Wall Street spent the first phase of the AI boom asking one question:
Who can make enough GPUs?
Then the bottleneck moved.
The chips needed data centres.
The data centres needed electricity.
The electricity needed generation, transformers, switchgear, substations, cooling and grid connections.
Now I think we may be entering another phase.
Governments and electricity regulators are increasingly asking:
If AI creates the infrastructure problem, who should pay to solve it?
And the emerging answer may be:
AI itself.
That sounds like a regulatory story.
I think it could become an industrial investment story.
Because if hyperscalers increasingly have to fund the physical infrastructure required to support their own electricity demand, then the AI capex boom may spread much further beyond semiconductors than the market initially expected.
The data centre could start looking less like a building full of servers.
And more like a private utility with GPUs attached.
⸻
🇦🇺 VICTORIA JUST PROVIDED THE LATEST RECEIPT
On 22 September, Victoria announced a new framework for data-centre development.
New facilities will be expected to match their electricity requirements with additional renewable generation and storage.
Developers will have to pay their own grid connection and necessary network-upgrade costs.
They will face requirements around recycled or non-drinking water for cooling.
And tighter planning rules will affect where facilities can actually be built.
Why is this happening?
Because the load is becoming too large to ignore.
Data centres currently consume roughly 5 TWh annually across Australia’s National Electricity Market.
By 2035-36, that is projected to reach approximately:
34 TWh
Their share of total NEM electricity consumption could rise from roughly:
3%
to around:
13%.
At that scale, AI infrastructure stops being just another electricity customer.
It starts affecting the planning of the electricity system itself.
⸻
🇺🇸 TEXAS IS SOLVING A DIFFERENT VERSION OF THE SAME PROBLEM
Texas has been flooded with enormous proposed data-centre and industrial electricity loads.
But there is a problem.
Not every announced gigawatt is real.
Some projects may appear multiple times in connection queues.
Some lack financing.
Some lack committed customers.
Some may never be built.
So Texas has started making large loads prove they are serious.
For projects of at least 75 MW, developers now face financial security requirements of:
$50,000 PER MW
plus a $100,000 study fee, site-control requirements and additional operating disclosures.
For a hypothetical 1 GW project, the security requirement alone becomes:
$50 MILLION
before we even start talking about building the actual data centre.
That distinction matters.
Because investors love counting announced gigawatts.
The grid needs to know which gigawatts will actually show up.
⸻
POWERPOINT GIGAWATTS VS REAL GIGAWATTS
This may become one of the most important distinctions in the AI infrastructure boom.
A developer can announce:
2 GW!
Fantastic.
But before those 2 GW produce a single AI token, the project has to survive:
FINANCING
↓
GRID CONNECTION
↓
GENERATION
↓
TRANSFORMERS
↓
SWITCHGEAR
↓
COOLING
↓
WATER
↓
PERMITTING
↓
COMMUNITY ACCEPTANCE
↓
ENERGISATION
Only then do the GPUs start earning money.
So the scarce AI asset may not actually be a gigawatt.
It may be a permitted, financed, equipped and energised gigawatt.
⸻
AND AMAZON ISN’T WAITING AROUND
This is where the corporate behaviour becomes interesting.
Amazon recently entered a long-term agreement with Generac involving approximately:
$2.4 BILLION
of expected initial generator deliveries across 2027 and 2028.
The agreement also contains warrants that can vest progressively as qualifying Amazon purchases increase.
That does not mean Amazon has committed to the maximum potential purchase amount.
But strategically, the message is difficult to miss.
A hyperscaler is locking down physical power equipment years ahead.
Why?
Because a GPU without reliable electricity is not an AI asset.
It’s extremely expensive furniture.
⸻
THE AI SUPPLY CHAIN IS GETTING WEIRD
We traditionally thought about AI infrastructure like this:
NVIDIA → SERVER → DATA CENTRE
But the real chain increasingly looks more like:
GPU
↓
CPU + NETWORKING + MEMORY
↓
LAND
↓
ELECTRICITY
↓
GENERATION
↓
TRANSFORMERS
↓
SWITCHGEAR
↓
SUBSTATION
↓
BATTERY STORAGE
↓
BACKUP POWER
↓
COOLING
↓
WATER
↓
GRID CONNECTION
↓
PERMIT
↓
COMPUTE
Suddenly companies that looked completely unrelated to artificial intelligence are sitting inside the AI supply chain.
⸻
THE SECOND-ORDER THESIS
The market sees:
AI NEEDS MORE POWER
and immediately thinks:
build more power stations.
I think the deeper consequence could be:
AI DEVELOPERS INCREASINGLY HAVE TO BUILD OR FINANCE THEIR OWN INFRASTRUCTURE ECOSYSTEM.
That potentially pushes hyperscaler capex into:
⚡ Generation
🔋 Battery storage
🔌 Transformers
⚙️ Switchgear
🏭 Substations
🔥 Behind-the-meter generation
❄️ Liquid cooling
💧 Water infrastructure
🧠 Power-management software
🏗️ Construction
💰 Project finance
AI companies thought they were building compute.
They may increasingly find themselves building industrial infrastructure.
⸻
POWER NOW VS POWER FOREVER
There is another market emerging inside this shortage.
Speed-to-power.
A hyperscaler may know exactly what its ideal long-term electricity solution looks like.
Maybe it’s nuclear.
Maybe it’s a dedicated gas plant.
Maybe it’s renewables plus storage.
Maybe it’s a huge utility agreement.
But that infrastructure can take years.
The GPUs are arriving now.
