastructure professional, AI Grid Report is designed to help you understand where power is moving, where capital is following, and where opportunities may emerge next.

Subscribe now

The full report examines the regional data, infrastructure constraints, transmission realities, capital flows, and power-market dynamics shaping the next generation of AI infrastructure development.

Paid subscribers can continue below.

Read more

Several readers asked how we arrived at the ranking in Issue #11, The Most Overlooked Winner in the AI Infrastructure Race.

The main article focused on a question that is becoming increasingly important as AI infrastructure development accelerates:

Which power markets appear best positioned to support large-scale AI and data-center growth over the next decade?

Rather than focusing exclusively on electricity prices, tax incentives, or headline announcements, the analysis evaluated the structural characteristics that influence whether proposed demand can realistically reach energized operation.

The conclusion surprised many readers.

Several regions that dominate industry discussion did not rank at the top of the framework. Meanwhile, one region that rarely appears in conversations about AI infrastructure emerged as the strongest overall candidate.

That result generated a number of questions:

• How were the regions evaluated?

• Why did certain markets score higher than others?

• What role did queue composition, demand pressure, transmission requirements, and infrastructure readiness play in the final ranking?

• Why did historical performance matter less than current queue structure?

This research supplement provides the supporting analysis behind the conclusions presented in Issue #11.

Inside, you’ll find:

• The complete regional evaluation framework
• Queue composition analysis for each finalist region
• Demand-pressure calculations and supporting data
• Transmission and infrastructure considerations
• Historical performance comparisons
• Supporting charts and source documentation
• Additional analysis that did not fit into the main publication

If you have not yet read Issue #11, I recommend starting there first:

Issue #11 — The Most Overlooked Winner in the AI Infrastructure Race

The full research supplement is available below for paid subscribers.

If you’re already a subscriber, thank you for supporting this work.

AI Grid Report exists to help investors, developers, operators, utilities, and energy-market participants understand how AI infrastructure, power markets, grid constraints, transmission planning, and energy policy are reshaping capital allocation across the United States.

Each week, we analyze the developments that may matter before they become consensus.

The goal is simple:

Identify where the energy system is changing, what those changes mean for capital, and where opportunities may emerge as a result.

If that appeals to you, we hope you will join our community.

Subscribe now

Premium subscribers can download the full Research Supplement #11 below, including the methodology, supporting data, regional analysis, and source documentation behind this week’s ranking.

Read more

Everyone assumes the AI infrastructure race will be won by the same handful of regions.

ERCOT.

PJM.

Maybe MISO.

Those are the regions attracting the headlines, the investment dollars, and the bulk of developer attention.

But attention and opportunity are rarely the same thing.

Over the past several weeks I’ve been comparing the smaller U.S. grid regions through a simple lens:

If you needed a large amount of power for a new AI project, which grid would give you the shortest credible path to energization?

Figure 1 — The Three Contenders

On paper, the outcome looks obvious.

One has enormous renewable resources.

Another sits next to one of the largest concentrations of demand in North America.

The third is smaller, quieter, and rarely appears on anyone’s shortlist.

If you ranked them by reputation alone, you’d probably put that third contender last.

The data says otherwise.

New to The AI Grid Report?

The AI Grid Report helps investors, developers, bankers, and energy executives understand the forces shaping AI infrastructure, power markets, and capital allocation.

Each week we combine proprietary research, grid analysis, and market intelligence to answer a simple question:

Where is the next infrastructure opportunity emerging—and what are most people missing?

If you’re interested in the intersection of AI, energy, and investment, subscribe below to receive future issues.

Subscribe now

Paid subscribers can continue below for the full analysis, including:

• The complete comparison of SPP, NYISO, and ISO-NE

• The demand-pressure framework driving the rankings

• The full Speed-to-Power analysis

• Our conclusion on which overlooked grid offers the strongest path to new AI infrastructure

Read more

Following this week’s AI Grid Report issue 10 on the queue behind the queue, we continued refining the framework, incorporating additional sources, feedback, and new data. The result is the a research supplement about the AIGR Speed-to-Power Index.

