Last week a hyperscaler signed a twenty-year contract for the entire output of a nuclear plant that has not been built yet. The plant will come online in 2032. The contract was signed anyway. Read that again. A company is paying today for power that will not exist for six years, and they are doing it because the alternative (waiting, bidding, hoping) means losing the ability to train and serve frontier models in the second half of this decade.
This is the new shape of corporate strategy, and almost no one outside of three or four boardrooms has noticed.
The story you have been told about AI is a story about chips. About models. About data. About talent. All of these matter. None of them is the binding constraint anymore. The binding constraint is electricity, and the people who run the largest AI companies in the world figured this out roughly eighteen months ago. They have been quietly buying every watt they can get their hands on ever since. Most boards I speak with are still arguing about which model to use for their customer support pilot.
You are playing a different game than they are. You do not yet know it.
The constraint moved
For most of the modern computing era, the binding constraint on what a company could do with software was talent. Engineers were scarce, machines were cheap, and the limit of your ambition was how many smart people you could hire and retain. Cloud computing slowly inverted part of this. Capital became a way to convert dollars into compute on demand, and for about fifteen years the bottleneck shifted to whether you could afford the bill.
Then the bill stopped being the problem. The grid became the problem.
The training run for a single frontier model now consumes the annual output of a small city. Inference, which everyone assumed would be a rounding error, turned out to be larger than training in aggregate, because billions of queries each day add up faster than a few quarterly training jobs. A single agentic workflow can consume more compute in one autonomous task than a human user generated in a year of chat. Multiply that by the number of agents that will be running by 2027 and the curve goes vertical.
The transmission grid in the United States cannot be expanded fast enough to meet the projected demand. Not because the technology does not exist. Because of permitting, interconnection queues, transformer shortages, and the simple physical reality of running new high voltage lines across land owned by people who do not want them there. The current interconnection queue for new generation in the US sits at over two terawatts. The wait time in some regions is now seven years.
Seven years. From now. To plug a power plant into the grid.
This is the wall that AI strategy has just hit, and the companies that understood it first are the ones now signing twenty-year offtake agreements for plants that do not exist.
What the hyperscalers actually bought
Look at what has happened in the last thirty days, not the last thirty months. One hyperscaler restarted a previously decommissioned nuclear reactor on Pennsylvania farmland and bought every electron it will produce. Another locked up small modular reactor capacity in three states for delivery in 2030. A third purchased a stake in a fusion company that will not produce commercial power for at least a decade. A fourth signed a contract for geothermal capacity in Nevada that nobody can build at scale yet.
These look like clean energy stories. They are not clean energy stories. They are corporate strategy stories disguised as clean energy stories.
What these companies are buying is not power. What they are buying is the right to compute at scale in 2030. Every watt they secure now is a watt their competitors cannot secure later. The hyperscaler that controls 200 terawatt-hours of dedicated generation by 2030 will be able to run an AI business of a certain size. The hyperscaler that controls 50 will not. There is no software trick that closes that gap. There is no algorithmic efficiency improvement that lets you run a frontier model on power you do not have.
The race for talent in 2015. The race for chips in 2023. The race for electrons in 2026.
We are watching the third great resource scramble of the modern computing era happen in real time, and the boards I speak with most weeks are still treating power as a line item on the operating budget rather than a strategic asset on the balance sheet.
Why this changes the shape of every company, not just the ones with GPUs
Here is the part most executives miss, and the part that actually matters for any business that is not a hyperscaler.
If frontier compute is gated by physical electron supply, then the cost of intelligence is going to bifurcate violently over the next four years. The companies that locked up cheap dedicated generation will be able to offer inference at a price point that no spot-market buyer can match. Everyone else will pay variable rates that move with grid stress, weather, and seasonal demand. We are heading for a world where running an AI agent in August during a Texas heat wave costs four times what it cost in March, and where some agentic workflows will be paused not because of bugs but because the underlying compute provider has been throttled by the grid operator.
This is not speculation. This is already happening at the wholesale level. Several major AI labs have been quietly throttling non-priority inference during peak grid hours since last summer. Your customer service agent slowed down in July and you blamed the model. The model was fine. The grid was not.
Your strategy has to account for this. If your business plan assumes that inference costs decline smoothly forever (the assumption baked into nearly every AI-native business case I have reviewed in the last twelve months), you are about to get a very expensive lesson in physical infrastructure. The cost of token generation is not going to follow a smooth Moore's Law style decline. It is going to step. It is going to oscillate. It is going to depend on whether your provider has dedicated power or is buying on the spot market. And the providers with dedicated power are going to charge a premium for predictability, because predictability is suddenly the rarest thing in compute.
The three classes of company that are about to emerge
There are three positions a company can occupy in the new compute geography. Each one implies a completely different operating posture.
The first class owns generation. Hyperscalers and a handful of vertically integrated AI labs. They will sell intelligence at a price floor set by their power costs and capture the margin on everything above it. Their strategic problem is regulatory and political, not technical. They need permits, they need grid interconnects, they need local communities to not revolt. Most of them will fail at one of these. The ones that succeed will own the next decade.
