GM. This is Milk Road AI, the newsletter that helps you understand the biggest AI stories before Wall Street fully prices them in.

​Here’s what we’ve got for you today:

Nexo is back in the U.S. - and new clients get 30 days of Wealth Club Premier perks! Higher yields, lower borrowing rates, and crypto cashback - start here.

THE AI BOOM NEEDS LANDLORDS

In 1973, a 38-year-old real estate lawyer named Jay Pritzker flew to Memphis to look at a struggling motel chain called Hyatt House.

The hotels themselves were nothing special, but what Pritzker saw was the land under them.

So he bought the chain and owned the real estate in the right cities before anyone else understood that was the game.

Forty years later, Hyatt is worth roughly $13B.

The AI industry in 2026 is at a similar stage to commercial real estate in 1973.

Everyone is focused on the buildings, the models, the chips, the benchmarks, the billion-dollar fundraising rounds.

OpenAI has raised $122B, and Microsoft is spending $190B on infrastructure this year alone. 

But underneath all of it, there is a question nobody is asking enough.

Where does all of this actually live? Because the dirty secret of the AI arms race is that the most powerful models on earth are completely useless without one thing: a building that can keep the lights on.

Not a metaphorical building but rather a literal one, with liquid cooling piped into the chips, massive power demands per rack, and utility deals that can take years to secure.

Those buildings do not exist at the scale the industry needs.

And a tiny New York company with a market cap smaller than some people's Hamptons real estate portfolios is quietly becoming the Jay Pritzker of the AI era.

The company is called WhiteFiber, WYFI.

And just like Pritzker, they are not trying to build the best AI, they are buying the land underneath it.

First, let's talk about the actual problem

Here is the part of the AI story that doesn't make great press releases but determines everything about who actually wins.

You can build the world’s smartest model, hire the best researchers, and design the perfect transformer architecture and then your AI has absolutely nowhere to live.

Training a single frontier AI model like GPT-5 or Gemini Ultra requires tens of thousands of the most expensive chips on earth running continuously for months.

One rack of NVIDIA's latest GB200 NVL72 GPUs consumes 120 to 130 kilowatts of electricity.

To put that into perspective, your entire house uses roughly 1.2 kilowatts to power everything at once.

A single AI GPU rack consumes more electricity than 100 homes, while a cluster of 10,000 GPUs, a modest setup by frontier model standards, needs 35 to 50 megawatts of continuous power.

That's enough electricity to power 35,000 homes, running nonstop, just to teach a machine how to write a cover letter.

Now here's where it gets painful.

The buildings that can handle this kind of power density essentially do not exist at the scale that's needed.

Most data centers were built for a world where a server rack consumed 10 to 15 kilowatts, but the new AI workloads need 10 to 15 times that. 

You can't just retrofit a regular building any more than you can plug a Tesla Supercharger into a 1950s wall outlet.

You need liquid cooling, massive electrical infrastructure, years-long utility agreements, and the right land near fiber and cheap power.

Oh, and every one of the 50 other companies trying to do the exact same thing is competing with you for all of those resources simultaneously.

This is the gap WhiteFiber is printing money into.

So who actually is WhiteFiber?

WhiteFiber has a market cap of roughly $867M and generated $82.7M in trailing twelve-month revenue, up 49% year over year.

You've probably never heard of it, and that's the point.

WhiteFiber was carved out of a company called Bit Digital, a former Bitcoin mining operation that looked at the AI infrastructure buildout in late 2023 and made a decision that, in retrospect, looks like one of the smarter pivots in recent memory.

They rebranded their data center and GPU infrastructure business as WhiteFiber, took it public in August 2025 at $17 per share, the top of their range, raised $159M, and got to work building what I'm going to argue is the landlord business of the AI era.

The business model is simple enough to explain to your grandmother, which is either a great sign or a deeply ironic one, given the complexity of what they actually have to build.

WhiteFiber either rents GPU clusters by the hour to AI companies that need compute and don't want to buy $150,000 chips they'll only use 60% of the time, or signs long-term leases with companies that want to park their own hardware somewhere that can actually handle the power draw.

Their segment one is the GPU cloud business. 

They buy NVIDIA H100s, B200s, and GB200s, deploy them in data centers, and rent them out at $2.50 to $2.55 per GPU hour on contracts that typically run 12 to 24 months. 

Gross margins on this business are roughly 65%, which is exceptional for an infrastructure company.

Segment two is colocation. 

They build the building, handle the power, handle the cooling, and the tenant brings their own servers, and WhiteFiber charges per megawatt of IT load delivered. 

The tenant pays the electricity and property taxes as a pass-through and they collect rent.

Right now, they have facilities operating in Iceland, two campuses outside Montreal, and one very important campus in Madison, North Carolina, with a leased expansion site in Atlanta ramping up this year.

