By Tony Grayson, Tech Executive (ex-SVP Oracle, AWS, Meta) & Former Nuclear Submarine Commander

Last week, Fermi lost a third of its value in a single session. The prospective anchor tenant withdrew its $150 million funding agreement, and the stock declined by 46% intraday before settling around 33%. If you're wondering what went wrong, the answer isn't complicated. It's physics.

The Fermi stock crash is the same physics I've been writing about for months.

What Actually Happened

Fermi (NASDAQ: FRMI) is a development-stage REIT developing "Project Matador"—an 11-gigawatt campus near Amarillo, anchored by the "Donald J. Trump Generating Plant." That's the actual name in the NRC docket. Former Energy Secretary Rick Perry co-founded the company. The team is heavy on political connections and light on people who've actually operated utilities at scale.

In October 2025, Fermi went public at $21/share, implying a valuation of around $12.5 billion. The stock popped in early trading, briefly topping $19 billion. Pre-revenue. No delivered power. No binding tenant leases. Just a pitch deck and a 99-year ground lease on 5,800 acres in the Texas Panhandle.

In November, Fermi announced a $150 million Advance in Aid of Construction Agreement (AICA) with their first prospective tenant. The company called it "a major step forward." They emphasized the tenant was "a very creditworthy counterparty."

On December 11, the tenant terminated the AICA. The exclusivity window had expired, and they walked.

No funds were ever drawn. The AICA was a commitment to fund shared-site work—roads, substations, and infrastructure. It wasn't a rent or a binding lease, just a handshake with an expiration date.

The company says lease negotiations may continue. Maybe they will. But the market didn't care about "may." The market was concerned that the first real test of demand returned negative results.

A press release isn't a lease. An MOU isn't a wire transfer.

This Isn't a One-Off

Fermi isn't an outlier. It's the template for what's coming.

Right now, there's a flood of capital chasing anything with "AI" and "Nuclear" in the pitch.

According to Goldman Sachs, data center power demand will increase 165% by 2030. The IEA projects global data center electricity consumption will double to 945 TWh—equivalent to Japan's entire electricity consumption.

Those numbers are real. But most of the capital chasing them is betting on a future that doesn't survive contact with engineering reality.

The pattern repeats: Developer announces ambitious project with a politically connected team, market assigns a massive valuation based on TAM slides, and then the first real commitment test fails. Stocks repriced overnight.

We saw it with Kemper. We saw it with UAMPS/NuScale. We're going to see it again, because the underlying thesis, that you can timeline-hack nuclear power to match AI demand, is physically impossible.

The Physics

I spent my naval career sleeping 50 feet from a reactor. I'm a massive proponent of nuclear energy. But I also know that the atom doesn't care about your quarterly guidance.

Fermi has site control. That's real, a 99-year ground lease with Texas Tech for about 5,800 acres. But site control isn't power delivery.

They're promising "bridge power" from natural gas by 2026. Maybe. But gas turbines have 18-24 month lead times right now, and they haven't disclosed an EPC contractor or a firm pipeline interconnect. So even the bridge has gaps.

The valuation isn't priced for gas though. It's priced for nuclear. And I've written extensively about why the Gen IV SMR timeline is 2035—if you're lucky. The ASME code testing alone for Gen IV reactor materials takes years. HALEU fuel supply is constrained. The NRC licensing process hasn't been meaningfully accelerated despite what the press releases say.

Anyone promising you a new reactor in four years either hasn't read an NRC docket or is hoping you won't.

The Financing Problem

This isn't just an engineering problem. It's a financing problem.

I wrote about this in The 3-Year Tenant in a 30-Year Building. AI tenants are signing shorter leases because they know their hardware will be obsolete in 36 months. Meanwhile, developers are financing 30-year assets against those short-term commitments.

GPUs depreciate in 3 years. The Blackwell cluster you're building for today will be legacy hardware by 2028. Meanwhile, SMRs take 10+ years to license and build—even the optimistic NRC timelines don't deliver commercial power before the mid-2030s. Try getting a debt market to underwrite that spread. The risk premiums alone kill the economics.

You're asking lenders to bet that demand stays constant for a decade, no technological disruption changes compute requirements, the political environment stays friendly, and you hit every milestone on time. Miss any of those and you're holding stranded assets with debt service you can't cover.

Fermi's AICA termination is a preview. The tenant reviewed the timeline, their own roadmap, and concluded the risk wasn't worth it. That calculus will recur across dozens of projects in 2026 and 2027.

I call this the Kemper Trap, in which political pressure and capital attempt to override thermodynamics. Kemper was intended to demonstrate that clean coal could work. It cost $7.5 billion and never operated as designed. The parallels to today's AI nuclear plays are uncomfortable.

