Google’s $4.3B Constellation Deal Buys 890 MW—and Makes Power the AI Moat
AI is not being constrained by cleverness anymore. It is being constrained by electricity—and Google just spent 20 years proving that power is now the real competitive advantage.
Google is not worried about running out of AI ideas. It is worried about running out of electricity.
That is why Alphabet has signed a 20-year power deal with Constellation Energy that supports more than $4.3 billion of nuclear investment, adds 890 megawatts of new capacity to the PJM grid and locks in another 2,700 MW of supply from Constellation’s existing fleet. The first new capacity is expected from 2028.
Forget the shiny chatbot demos for a minute. This is the AI race in its grown-up form: whoever secures power, permits, land, chips and grid access gets to keep building while everyone else writes another LinkedIn post about “innovation.”
Google Has Bought the Boring Bit That Matters
On October 6, Google and Constellation announced a long-term clean-energy collaboration centred on 11 nuclear units across Illinois, Pennsylvania and New Jersey.
The headline number is 890 MW of incremental nuclear generation. That is not a vague green pledge or a press-release tree-planting exercise. It comes from upgrading equipment and technology at existing reactors to improve thermal and electrical efficiency. In plain English: get more reliable output from plants that are already operating, rather than waiting a decade-plus for a brand-new nuclear project to crawl through politics, approvals and construction.
Constellation says it will invest more than $4.3 billion. The deal is intended to sustain roughly 4,400 existing jobs and create about 7,200 construction jobs during the work.
There is a second piece that deserves just as much attention. Google has also agreed to a 15-year energy-supply arrangement covering an additional 2,700 MW from Constellation’s PJM fleet. That matters because it gives the generator revenue certainty to keep existing plants economically viable and online.
The distinction matters. The 890 MW is new capacity. The 2,700 MW is about keeping proven, dispatchable capacity in the system. One expands supply; the other stops supply disappearing just as demand goes ballistic.
Google and Constellation have also expanded their technology alliance for five years. Constellation plans to use Google Cloud and Gemini Enterprise in areas including site selection, grid modelling, plant operations and critical-infrastructure security.
Naturally, every company involved will say AI can make all of this faster and cheaper. Fine. Maybe it can. But the far more important fact is that Google is putting real, long-dated commercial backing behind physical generation.
That is the bit to watch.
The AI Boom Has Hit a Very Old Constraint
For years, tech sold us a lovely story: software eats the world, platforms scale infinitely, marginal costs fall toward zero.
That story works brilliantly until your software needs warehouses full of specialised chips consuming industrial volumes of power around the clock.
Training and running frontier AI models is not like launching another lightweight SaaS product. It requires data centres, transmission lines, substations, cooling, chips and reliable electricity. You cannot prompt-engineer your way around a congested grid.
PJM is the grid operator serving 67 million people across parts of the Mid-Atlantic and Midwest. It is precisely the sort of region where data-centre growth is colliding with an ageing energy system, slow interconnection processes and the hard public reality that households do not enjoy paying more because a tech giant needs another cluster.
That is why this deal is more intelligent than simply buying renewable-energy certificates and congratulating yourself. Google is backing actual supply expansion while also agreeing to demand-response and load-shaping measures for non-critical use during high-stress grid events.
That is what responsible growth looks like: bring your own capacity, do not just turn up with a massive load and send the bill to everyone else.
Amazon is moving in the same direction. The Wall Street Journal reported that Amazon recently agreed to help expand output at Constellation’s Calvert Cliffs nuclear site in Maryland by 190 MW, while securing up to 690 MW of long-term power from Constellation’s fleet. Microsoft has tied itself to the restart of Three Mile Island. Meta has also made nuclear-power agreements.
The pattern is obvious. Big Tech is becoming an infrastructure buyer because its next decade of revenue depends on infrastructure it does not control.
The Second-Order Effect: AI Gets More Expensive to Compete In
Here is the uncomfortable bit for founders and investors: the AI market may look open at the application layer, but it is getting brutally concentrated underneath.
Anyone can call a model API. Anyone can build a slick wrapper in a weekend. But very few companies can sign a 20-year nuclear agreement, bankroll grid upgrades, secure multi-gigawatt power supply and negotiate with utilities, regulators and local communities.
That means the biggest AI moat may not be the model itself. It may be the boring collection of contracts beneath it.
Power contracts. Chip allocations. Data-centre leases. Water access. Transmission rights. Long-term relationships with utilities. The willingness to spend capital years before revenue is guaranteed.
This is a familiar pattern in business. The flashy layer gets attention first. Then the real money accrues to whoever controls the bottleneck.
In the gold rush, it was often not the bloke swinging the pickaxe who won. It was the person controlling transport, equipment, finance or land. AI is following the same script. The winners will not merely have smarter models. They will have more dependable access to the inputs that let them keep improving those models when everyone else hits a ceiling.
That should change how investors assess “AI exposure.” A business claiming to be powered by AI is not automatically defensible. Ask what it actually owns. Ask what it can access that competitors cannot. Ask whether its margins survive if model costs, cloud costs or electricity costs rise.
If the answer is “we have a clever prompt,” mate, you do not have a moat. You have a feature with a very short half-life.
The Overlooked Angle: This Is Better Than Waiting for a Miracle
The contrarian point is that the nuclear story is not really about futuristic small modular reactors or some sci-fi moonshot.
It is about squeezing more output from assets already connected to the grid.
New nuclear plants can take decades and cost tens of billions. That does not solve an electricity crunch arriving this decade. Uprates at existing plants are less glamorous, but they are far more useful because the sites, workforce, fuel systems, operating expertise and grid connections already exist.
Operators should take note. Too many people get seduced by the grand transformation plan and ignore the valuable asset sitting in front of them.
The better business question is often not, “What revolutionary new thing should we build?” It is, “What underused asset do we already control, and what would unlock more output from it?”
Google and Constellation are applying that logic at industrial scale.
There is also a political lesson. Communities will tolerate growth more readily when the growth pays for things they can see: reliable supply, durable jobs and a grid that does not become more expensive or less reliable for everyone else. The companies say this structure is designed to add capacity without shifting costs to residential customers.
That claim will be tested in the real world, as it should be. But it is a far better starting point than the usual Silicon Valley posture of arriving late, demanding speed and treating the local consequences as somebody else’s admin problem.
What This Means for You
If you are a founder, stop treating infrastructure as someone else’s problem. You may not need a nuclear contract, obviously. But you do need to know your critical dependencies: cloud concentration, model providers, data rights, payments, logistics, hardware and regulation. Make a one-page dependency map this week. Mark the three things that could halt your product or destroy your margin. Then start reducing exposure before it becomes urgent.
If you are an operator, look for capacity hiding in plain sight. It might be an underused sales channel, an existing customer dataset, a dormant partnership, a process full of manual waste or a piece of equipment producing below potential. Before you launch a shiny new initiative, ask whether you have earned the right to ignore what you already own.
If you are an investor, separate AI beneficiaries from AI tourists. The beneficiaries control scarce inputs, sell indispensable tools or own distribution. The tourists use “AI” as branding while renting everything important from someone else.
And if you are simply trying to get sharper about business, remember this: every boom eventually runs into something physical. Capital. Talent. Supply chains. Power.
Google’s 20-year Constellation deal is a reminder that the next fortune is often made not by predicting the future, but by securing the dull, expensive thing the future cannot work without.