What the Index Measures

The AIGR Speed-to-Power Index measures a question that traditional interconnection data often misses:

How quickly can a large-load project move from announced demand to energized operation?

Most public discussion focuses on queue size.

We focus on how quickly projects move from demand to energized operation.

The index evaluates the physical and procedural factors that determine whether a project can move from concept to power on a realistic timeline.

Among other factors, the framework considers:

• Large-load queue magnitude

• Post-approval energization delays

• Transmission build requirements

• Substation and infrastructure readiness

• Regional pathways for co-located generation and alternative interconnection structures

The goal is not to estimate total demand but rather to identify where demand is most likely to reach power first.

Why Speed-to-Power Matters

For much of the last decade, investors could assume that approved demand would eventually become operating demand.

That assumption is becoming less reliable.

AI infrastructure is forcing investors to confront a question that rarely mattered at this scale before:

How much time sits between demand and delivered power?

A signed agreement no longer guarantees a predictable timeline to energization.

The gap between announced capacity and energized capacity is beginning to influence project economics directly.

Projects that can reach power sooner may command a premium over projects with larger headline demand but weaker execution certainty.

In that environment, speed-to-power becomes more than an operational metric.

It becomes a capital allocation metric.

Three Findings That Stood Out

1. Queue Size Is Becoming a Poor Proxy for Deliverable Power

Several regions continue to report exceptionally large demand pipelines.

Far fewer projects have demonstrated a clear path from approval to energized operation.

The difference between those two numbers may become one of the most important metrics investors track over the next several years.

2. The Delay Often Begins After Approval

Interconnection approval remains important.

But in several regions, transmission upgrades, substation capacity, equipment availability, and energization timelines now represent a second waiting period that can stretch for years.

Approval can arrive years before usable power.

3. Alternative Paths Are Becoming Strategically Important

Grid operators are increasingly evaluating frameworks that allow developers to pair generation with load, accept flexible service arrangements, or pursue alternative energization structures.

The regions creating workable pathways around traditional bottlenecks may attract projects that cannot afford to wait for conventional timelines.

The full Speed-to-Power Index, methodology, regional scoring framework, and supporting analysis are available below for paid subscribers.

If you’re already a subscriber, thank you for supporting this work.

AI Grid Report exists to help investors, operators, and developers understand how AI infrastructure, power markets, grid constraints, and energy policy are reshaping capital allocation across the United States.

Each week, we track developments across ERCOT, PJM, MISO, utilities, transmission planning, tax-credit markets, and large-load growth to identify the trends that may matter before they become consensus.

If you’re not yet a subscriber, I invite you to join us.

Subscribers receive weekly analysis, proprietary research, original frameworks, and occasional research supplements like the Speed-to-Power Index.

The goal is simple: help readers understand where the energy system is changing, what it means for capital, and where opportunities may emerge as a result.

Subscribe now

Read more

While the market is focused on how much power is waiting for approval, the real investment race begins after approval.

Most discussions around AI infrastructure center on generation interconnection: how much power is waiting in the queue, where projects sit in line, and how quickly new supply can reach the grid.

Across ERCOT, PJM, and MISO, another backlog is taking shape behind the public one.

The market still talks about AI like a software story.

The grid is turning it into a power story.

And that may determine who captures value first.

Welcome to AI Grid Report.

We spend an unusual amount of time reading Independent System Operator (ISO) filings, grid queue reports, and transmission planning documents so you don’t have to.

Then we connect those dots to AI infrastructure and where capital is being deployed.

Paid subscribers receive the full analysis—and my personal view on where the edge is forming before it becomes obvious.

If you’re already here: thank you.

If you’re new: you picked a good week.