The Hidden Stack

Here's what the slide decks don't show you.

Real infrastructure isn't about who you know in Washington. It's about what I call the Hidden Stack, the stuff that actually determines whether a project gets built.

Water. Thermal power plants need cooling. In the Texas Panhandle, that means millions of gallons in a region that's already stressed. Fermi announced an MOU with an unnamed Hungarian company for "hybrid dry-wet cooling towers." An MOU. With a company that has to ship technology across an ocean. For a site that doesn't have firm water rights.

Interconnection. You can build all the generation capacity you want. If you can't get it to the grid (or if you're behind-the-meter, if you can't get utility backup power) you don't have a project. This is why physical infrastructure vulnerability is the real constraint.

Fermi announced an Electric Service Agreement with Xcel's Southwestern Public Service for up to 200 MW. That's real progress. But 200 MW is a rounding error against an 11 GW vision. The interconnection queue for serious capacity is measured in years, not months.

EPC. This is where the hype usually dies. If there isn't a fixed-price contract with a Tier-1 builder, you don't have a project. You have a PowerPoint. Cost-plus contracts are how Kemper happened. They're how Vogtle went from $14 billion to $35 billion.

Fermi hasn't disclosed an EPC. They haven't disclosed contract terms. They haven't disclosed who's actually going to build this thing.

How to Spot the Next One

Fermi won't be the last. This playbook will continue through 2026 as capital continues to chase the narrative.

Before you write a check, find out if the tenant agreement is actually binding. AICAs and MOUs can be terminated without penalty. Binding leases have take-or-pay minimums and termination fees. If a tenant can walk without writing a check, the "commitment" isn't worth the paper it's printed on.

Ask if they have an interconnection agreement. Not "discussions." Not "applications pending." An executed agreement with a queue position and a timeline.

Check whether the delivery date matches reality. If they're promising nuclear power before 2035, ask which reactor design, what stage of NRC licensing, and where the HALEU fuel is coming from. If they can't answer, the timeline is fiction.

Find out who the EPC is. If they won't tell you who's building it and whether the contract is fixed-price, you're betting on a concept, not a project.

Figure out who funds the gap. Early site work costs real money. If the capital structure depends on tenant AICAs that can evaporate, what happens when they do?

Fermi failed the first test. Most of these projects will fail one of the others.

What Real Looks Like

I don't want to just throw stones. There are ways to build AI infrastructure that actually work.

The answer isn't 10-year mega-projects that gamble on political winds and reactor timelines. It's phased, industrialized buildouts that match the speed of the chip cycle.

Gas-to-nuclear bridge strategies that deliver power in 2026-2027 while leaving room for SMRs later. Modular builds you can deploy in months—and redeploy when requirements change. Projects sized to actual demand, not 11 GW fantasies. Distributed edge compute that puts inference closer to users instead of betting everything on mega-campuses.

The training-to-inference shift is changing the economics anyway. The massive centralized "AI factories" that everyone's racing to build may not be what the market needs in three years. Edge inference, distributed compute, and efficiency gains are going to reshape demand in ways that make today's mega-projects look like stranded assets.

As I wrote in From Parameters to Physics, the metric that matters now is watts per token, not total megawatts. The winners won't be the developers with the best Rolodex. They'll be the ones who can pour concrete, deliver power, and adapt when the market shifts.

The So What?>

Fermi is a canary.

Capital is flowing to narratives rather than to execution. Political connections are being valued over engineering capability. And the market is pricing in futures that don't survive contact with physics.

I'd love to be wrong. Show me an executed lease with take-or-pay minimums. Show me an interconnection agreement. Show me a financing structure that doesn't collapse the second COD slips by two years. I'll write the correction myself.

Until then, we're just trading futures on power plants that don't exist.

We keep funding Apollo Programs every ten months and acting surprised when they don't land. At some point we have to ask whether the people pitching these projects have ever actually built anything, or whether they're just really good at building slide decks.

Focus on the engineering.

FAQ

Who founded Fermi America?

Fermi America was co-founded by former U.S. Energy Secretary Rick Perry, along with co-CEOs Michael Goff and David Gross. Perry serves as Executive Chairman. The company went public in October 2025 at $21 per share, briefly creating three paper billionaires despite having zero revenue. The political connections helped generate significant media coverage, but couldn't prevent the December 2025 crash when commercial reality caught up with the narrative.

What is Project Matador?

Project Matador is Fermi America's flagship development—a planned 11-gigawatt AI data center campus in the Texas Panhandle, powered by nuclear energy. The project was the centerpiece of Fermi's IPO story. When the anchor tenant terminated their AICA agreement in December 2025 without drawing any funds, Project Matador lost its primary commercial validation, triggering the 33% stock crash.