But to understand the real story, you have to understand one specific piece of real estate.

THE AI BOOM NEEDS LANDLORDS (P2)

There is a 96-acre plot of land in Madison, North Carolina, and it’s called NC-1.

And here are the facts.

One million square feet of AI data center space, 54 megawatts of live power today, a secured 99-megawatt deal with Duke Energy, and infrastructure built for ultra-dense AI workloads.

The anchor tenant is a company called Nscale Global Holdings.

In November 2025, Nscale signed a 10-year lease worth $865M in total contract value, with 3% annual rent escalators built in, so the check gets bigger every year regardless of what happens in the AI market.

At full Phase 1 and Phase 2 billing, the annualized revenue from this single contract exceeds $87M per year.

WhiteFiber's entire 2025 revenue was $79.5M.

One contract is more annual revenue than the whole company made last year, and NC-1 still has 60 megawatts of capacity left after Nscale's 40 megawatts are filled.

Here is the number that should genuinely make you stop scrolling.

As of March, WhiteFiber reported remaining performance obligations, the fancy accounting term for "money we are contractually owed” of approximately $921M, and their trailing twelve-month revenue is $83M.

They have contracted future revenue equal to more than 11 times what they currently make in a year.

For context, most companies of this size would throw a party if their backlog was three times annual revenue, WhiteFiber is sitting on 11 times.

The Iceland gem nobody is talking about

Before we get to the risks, there's one part of this story the analyst reports keep skipping over that I think deserves more attention.

WhiteFiber has a facility in Iceland.

An operational data center running NVIDIA B200 and GB200 GPU clusters, powered 100% by renewable hydroelectric electricity from the Blanda Hydropower Station.

Why does this matter? Because Iceland has two things that the rest of the world's data center market would kill for: extremely cheap electricity and free cooling.

The average temperature in Iceland means you don't need air conditioning for the majority of the year.

For GPU clusters that throw off enormous heat, it's a structural cost advantage that compounds at scale.

And WhiteFiber deployed something even more interesting there, a networking architecture from a company called DriveNets that uses a single Ethernet fabric for both GPU-to-GPU communication and storage networking simultaneously.

This matters because in most large GPU clusters, the network connecting the chips together is the bottleneck. 

GPUs sit idle waiting for data, and WhiteFiber's Iceland setup reduces that bottleneck while also enabling multi-tenancy, meaning multiple customers can share the same infrastructure without stepping on each other's workloads.

As the AI market shifts from training, which is expensive and clustered around a few frontier labs, toward inference, which is happening everywhere all the time for everything, Iceland becomes more strategically valuable.

Gartner estimates that by 2027, inference will account for 90% or more of enterprise AI compute spending, and Iceland is WhiteFiber's inference fortress. 

Also, they have a Cerebras deployment.

WhiteFiber is one of the only infrastructure providers on the planet currently hosting Cerebras Systems' CS-3 wafer-scale processors.

Cerebras is doing something fundamentally different from NVIDIA.

Instead of linking thousands of small chips together, Cerebras built one massive wafer-sized chip with ultra-fast, low-latency inference.

This is deployed at MTL-3, WhiteFiber’s Montreal campus, on a 5-year contract, giving them exposure to an alternative high-performance AI inference architecture alongside NVIDIA systems.

If Cerebras gains real traction in AI inference, WhiteFiber already has the infrastructure in place, giving it meaningful upside optionality for a company of its size.

The bear case gets uncomfortable

Here is the part of the story I'm legally and morally obligated to tell you.

Because anyone who reads the bull case on WhiteFiber and doesn't immediately feel a cold sweat forming hasn't looked at the balance sheet.

The risks are real, and you should take them seriously.

Let's start with the one that almost gave me a headache when I read the 10-K.

One unnamed customer represented 70.7% of WhiteFiber's 2025 total revenue.

That is a single counterparty that, if they sneeze, WhiteFiber catches pneumonia, and that customer has paused services.

Renegotiation is ongoing but no definitive resolution as of the most recent report date.

A second customer, DNA Fund, representing 11.5% of 2025 revenue, had their contract terminated. 

They still owe WhiteFiber $7.3M, and only $2.1M has been collected.

So let's be completely clear about what's happening in the cloud segment simultaneously with the glorious NC-1 narrative:

The two largest cloud customers, representing more than 80% of prior cloud revenue, are either paused, terminated, or delinquent.

The company states they have redeployed GPUs to replacement customers, including a $17M contract with a company called Hyperbolic, but the math on whether that fully covers the gap is not yet confirmed.

WhiteFiber's cash burn surged 270% in the prior year, totaling approximately $223M, against a December 2025 cash balance of $114M.

Do the math on that runway, and it gets uncomfortable fast.