The second class buys generation under long-term contract. This is the layer most people have not realized exists yet. A small number of large enterprises (banks, pharmaceutical companies, defense contractors, a few industrial players) are starting to sign their own power purchase agreements directly with utilities and independent generators. They are doing this because they have figured out that being a price-taker in the inference market is incompatible with running mission-critical AI workloads. If your fraud detection model needs to run in 80 milliseconds and your inference provider is being throttled, your fraud detection does not run. So you go upstream and you secure your own electrons.
The third class buys intelligence on the spot market. This is where almost every company reading this currently sits, and it is the most dangerous position to be in for the next four years. You will face pricing volatility you did not plan for. You will face availability constraints during peak demand. You will discover that the agent workflows you built in 2025 do not have the same unit economics in 2027 because your provider passed through a 60% increase in compute costs that came from grid pricing rather than model pricing. And you will have no recourse, because you signed a clickwrap.
The strategic question for every executive reading this is which class you want to be in. If you are running a company with serious AI exposure (and almost every company will have serious AI exposure within 24 months), staying in the third class is a slow loss. You need to figure out whether you can credibly move to the second.
What "moving to the second class" actually looks like
This is where the work begins, and where most consultants will tell you something useless.
Moving to the second class does not mean building a data center. Almost no company should build a data center. What it means is restructuring your AI architecture so that you have optionality across multiple compute providers, with contractual guarantees on availability and pricing, and a clear understanding of which workloads can tolerate spot-market pricing and which cannot.
This requires three things, all of which most companies do not currently have.
You need a real workload classification. Which of your AI processes are latency-sensitive, which are availability-sensitive, which are cost-sensitive, and which are all three. Most companies have never done this exercise. They have one AI budget and one AI provider and they hope.
You need an architecture that lets you route workloads dynamically. If your customer-facing chatbot needs guaranteed availability and your batch document processing can tolerate four-hour delays during grid stress, those should not be running on the same compute contract. They should not even be running through the same abstraction layer. Today, in most companies, they are.
You need contractual relationships with at least two compute providers, with one of them offering capacity guarantees rather than spot pricing. The premium for guaranteed capacity is going to look expensive in 2026 and look like a bargain in 2028. The companies that lock it in early will have a structural cost advantage over the ones that wait until they need it.
This is architectural work. It cannot be bought as a product. It cannot be solved by a vendor pitch deck. It requires sitting down and looking at every AI workflow your company runs, asking what its physical compute requirements actually are, and then designing an infrastructure posture that maps those requirements to a power-aware procurement strategy.
The second-order effect nobody is talking about
There is a deeper pattern here that the power story makes visible.
For thirty years, the assumption underneath corporate strategy has been that digital resources are abundant and physical resources are scarce. You optimized your supply chain because steel and trucks and warehouse space were the constraints. Software was free, in the sense that you could always buy more of it. The cloud reinforced this. Compute felt elastic.
That assumption is dead. Compute is now a physical resource constrained by physical infrastructure that takes physical years to build. Your AI strategy is now a physical strategy. The board that treats it as a software question is going to be outmaneuvered by the board that treats it as an industrial question.
This has implications that ripple outward. It means M&A targets get evaluated partly on their compute contracts. It means CFOs need to understand power purchase agreements the way they currently understand interest rate hedges. It means that the geography of where you run your business starts to matter again, after thirty years of geography supposedly not mattering. Companies in regions with cheap, abundant, low-carbon power are going to have an inference cost advantage that compounds. Companies in grid-constrained regions are going to find themselves paying a premium that nothing in their software architecture can fix.
The location of your inference is going to become a strategic variable on par with the location of your manufacturing was in 1995. Most strategy teams have not yet woken up to this.
The window is closing
The companies signing twenty-year power contracts today are doing so because they understand something most executives do not. The window to lock in cheap, predictable AI compute is closing. By 2028, the spread between contracted and spot-market inference pricing will be visible enough that the strategic question will be obvious to everyone. By then, most of the contractable capacity will already be spoken for.
You have roughly 18 months to position. After that, you are buying retail in a market where your largest competitors have been buying wholesale for three years.
This is not a story about technology adoption. This is a story about industrial strategy in a world where intelligence has become a physical commodity. The executives who treat it as the former will be outcompeted by the ones who treat it as the latter, and the gap will be visible in operating margins by 2028.
What you should actually do
Stop buying AI tools. I mean it. Stop scrolling vendor decks. Stop running another pilot.
Sit down with your CFO and your CTO and ask three questions. What are our actual AI workloads going to look like in 36 months, broken down by latency, availability, and cost sensitivity? What is our current exposure to spot-market compute pricing, and what does our P&L look like if that pricing increases 80% over two years? What contractual relationships do we have, or could we have, that would give us guaranteed capacity at predictable prices for the workloads that need it?
If you cannot answer these questions, you do not have an AI strategy. You have an AI shopping list. Those are different things, and the difference is going to determine which side of the cost curve you sit on for the rest of this decade.
This is the work Agor AI Advisory does. Not vendor selection. Not chatbot deployment. The architectural decisions that determine whether your AI infrastructure is an asset or a liability when the constraint moves from software to physics. We sit at the intersection of strategy and engineering because that is where the actual decisions get made, and we have been telling our clients about the watt ceiling for a year because we saw it coming.
The hyperscalers signed their contracts. The largest enterprises are signing theirs. The window for everyone else is open right now and will not stay open. If you want to architect your company for the new shape of compute (rather than waking up in 2028 to find that your competitors did and you did not), the time to start is this quarter.