What's the difference between an LOI and a binding lease?

A Letter of Intent (LOI) is a non-binding expression of interest to negotiate—it commits neither party to anything. A binding lease is a legally enforceable contract with specific terms, payment obligations, and remedies for breach. The distinction matters enormously: Fermi's valuation was built on LOIs and MOUs, not signed leases. When the LOI exclusivity expired and negotiations failed, Fermi had no legal recourse. An LOI is a handshake; a lease is a commitment.

How much power does an AI data center need?

Modern AI training clusters require 50-150 megawatts per facility, with the largest planned campuses targeting 1-2 gigawatts. For context, a typical nuclear reactor produces about 1 gigawatt. Fermi's 11-gigawatt vision would require the equivalent of 11 large nuclear plants—more than the entire installed nuclear capacity of most countries. These power demands are why AI infrastructure projects face physical constraints that financial engineering cannot overcome.

Why are hyperscalers backing out of data center deals?

Hyperscalers are rationalizing AI infrastructure commitments as the gap between capital expenditure and AI revenue becomes harder to ignore. GPU depreciation cycles (3-5 years) don't align with 15-20 year power infrastructure investments. When a hyperscaler can't project sufficient utilization to justify a multi-billion-dollar facility, they walk—even if it means forfeiting exclusivity deposits. Fermi's tenant performed this calculation and concluded the numbers didn't work.

What is GPU depreciation, and why does it matter for data centers?

GPU depreciation refers to the rapid decline in value of AI accelerator hardware as new generations launch. NVIDIA releases new architectures every 18-24 months, and each generation offers 2-3x performance improvements. This creates a mismatch: data center infrastructure is financed over 15-20 years, but the GPUs inside depreciate in 3-5 years. Tenants are reluctant to sign long-term leases when the hardware they'd deploy will be obsolete before the lease term ends.

What happened to other nuclear data center deals?

Nuclear-powered data center deals have faced repeated setbacks. Amazon's Talen Energy deal at Susquehanna was rejected by FERC in November 2024. Microsoft's Three Mile Island restart faces years of NRC review before any power flows. Constellation's Byron and Braidwood co-location proposals face similar regulatory hurdles. The pattern is consistent: announcements generate headlines, but regulatory and technical realities delay actual power delivery by years. Fermi's crash is the latest example of narrative outrunning execution.

What is a data center REIT?

A data center REIT (Real Estate Investment Trust) is a company that owns, operates, or finances data center properties and passes rental income to shareholders. Established REITs like Equinix and Digital Realty have decades of operating history and contracted revenue. Fermi structured as a REIT but had no operating properties or revenue at IPO—making it fundamentally different from traditional data center REITs despite using the same corporate structure.

How long does NRC nuclear licensing take?

NRC licensing for a new nuclear reactor typically takes 3-5 years for design certification and another 3-5 years for combined construction and operating licenses. First-of-a-kind reactor designs face additional scrutiny. The NRC approved NuScale's SMR design in 2020, but the UAMPS project using that design was cancelled in 2023 due to cost overruns, five years after the announcement. Anyone building a nuclear timeline should start with NRC.gov, not investor presentations.

Will Fermi stock recover?

Recovery depends entirely on Fermi's ability to sign binding leases with creditworthy tenants. The company still has its land position, equipment agreements, and political relationships—but those assets were already priced into the $19 billion IPO valuation. The market is now demanding proof of commercial traction before re-rating the stock. Until Fermi announces a signed lease (not an LOI, MOU, or AICA), the stock will trade on speculation rather than fundamentals.

What should investors watch for with AI infrastructure stocks?

Watch for the difference between announcements and commitments. Press releases about MOUs, LOIs, and "partnerships" aren't revenue. Key questions: Does the company have signed leases? What's the weighted average lease term? Are tenants investment-grade credits? Is there contracted revenue, or just pipeline? The Fermi crash is a case study in what happens when investors stop asking these questions.

____________________________________

Tony Grayson is a recognized Top 10 Data Center Influencer, a successful entrepreneur, and the President & General Manager of Northstar Enterprise + Defense.

A former U.S. Navy Submarine Commander and recipient of the prestigious VADM Stockdale Award, Tony is a leading authority on the convergence of nuclear energy, AI infrastructure, and national defense. His career is defined by building at scale: he led global infrastructure strategy as a Senior Vice President for AWS, Meta, and Oracle before founding and selling a top-10 modular data center company.

Today, he leads strategy and execution for critical defense programs and AI infrastructure, building AI factories and cloud regions that survive contact with reality.