Subscribe to unlock the full issue →

Read more

A month ago I wrote the first recap. The premise was that capital is pouring into AI infrastructure on the assumption that power, transmission, and interconnection are solvable inputs. They aren’t: they are the constraint, and the constraint is neither fully capitalized nor properly priced.

The second month tested that argument against the regions themselves. Four issues plus bonus content, working outward from a single auction print toward a national map. The capacity auction that printed at the cap for the second time, with FERC extending the ceiling rather than letting the market discover what an uncapped print would look like. A Midwest retirement schedule that erases roughly a third of average demand on a timetable the replacement queue can’t meet. A pan-regional review showing nine grids making nine different bets on how to firm the same renewable buildout. A political backlash that is no longer stopping the build but is increasingly choosing where it survives. And, as bonus content, a discussion of how the transferable tax credit market has quietly become the policy tool deciding which physical systems are financed.

What follows is the second month, condensed. The full archive is at www.theaigridreport.com.

ISSUE 06 · The Market That Can’t Settle

PJM’s December 2025 capacity auction settled at $333.44 per megawatt-day. That was the cap. It was also the second consecutive PJM auction to print at the cap, after the July auction settled at $329.17. Together the two auctions secured 134,479 MW for $16.4 billion in annual payments — at a procured reserve margin of 14.8%, the lowest in PJM’s history. The cap was supposed to be temporary. On April 28, FERC extended it through the 2029/30 delivery year. The market still doesn’t know what an uncapped print would look like.

The mechanics underneath the print are the more interesting story. Roughly 5,500 MW of fossil capacity is scheduled to retire — Brandon Shores, H.A. Wagner, Conemaugh, Keystone — but none of it is leaving. Brandon Shores and Wagner are operating to May 2029 under FERC-approved RMR agreements; Keystone and Conemaugh were held open through 2032 under a Pennsylvania consent decree. PJM connected 2,117 MW of new generation across all of 2025, almost entirely solar, while forecast load growth of roughly 5,250 MW arrives over the same window, primarily in Northern Virginia. The capacity market is procuring the gap the system cannot fill, and FERC has set the price at which that procurement may occur.

The PJM-versus-ERCOT comparison anchors the piece. Same demand, same generation technologies, different process. ERCOT’s connect-and-manage interconnection, energy-only market, and single-state regulator move the queue at near-zero withdrawal. PJM filters two-thirds of projects out before they reach completion. PJM can’t copy ERCOT’s structure, but it can copy ERCOT’s cadence — and until it does, the cap is the auction, and the auction is a political instrument with FERC’s overseeing it.

ISSUE 07 · MISO Has a Retirement Problem

MISO’s headline number is a posted retirement schedule, not an auction price. Roughly 25 GW of operating generation is scheduled to retire — about 37% of MISO’s average demand, with no other major ISO close. SPP sits at roughly half that exposure; PJM is in the mid-teens; ERCOT remains below 3%. Of the 25 GW, roughly 14 GW is subbituminous coal and 8 GW is natural gas. The dispatchable, fuel-secure capacity that operates through extended weather events and low-renewable-output stress is the part of the fleet leaving the system.

Michigan is the version of the problem visible at a single regulator. DTE Energy has disclosed 8.4 GW of data-center load — Oracle (1.4 GW, approved), Google (1.0 GW, decision due September 2026), ~2 GW in advanced discussions, and 3–4 GW earlier-stage. Over the same window, ~4.6 GW of Michigan coal is scheduled to retire: Monroe (3.3 GW, 2028 and 2032) and J H Campbell (1.3 GW, operating under its fourth federal 202(c) emergency order). The same balance sheet is selling the retiring capacity and buying the new load, and the regulator can watch the arithmetic in a single docket. Campbell is the precedent the rest of the 2028 fleet now has to underwrite to — a coal unit nominally retired a year ago that is still operating, with roughly $200 million in operational losses being contested before FERC.

