To The Moon

The market thinks AI has a power problem.

I think that’s becoming outdated.

AI increasingly has a bottleneck migration problem.

When the grid became too slow, hyperscalers started looking behind the meter.

Problem solved?

Not quite. 😅

Because building your own electricity means you suddenly need:

⚙️ TURBINES

🔋 GENERATORS

🔥 GAS

🛢️ PIPELINES

⚡ TRANSFORMERS

🔌 SWITCHGEAR

🔋 BATTERIES

❄️ COOLING

💰 FINANCING

📋 PERMITS

👷 LABOUR

And something fascinating is beginning to happen.

🚨 EVERY TIME AI ESCAPES ONE BOTTLENECK, IT RUNS STRAIGHT INTO THE NEXT ONE.

📸 

⚡🏗️ THE FIRST ESCAPE: BUILD YOUR OWN POWER

For years, the data-centre model was relatively straightforward:

UTILITY GRID → DATA CENTRE → COMPUTE

AI broke that model.

Independent research from Cleanview estimates Amazon has 44 GW of US data-centre capacity under development across 16 states.

But here’s the extraordinary part.

Cleanview identified:

🔥 17.5 GW of new gas generation being built to serve Amazon

🌱 25.6 GW of contracted carbon-free electricity

⚛️ 3.12 GW of purchased nuclear capacity

Those numbers are reconstructed from hundreds of project records rather than Amazon guidance, so “under development” should not be confused with guaranteed operating capacity.

But look at what the physical infrastructure is telling us.

Amazon isn’t simply buying electricity anymore.

⚡ IT IS ASSEMBLING A POWER PORTFOLIO.

And Amazon isn’t alone.

The AI race is turning technology companies into some of the world’s most aggressive buyers of physical energy infrastructure.

🚧 THEN THE WORKAROUND HIT ANOTHER BOTTLENECK

So build gas generation.

Easy.

Except everyone else had the same idea.

GE Vernova reported that its gas-power equipment backlog and slot reservations reached approximately 116 GW in Q2, up from 100 GW just one quarter earlier.

Its total backlog reached approximately $176 billion.

And data-centre-related Electrification orders exceeded $5 billion in the first half, already more than double the company’s entire 2025 total.

GE Vernova has also been expanding annual gas-turbine output toward:

⚙️ 20 GW in 2026

⚙️ 24 GW in 2028

⚙️ 30 GW in 2030

Think about what that means.

AI developers tried to escape the grid queue.

🚨 THEY CREATED A TURBINE QUEUE.

The scarce asset isn’t necessarily just the turbine anymore.

It can be your place on the manufacturing calendar. 📅

🔒⚡ SO DEVELOPERS ARE RESERVING MEGAWATTS BEFORE THEY EVEN NEED THEM

This is where it gets properly interesting.

Energy Vault announced it had secured and financed 275 MW of Rolls-Royce MTU reciprocating-engine generation capacity for delivery across 2027 and 2028.

Why reserve generators before deployment?

Two constraints:

Equipment availability + access to capital.

Energy Vault is discussing multi-gigawatt AI projects and wants generation capacity ready when customers need it.

That’s a major change.

Hyperscalers aren’t the only ones reserving infrastructure anymore.

Companies are beginning to reserve future megawatts themselves.

Energy Vault calls its model Powered Land:

🏗️ LAND

⚡ GRID ACCESS

🔥 GENERATION

🔋 BATTERIES

🔌 ELECTRICAL INFRASTRUCTURE

🧠 CONTROLS

Essentially:

DON’T WAIT FOR THE POWER. BUY A SITE WHERE THE POWER SOLUTION ALREADY EXISTS.

Speed-to-power is becoming a product. ⚡⏱️

🔥 BUT BUILD YOUR OWN POWER AND YOU NEED SOMETHING ELSE…

Fuel.

And now we get one of my favourite receipts from this entire rabbit hole.

Woodway Energy Infrastructure announced definitive agreements to build, own and operate a:

🛢️ 22-MILE, 30-INCH NATURAL-GAS PIPELINE

specifically providing dedicated gas transportation to behind-the-meter generation for a hyperscale data-centre development.