They subsequently raised $230M in convertible senior notes at 4.5%, maturing in 2031, to stay capitalized.

After paying $120M for a hedging structure to protect against dilution from the conversion, they ended up with roughly $101M in net new cash available.

NC-1 has already consumed $150M in equity investment, and the project-level credit facility expected to fund the next phase of construction had not yet been confirmed as of Q1 2026 earnings.

They also deferred construction of MTL-2, their fourth Montreal campus, to preserve capital.

When a company starts deferring planned projects to preserve cash, that is worth paying attention to.

Here is the last bear case item, and it's the one that keeps the whole risk picture honest.

There are currently more than 100 GPU cloud providers operating globally, and analysts and industry observers expect a major shakeout in 2026 and 2027.

The winners will control 80% or more of the market, and the losers will either get acquired at distressed valuations or simply run out of money.

CoreWeave is at $54B, Nebius is at $50B, and WhiteFiber is at $867M.

If WhiteFiber cannot secure additional capital, maintain GPU supply relationships with NVIDIA, and close new contracts at NC-1 beyond Nscale, it risks being a subscale participant in a winner-take-most market.

The honest probability-weighted verdict

So what do you actually do with all of this? Here is how I think about the range of outcomes, honestly.

There is about a 20% chance this becomes a transformational winner. 

NC-1 executes flawlessly, new tenants fill the remaining 60 megawatts, the cloud customer situation resolves without catastrophic revenue loss, and WhiteFiber re-rates toward the multiples Applied Digital trades at. 

In that world, you're looking at a stock that could be at $80 to $150 in three to five years from today's $22 to $25.

There is about a 30% chance this becomes a solid compounder. NC-1 delivers roughly on schedule, cloud headwinds are partially absorbed by new contracts, revenue reaches $300M by 2028, and the stock grinds to $40 to $80. 

Not glamorous, but a very good return from current prices.

There is about a 30% chance of a muddle-through scenario.

NC-1 gets delayed six to twelve months, the cloud segment stagnates at current levels, they need to raise additional dilutive equity, and the stock spends two years between $15 and $35 while execution catches up to the narrative.

There is a 15% chance of significant disappointment. 

The Initial Customer exits entirely, NC-1 financing takes longer than expected, GPU price deflation crushes cloud margins, and the stock is at $5 to $15.

And there is a 5% catastrophic scenario where multiple things go wrong simultaneously, and this becomes an expensive lesson.

After all of that, here is what I keep coming back to.

Jensen Huang, the CEO of NVIDIA, has projected over $1T in AI chip revenue from the Blackwell and Rubin GPU architectures between 2025 and 2027.

Every single one of those chips needs a physical home.

The global AI data center market is projected to grow from $236B in 2025 to $934B by 2030.

That is a 31.6% compounding annual growth rate for five straight years.

The neocloud GPUaaS market alone is expected to reach $250B by 2030.

By 2035, experts estimate that new AI infrastructure may require power equivalent to 125 nuclear reactors.

The demand for AI compute is not going anywhere and there are very, very few people who own the buildings.

WhiteFiber has real assets, a 96-acre NC-1 campus, a 99-megawatt Duke Energy agreement, $921 million in contracted revenue, operational facilities across four sites, and a 1,500 megawatt development pipeline they're actively evaluating.

It also has real risks such as customer concentration, cash burn, hyperscaler competition, and a cloud segment in active disruption.

That balance between asymmetric upside and very real execution risk is exactly why I find this story interesting.

Personally, I will likely add this to my watchlist here and alert our Pro members inside Discord if I do end up taking a position. 

The question is not whether the AI infrastructure market is real.

It is whether WhiteFiber specifically is the company that captures a meaningful piece of it before running out of runway or getting crushed by a larger competitor.

The next two earnings reports, Q2 2026 and Q3 2026, are the most important data points in this story.

If Nscale billing has commenced, if the Initial Customer situation resolves cleanly, and if the NC-1 project-level financing closes, you have the foundation of a multi-year compounder.

If those three things don't happen on schedule, the risk/reward deteriorates fast.

Alright, that's it for this edition of Milk Road AI. We want to hear from you.

BITE-SIZED COOKIES FOR THE ROAD 🍪

Anthropic snapped up Stainless, the dev tools startup behind OpenAI & Google SDKs, for $300M+. Rivals lose access to the SDK builder they relied on, Anthropic just locked in a key moat.

OpenAI is lawyering up against Apple over a buried ChatGPT deal gone cold. Big Tech "partnerships" have fine print, and OpenAI wants its promised subscribers back.

Google I/O 2026 kicked off yesterday with Gemini 4.0, XR glasses, and AI agents. Google is betting its entire Android future on Gemini being everywhere, all at once.

MILKY MEMES 🤣

ROADIE REVIEW OF THE DAY 🥛