The queue does not solve this on a one-for-one basis. MISO has ~230 GW active in queue, and it looks like an order-of-magnitude replacement for the retirements until you apply withdrawal rates and accreditation. Coal operates at 45–55% capacity factor; solar at ~25%. Once the cascade compresses the nameplate into accredited firm capacity, the apparent 9:1 replacement ratio falls to roughly 1.6:1. The retirement schedule prints either way. The replacement supply does not.

ISSUE 08 · What America Is Trying to Build

The interconnection queue is more than a backlog. Every project in it has committed capital and accepted years of transmission-upgrade exposure before delivering a megawatt. The composition is therefore a map of what developers believe is worth building — and it now reveals a national power system converging around one consensus and diverging around one unresolved question.

The consensus is the disappearance of coal. Across more than 1 TW of active queue volume in the seven major RTOs, roughly 0.5 GW of new coal appears. The existing fleet still operates, and in PJM still settles in the capacity auction; no one is willing to spend years queuing new coal through interconnection. Solar plus storage has become the national default, with roughly 956 GW of solar and 890 GW of batteries sitting in active queues at end-2024. A decade ago the storage figure was close to zero. The queue is now a renewable-plus-battery expansion in aggregate, regardless of the regional politics surrounding it.

The divergence is the firming layer. PJM still leans on gas and a long-dated nuclear bet — including a roughly 10 GW nuclear cluster concentrated in restarts and uprates rather than greenfield. ERCOT is attempting to firm renewable growth primarily through storage, with 176 GW of batteries in queue, the largest in the country, alongside a 233 GW large-load queue that is more than 70% data centers. MISO has shifted back toward gas as coal retires. CAISO, NYISO, and ISO New England show almost no new gas at all and are implicitly assuming either slower hyperscaler arrival, lower reliability tiers, or longer-than-planned operation of the existing thermal fleet. None of those assumptions is obviously wrong, and none is obviously safe. Same continent, different bets.

ISSUE 09 · Data Center Opposition Is Deciding Who Wins

Local rejections of US data-center projects totaled 49 across all of 2025. The first five months of 2026 produced 89. Q1 alone saw 20 cancellations representing more than $41.7 billion in shelved investment. The slope is the story: the rate of refusal is now rising faster than the rate of construction, even as the pipeline of roughly 2,788 facilities implies a two-thirds expansion off the installed base.

Both sides of the argument have real numbers. Virginia’s data-center industry supports roughly 74,000 jobs and more than $9 billion in annual GDP contribution; the state’s sales-tax exemption returns 48 cents per dollar foregone, well above the 17-cent average for other Virginia incentives. JLARC (Joint Legislative Audit and Review Commission) still classified the benefit as moderate and concluded the exemption does not fully pay for itself. The 833% year-over-year jump in PJM’s capacity-auction print, and the $11.24 monthly residential bill increase Virginia’s SCC approved for 2026, are the kitchen-table version of the same fight. The percentage of Virginia residents comfortable with a new local data center fell from 69% in 2023 to 35% in 2026. Support for state tax breaks fell from 61% to 37%. A 34-point swing in three years is what gives legislators cover to move.

The resistance is selecting the composition of the buildout rather than halting it. MISO’s data-center footprint has grown at 43% per year since 2020, driven by Virginia saturation and developers routing toward jurisdictions where opposition has not yet organized. The political-risk filter is converging on the same answer as the power-economics filter from Issue 07: the projects most exposed to community resistance are the same ones weakest on supply-side fundamentals. The version of the buildout that survives is the version whose power is sited where no one is close enough to object.

BONUS CONTENT · Transferable Tax Credits Are Quietly Reshaping Infrastructure Capital

The transferable tax credit market reached $42 billion in 2025, up 27% from $32 billion in 2024. With tax equity and preferred equity included, total tax credit monetization hit $63 billion. The market grew through OBBBA reconciliation, not despite it. Eligibility tightened, the qualifying pool narrowed, and the buyer-side market remained intact. The §6418 transferability provision that opened the market in 2022 is still the policy tool deciding which physical systems are financed.