Not an existing pipeline that happens to pass a data centre.

A dedicated pipeline supporting the power system serving it.

The AI infrastructure chain has now physically moved upstream:

🤖 GPU

↓

🏢 DATA CENTRE

↓

⚡ ELECTRICITY

↓

⚙️ GENERATOR

↓

🔥 NATURAL GAS

↓

🛢️ PIPELINE

AI started by renting servers.

Then it started building data centres.

Then power plants.

NOW THE SUPPLY CHAIN CAN EXTEND ALL THE WAY TO THE PIPE FEEDING THE POWER PLANT.

But surely that solves it?

😂

Not so fast.

🛑📋 WELCOME TO THE PERMISSION QUEUE

Texas has moved to restrict new data-centre approvals while regulators examine the industry’s electricity and water impacts.

The scale of the development queue helps explain why.

More than 470 GW of large-load projects have reportedly sought Texas grid connections.

Important caveat:

🚨 470 GW QUEUED DOES NOT MEAN 470 GW GETS BUILT.

There will be speculative and duplicative applications.

But that actually improves the thesis.

Because the next phase may increasingly separate:

POWERPOINT GIGAWATTS 📊

from

BANKABLE GIGAWATTS. ⚡

Projects surviving the funnel need more than a render and some land.

They need:

💰 Capital

⚙️ Equipment

⚡ Power

💧 Water

📋 Permits

👷 Labour

🏘️ Community acceptance

🇦🇺⚡ AND THIS ISN’T JUST TEXAS

Something remarkable happened here in Victoria.

The Victorian Government’s Sustainable Data Centre Action Plan says new data centres will be required to:

“POWER THEMSELVES.”

Developers must bring their own renewable electricity supply and storage, while further requirements cover infrastructure impacts and community consultation.

Victoria already has more than 50 operating data centres, while Melbourne represents roughly 6% of the Asia-Pacific development pipeline.

So “bring your own power” isn’t merely becoming a clever workaround.

🚨 IN SOME MARKETS IT IS BECOMING POLICY.

And that changes the economics again.

The hyperscaler isn’t merely financing:

GPUs + BUILDING

It’s increasingly financing some combination of:

GPUs + BUILDING + GENERATION + STORAGE + GRID EQUIPMENT + WATER INFRASTRUCTURE + COMMUNITY OBLIGATIONS

💰⚡ The hyperscaler balance sheet is becoming an infrastructure balance sheet.

📸 

🐯 THE REAL THESIS

This is where I think the market is looking at AI infrastructure incorrectly.

Everyone wants to identify the bottleneck.

GPUs.

Then power.

Then transformers.

Then turbines.

But there may never be one permanent bottleneck.

🚧 THE BOTTLENECK MOVES.

Solve grid access with onsite generation?

➡️ You need turbines.

Turbines booked out?

➡️ Use reciprocating engines.

Build enough engines?

➡️ You need gas.

Secure the gas?

➡️ You need pipelines.

Build the pipeline?

➡️ You need permits.

Secure everything?

➡️ You still need transformers, switchgear, cooling, batteries, financing and skilled labour.

And every scarcity creates an incentive for someone to engineer around it.

That’s the game. ♟️⚡

🔍💰 THE SECOND-ORDER TRADE

The biggest AI infrastructure winners may therefore not simply be companies exposed to more megawatts.

They may be companies positioned at whichever constraint becomes hardest to bypass.

My research universe:

⚙️ $GEV: gas turbines, grid equipment, electrification

🔌 $ETN: switchgear, electrical distribution, power management

❄️ $VRT: power and thermal infrastructure inside the AI factory

⚡ $GNRC: distributed generation and onsite/backup power

🔥 $BE: onsite fuel-cell generation and potential DC-native architecture

🔋 $NRGV: powered-land and integrated onsite-power infrastructure

Further upstream, gas and midstream infrastructure become worth watching where direct data-centre contracts actually emerge.