And billions of dollars of accelerators sitting idle for three years waiting for the perfect electricity solution could be economically disastrous.
So developers are increasingly exploring interim solutions:
modular gas generation
existing grid supply
fuel cells
batteries
microgrids
temporary or bridge-power contracts
This creates an interesting equation.
The relevant question isn’t always:
What is the cheapest electricity?
It may be:
What is electricity available two years earlier worth?
If expensive bridge power allows billions of dollars of GPUs to begin generating revenue immediately, comparing only the cost per megawatt-hour misses the economic point.
Speed-to-power itself may be becoming a product.
⸻
THEN THERE ARE THE OLD FACTORIES
This is one of my favourite second-order consequences.
If new grid connections become expensive, slow and difficult to permit, what happens to old industrial sites that already possess enormous electrical connections?
Former:
aluminium smelters
coal plants
steelworks
paper mills
Bitcoin mines
may suddenly look very different.
The building might be obsolete.
The industrial process might be dead.
But the substation is still there.
The transmission infrastructure may already exist.
The land may already be industrially zoned.
There may already be gas, water and fibre nearby.
Suddenly yesterday’s stranded industrial asset becomes potential AI real estate.
The valuable thing isn’t necessarily the factory.
It’s the invisible electrical plumbing underneath it.
⸻
THE CRYPTO CONNECTION IS EVEN STRANGER
Bitcoin miners spent years searching the world for:
cheap electricity
large industrial sites
grid connections
substations
permitting
land
Then AI arrived and discovered electricity was becoming one of its scarcest inputs.
The ASICs themselves may have little value for AI.
But the powered land?
That could be enormously valuable.
Which creates one of the strangest consequences of the entire AI boom:
THE CRYPTO BOOM MAY HAVE ACCIDENTALLY FINANCED PART OF THE PHYSICAL INFRASTRUCTURE REQUIRED FOR THE AI BOOM.
Same megawatt.
Different digital economy.
⸻
BUT HERE IS THE PART THAT COULD BREAK THE THESIS
This is not automatically bullish for everything containing copper.
There is a serious bear case.
If developers are forced to pay for more of their own infrastructure, some projects simply won’t make economic sense.
A proposed 2 GW campus could become 1 GW.
A speculative development could disappear entirely.
AI efficiency could improve dramatically.
Models could require fewer resources.
Custom chips could improve inference economics.
Battery and demand-response systems could allow more compute to fit onto the existing grid.
Manufacturing capacity could expand enough to reduce today’s transformer and turbine shortages.
And enormous interconnection queues almost certainly overstate the amount of demand that will ultimately be built.
So I would not value an industrial company based simply on:
AI POWER DEMAND GOES UP = STOCK GOES UP.
The harder question is:
Which companies capture spending from projects that actually survive the funnel?
⸻
THE RESEARCH UNIVERSE
Not recommendations.
But these are some of the names I think deserve attention as the physical AI stack expands:
GEV, generation and grid equipment.
ETN, electrical distribution, switchgear and power management.
VRT, data-centre power and thermal infrastructure.
GNRC, distributed and backup generation.
BE, behind-the-meter power.
FLNC, battery storage and energy management.
CEG, firm electricity generation.
NRG, generation and increasingly data-centre exposure.
And then there are less obvious companies involved in modular power, logistics, water infrastructure and powered industrial land.
That is where I suspect some of the more interesting second-order research still sits.
⸻
WHAT WOULD CHANGE MY MIND?
I’d weaken this thesis if:
Grid interconnection times fall dramatically.
Transformer and switchgear lead times normalise rapidly.
AI electricity forecasts begin falling materially.
Hyperscaler capex slows sharply.
Governments retreat from forcing developers to internalise infrastructure costs.
Behind-the-meter generation proves uneconomic or impossible to permit.
Or AI efficiency improves so quickly that infrastructure investment consistently undershoots current expectations.
And there is one test above everything else:
ORDERS.
Press releases are easy.
Backlogs, signed contracts, deposits, equipment deliveries and energised campuses are harder to fake.
⸻
WHAT I’M WATCHING NEXT
I want to know:
How much hyperscaler capex migrates from chips into electrical infrastructure?
Who signs the next Amazon-style equipment agreement?
Do developers begin reserving transformers and generators years ahead?
Do more data centres build their own generation?
Do utilities charge increasingly large deposits for grid access?
Do powered industrial sites command premiums?
Does water availability begin determining AI geography?
And perhaps most importantly:
WHAT DOES IT COST TO PRODUCE ONE ENERGISED MEGAWATT OF AI CAPACITY?
Because announced gigawatts are easy.
Energised gigawatts are fucking hard.
⸻
THE BIGGER IDEA
AI started as software.
Then it became a semiconductor story.
Then a data-centre story.
Then an electricity story.
Now it is becoming something much larger.
AI MAY BE TURNING INTO ONE OF THE BIGGEST PRIVATE INDUSTRIAL INFRASTRUCTURE BUILDOUTS OF OUR GENERATION.
And if governments increasingly decide that ordinary electricity customers shouldn’t subsidise that buildout, hyperscalers may have to reach much deeper into their own pockets.
Wall Street keeps asking:
WHO MAKES THE BEST AI CHIP?
Maybe the next question is:
WHO SELLS EVERYTHING REQUIRED TO TURN IT ON?
Because the next scarce AI resource may not be the GPU.
It may be a permitted, financed, equipped and energised megawatt.
🐯⚡🚀
What do you think? Is power infrastructure becoming the next major layer of the AI trade, or is Wall Street already pricing too much growth into the companies supplying it?
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.
Comments