The framing matters more than the mechanism. A transferable tax credit is technically a financial instrument, but economically it is beginning to function as a directional signal for capital allocation into energy, manufacturing, grid infrastructure, and industrial capacity. A $5 million tax liability becomes $4.5–4.6 million in credits funding a real project with a placed-in-service date. The project exists before the capital arrives, which makes the accountability structurally different from an ESG score the rating agency can revise after the fact. For family offices and long-duration capital pools that can underwrite illiquid, politically shaped, operationally complex infrastructure cycles, the structure fits unusually well. The question has moved from how to minimize the tax bill to what physical system the tax bill should fund.

The Through-Line

Five pieces, one argument extended: the constraint is moving, and it is moving in a direction the headline numbers don’t capture.

PJM’s capacity auction shows the political ceiling rather than the physical scarcity price. MISO’s retirement schedule happens whether the queue connects or not. The pan-regional queue shows what developers believe — and what they no longer believe — about which firming technology will pay. The rejection database shows which jurisdictions have organized against the build and which have not yet. The transferable tax credit market shows which projects are financed at all. Each of these is a price the market is paying for a problem it has not solved, and each of them surfaces somewhere other than where the original constraint sat.

A month ago I wrote that the grid constraint had shifted. The second month sharpened the claim: every mechanism the system uses to manage the constraint — auction caps, RMR contracts, emergency orders, queue filters, siting moratoria, tax credit eligibility — relocates the cost rather than resolving it. The open question is the same as last month, with two more data points behind it. How long can the market continue talking about the headline before it is forced to price the outcome?

Go Deeper on Each Issue

– PJM printed at the cap for the second time, and FERC extended the ceiling rather than letting the price discover itself →

– MISO’s retirement schedule is 25 GW, and the queue replacing it is composed of different generation types →

– Nine regions, nine queues, one disappearance and one emerging firming-layer disagreement →

– The rejection rate is now rising faster than the construction rate, and it is choosing where the build survives →

– Transferable tax credits have evolved beyond accounting exercises into the financing layer for the physical buildout →

What Comes Next

The third month moves underneath the queue and into the wires. Transmission planning, the FERC co-location order, and the question of whether behind-the-meter generation is a way around the grid or a new layer of it. The Constraint Index continues to update weekly. The full archive, the index, and the paid deep dives are at theaigridreport.com.

The Fight That Everyone Wants to Have

Local rejections of US data-center projects totaled 49 across all of 2025.

The first five months of 2026 produced 89. Twenty projects were canceled in Q1 alone — more than $41.7 billion in planned investment shelved.

The political revolt against data centers is no longer a Northern Virginia curiosity.

It is a national pattern with a measurable slope, and the slope is the story: the rate of refusal is now rising faster than the rate of construction.

That one fact changes how you read every headline about AI infrastructure spending.

US data-center capacity has grown at 24% annually since 2020. The average facility coming online in 2025 was 80 MW — three times the size of five years ago. The latest campuses are sizing to 1 GW and beyond. Virginia alone hosts 665 facilities. Its data centers consumed roughly 40% of the state’s electricity in 2024. Northern Virginia’s “Data Center Alley” accounts for an estimated 14% of all data centers worldwide.

That concentration is the mechanism behind the backlash now taking shape.

This is the constraint this issue examines — not whether the buildout happens, but which version of it survives contact with the communities it lands in.

Below the paywall:

• who is right on jobs and water — and why both sides can cite real numbers

• why electricity-rate fights become political fast

• what the financing structure says about long-duration AI capex

• and why the projects facing the least political resistance may end up shaping the build

The resistance is not stopping the build.

It is increasingly deciding what gets built — and where.

Subscribe to continue reading.

Read more