Not every company benefits equally.

That’s precisely the point.

🐂📈 THE BULL CASE

AI capex keeps accelerating.

Grid construction remains slower than compute deployment.

Hyperscalers continue moving behind the meter.

Equipment lead times stay long.

Developers reserve generation years ahead.

Utilities and governments increasingly require AI projects to fund more of their own infrastructure.

Under that scenario, scarcity economics keep migrating through the physical AI supply chain.

The opportunity becomes identifying the next gate before Wall Street does. 👀

🐻📉 THE BEAR CASE

There’s a serious counterargument.

Scarcity creates supply.

GE Vernova is expanding turbine production.

Manufacturers are adding transformer capacity.

New generation technologies are emerging.

AI models and chips keep becoming more energy-efficient.

Flexible compute could shift workloads away from constrained grids.

And technologies such as native DC architectures could eventually remove pieces of today’s transformer and power-conversion stack.

That’s important.

🚨 TODAY’S BOTTLENECK WINNER CAN BECOME TOMORROW’S DISINTERMEDIATED SUPPLIER.

There’s another risk.

The gigantic data-centre queues may substantially overstate genuine demand.

If regulators force developers to prove financing and infrastructure commitments, hundreds of proposed gigawatts could disappear.

That would reduce future equipment demand.

🚨 WHAT WOULD CHANGE MY MIND?

I’d weaken this thesis if:

❌ Grid interconnection times collapse

❌ Turbine and transformer lead times normalise quickly

❌ Equipment capacity grows materially faster than AI demand

❌ Behind-the-meter projects fail economically

❌ Hyperscalers materially reduce capex

❌ AI efficiency improvements overwhelm growth in compute demand

Or most importantly:

❌ Announced gigawatts repeatedly fail to become contracted equipment orders.

That’s why I’m no longer interested in simply counting press-release gigawatts.

I want to count:

⚡ FINANCED, PERMITTED, EQUIPPED AND ENERGISED MEGAWATTS.

👀 WHAT I’M WATCHING NEXT

⚙️ Turbine delivery slots

🔌 Transformer lead times

💰 Hyperscaler equipment reservations

🔥 Behind-the-meter generation contracts

🛢️ Dedicated gas infrastructure

🔋 Battery and microgrid deployments

📋 Large-load financial-security requirements

🛑 Permitting delays

And the most important metric of all:

⏱️⚡ TIME TO POWER.

Because an AI campus with 100,000 GPUs and no electricity isn’t an AI campus.

It’s a very expensive warehouse. 😂

📸 

🧠⚡ THE BIGGER IDEA

The AI infrastructure race started with one question:

Who has the best GPU?

Then:

Who can get enough GPUs?

Then:

Who can find enough electricity to turn them on?

We’re entering the next phase.

🚀 WHO CAN ASSEMBLE THE ENTIRE PHYSICAL SYSTEM FAST ENOUGH TO TURN COMPUTE INTO REVENUE?

Amazon’s reported power portfolio spans gas, nuclear and renewables.

GE Vernova has roughly 116 GW of gas equipment and reservations in backlog.

Energy Vault is financing generators before customers deploy them.

A dedicated 22-mile gas pipeline is being built for hyperscale behind-the-meter power.

Texas is tightening the development gate.

Victoria is telling new data centres to bring their own power.

These aren’t six separate stories.

🧩 THEY’RE ONE STORY.

AI hasn’t solved its power problem.

It has pushed the problem deeper into the physical economy.

And that’s why I think the most important question in AI infrastructure has changed:

🐯⚡ WHEN EVERY SOLUTION CREATES ANOTHER BOTTLENECK, WHERE DOES THE BOTTLENECK MOVE NEXT? 👇🔥

ADZ5150 🐯

⚠️ Disclaimer: This post reflects my own personal research, opinions and interpretation of publicly available information. It is shared for discussion and educational purposes only and is not financial advice or a recommendation to buy or sell any security. I may hold positions in companies mentioned. Always do your own research and consider your own financial circumstances and risk tolerance before making investment decisions.