Big Tech is on pace to spend $650 billion on compute infrastructure in 2026. That's up 80% from last year's record.

Let that marinate.

OpenAI has asked the U.S. government to build 100 nuclear reactors’ worth of power capacity every year just to keep the lights on. Not over a decade. Every. Single. Year.

Something has to give. Spoiler: several things already have. Simultaneously.

Here's the thing: this isn't a tech story. It's a resource scarcity story dressed up in a Patagonia vest, pitching Series G, and it has a very investable conclusion.

Data center electricity demand is projected to surge from 460 TWh in 2024 to 1,300 TWh by 2035. That's not a growth curve. That's a vertical line with a prayer attached.

Solar and wind can't deliver a 24/7 baseload. You know it. The hyperscalers know it. So they did what hyperscalers always do. They wrote very large checks.

The Nuclear Scoreboard. All four majors are in:

Translation: the same companies that built their own chips are now building their own power plants. Vertical integration isn't a strategy anymore. It's a survival instinct.

ERCOT estimates Texas alone must double its 2024 power production by 2031 just to meet projected data center demand. Double. In seven years. In a state that can't keep the lights on during a cold snap.

These aren't speculative bets. These are 20-year PPA structures. Defensive cash flows locked in at scale. The kind of contracts that make CFOs weep with joy and competitors weep with everything else.

Most readers assume data centers just eat electricity. Reasonable assumption. Also wrong.

The hidden variable is cooling water. Hyperscalers deliberately traded electricity costs for water consumption. Evaporative cooling. Cheaper to run. Brilliant in 2018.

Then AI rack densities hit 60–120 kW per rack. Traditional racks run 5–10 kW. Do the math. Or don't. The water bill will take care of it for you.

Texas. The case study nobody wanted.

Let's review:

Read that list again. Slowly.

The state has a water crisis. It doesn't account for data centers in its water plan. And there are no limits on how much data centers can drink. This is the planning equivalent of building a swimming pool on top of a sinkhole.

Reality Check: This is an architectural choice, not a law of physics. Switch-style air-cooled facilities don't have this problem. Companies chose the cheap cooling option, and now the bill is coming due. In drought-prone Texas. During a megabuild. Brilliant.

And here's the kicker: Microsoft pledged zero water consumption on the data center side. Noble. Inspirational. They're also buying power from Three Mile Island. A pressurized water reactor. As in, a reactor that runs on water.

Irony doesn't get more structural than that.

Before you crown nuclear the clean savior of the AI age, let's talk about its dirty little secret. Water.

Traditional light-water reactors consume 400–720 gallons per MWh. Comparable to coal. Not exactly the "clean energy" brochure you were picturing.

The critical nuance: legacy plants withdraw enormous volumes but return roughly 99%. The issue is thermal pollution, not depletion. Hot water is being returned to rivers and lakes. The fish are not thrilled.

Now the plot twist. The reactors these hyperscalers are actually buying? They're not water-cooled:

These Gen IV designs break the nuclear-water trade-off entirely. That's not incremental improvement. That's a different physics.

The verdict: legacy nuclear in drought zones equals real stress. Next-gen SMRs equal the engineered solution to both problems at once. The hyperscalers aren't buying yesterday's reactors. They're buying tomorrow's. For once, the long bet might actually be the smart one.

Remember when chip shortages were a pandemic thing? Cute. That was the warm-up act.

IDC is calling the current memory shortage "a crisis like no other." Not "a challenge." Not "a headwind." A crisis.

DRAM spot prices have jumped nearly 700% in some categories over the past year. Seven hundred percent. Say it out loud so it lands.

Here's what happened: chipmakers are diverting production to High-Bandwidth Memory (HBM) for AI. One HBM wafer displaces two or more conventional DRAM wafers. The AI gods get fed. Everything else starves.

Memory prices are forecast to rise 40–50% in Q1 2026 alone.

Think about that next time you buy a laptop. Or a smartphone. Or a refrigerator. Or a car. Everything with a chip costs more now. Tim Cook has publicly flagged AI infrastructure costs pressuring Apple's hardware margins. When Apple starts sweating component costs, you know the supply chain is on fire.

PC and smartphone shipments are contracting in 2026 due to supply reallocation. Not weak demand. Supply reallocation. Your phone costs more because a data center in Iowa needed the silicon first.

Translation: this isn't a cycle. It's a structural reallocation. It won't self-correct until new fabs come online. That's 2027 at the earliest. Enjoy that price tag in the meantime.

Pop quiz: what do semiconductor fabs, MRI machines, and fiber optic cables all need? Helium. What just lost a third of its global supply? Also helium.

Iranian strikes on Qatar's Ras Laffan facility on March 4, 2026, took approximately 33% of the global helium supply offline via force majeure. One facility. One-third of the supply. Gone.

Helium is irreplaceable in semiconductor manufacturing. There are no substitutes for wafer cooling, fiber optic production, and MRI shielding. None. Zero. This isn't a "find an alternative" situation. This is a "pray the geopolitics calm down" situation.

Samsung and SK Hynix are holding approximately six months of inventory before yield impacts begin. Six months sounds comfortable until you remember the Ras Laffan facility isn't coming back online on anyone's optimistic timeline.

Spot prices already up 20–50%.

The flip side. The U.S. produces 54% of global helium. America just became the world's swing producer overnight. Not because of innovation. Because of a missile.

Avanti Helium is pushing for a U.S. critical mineral designation for helium. If granted, that opens federal support and institutional capital flows. The kind of policy catalyst that turns a niche commodity play into a front-page trade.

Sound familiar? It should. This is the lithium playbook. Again.

Let's step back. The AI buildout is simultaneously straining:

These aren't independent risks anymore. They're reinforcing constraints on the same buildout. Pull one thread, and the whole sweater unravels. Pull five threads at once, and you're standing in the cold, wondering who greenlit this.

Growth is structurally outpacing supply in every single input category. That's why consumer electronics, cloud services, and utility rates are all moving in the same direction. Up.

NexGen Energy's CEO compared hyperscaler uranium contracting to EV makers locking in lithium. The companies that secured supply early won. Everyone else paid the panic premium.

Here's the thing: the hyperscalers with the deepest pockets will get their share. Meta, Google, Microsoft, Amazon. They'll be fine. They always are.

The asymmetric opportunity isn't in owning the hyperscalers. It's in owning the companies that supply the inputs the hyperscalers can't function without.

Not buy recommendations. Research starting points. Do your own homework. We're not your financial advisor, and this isn't that kind of newsletter.

The AI boom isn't slowing down. The quarterly earnings calls sound great. The capex commitments are staggering. The future is bright.

But the physics of resource supply don't care about earnings calls. They don't care about your pitch deck. They don't read press releases.

The investors who win the next decade won't just own the hyperscalers. They'll own the pipe, the fuel, the gas, and the water that the hyperscalers can't function without.

Every shortage on this list is someone else's moat.

Build a watchlist from Section 8. Watch for catalysts in helium designation, SMR permitting, and ERCOT capacity reports. Those are your entry signals.

The infrastructure tax is here. Someone's going to collect it.

Might as well be you.