The Power-Compute Nexus: Compute Is the New Tollbooth
They built the forge before they had the fire to feed it — a great engine of iron standing cold, waiting on a river of heat that no one had yet learned how to carry. The furnace was never the hard part. The hard part was the fuel, and the road it had to travel.
The Demand Shock
For two decades, electricity demand in the United States was essentially flat. Efficiency gains cancelled out growth, utilities planned around a stable load, and power was the last thing anyone in technology thought about. Artificial intelligence ended that in about eighteen months.
The clearest way to see the scale is through what the hyperscalers are spending. Combined capital expenditure at the four largest — Amazon, Alphabet, Meta and Microsoft — ran near $226 billion in 2024, roughly $410 billion in 2025, and is guided at about $725 billion for 2026, a jump of roughly 77% in a single year. Fold in Oracle and the five-firm total pushes toward $700–900 billion, and analysts already pencil in more than a trillion dollars for 2027. Goldman Sachs puts cumulative hyperscaler capex across compute, data centres and power at something on the order of $7.6 trillion between 2026 and 2031. The Stargate venture alone — OpenAI, SoftBank and Oracle — carries a $500 billion headline. This is, by a wide margin, the largest private construction programme in history.
The number that matters for this section, though, is not the dollars — it is where they go. A rising share of every capex dollar now lands not on chips but on the physical shell around them: the buildings, the cooling, the substations, and the power contracts. Memory alone is set to absorb roughly 30% of hyperscaler data-centre spending this year, four times its 2023 share. The AI story, told honestly, is an infrastructure story wearing a software costume.
Why Power, Not Chips, Is the Bottleneck
The tell came from Microsoft, which disclosed an Azure order backlog it could not fill — not because it lacked chips, but because the GPUs it already owned sat idle in inventory, waiting for power. When the most valuable company in the world is capacity-constrained by electricity rather than silicon, the binding constraint has moved.
The physics explains why. A traditional server rack draws 5 to 15 kilowatts; an AI training rack draws 30 to more than 100. A single cluster of 100,000 GPUs pulls 70 to 80 megawatts — the continuous draw of a small city — from one connection point. Multiply that across the build-out and the load curve goes vertical:
That last comparison is the whole problem in one line. The compute can be built in a year; the power to run it cannot. And the queue to connect has become surreal: at the end of 2024, roughly 2,300 gigawatts of generation and storage sat waiting in US interconnection queues — more than the entire installed US generating fleet of about 1,280 GW. In Texas, the ERCOT large-load queue hit 410 GW by spring 2026, 87% of it data centres, of which under 2% had actually energised.
Most of that queued capacity will never get built; a large share is speculative, duplicative, or filed to hold a place in line. But the queue’s sheer size is the signal: the scarce, valuable thing is no longer the ability to build a data centre — it is the ability to power one. Whoever controls interconnected capacity, firm generation, or a site with secured power holds the real asset. Everyone else holds a waiting ticket. That is the single most important reframe in this entire section.
The Workarounds — and Where They Lead
Because the grid cannot connect fast enough, capital is routing around it. Three responses dominate, and each is a full topic in its own right:
- Skip the grid entirely. Site the data centre directly at a power source — behind the meter, on-site generation, or straight off a plant — and avoid the interconnection queue altogether. Developers going this route (xAI, Crusoe, Oracle) are energising in one to two years against four-plus on the grid.
- Revive firm power. Renewables can’t run a 24/7 training cluster alone, so the build-out is dragging gas turbines (now backordered for years) and nuclear — including the hyperscaler-nuclear deals — back to the centre of the conversation.
- Squeeze the existing grid. Reconductoring, grid-enhancing technologies and undergrounding aim to move more power through lines that already exist, faster than new transmission could ever be permitted.
Reading It Through the Frameworks
Run this through the primer’s models and the investable shape comes into focus quickly.
How does it get paid? The prize is contracted, investment-grade offtake — a long power-purchase agreement or a take-or-pay lease with a hyperscaler whose credit is impeccable. That is the low-risk, financeable end, and it is why infrastructure and private-credit capital is piling in. The danger zone is merchant exposure: speculative “neocloud” capacity built on the assumption that demand will show up to fill it. Same sector, opposite risk.
What stage is it at? Almost all of this is greenfield — new data centres, new generation, new lines. On the primer’s risk spectrum that lands in value-add and opportunistic territory, not core. Whatever else the AI build-out is, it is not a bond proxy, and pricing it like one is the first mistake.
Where does policy become the cash flow? Everywhere. Power-market rules set what generation earns; interconnection reform decides who connects and when; and the firm-power response leans heavily on tax credits for nuclear, storage and gas. Change any of those and the economics move.
Is the interconnection queue a temporary bottleneck, or a structural moat?
For a developer without power, it’s a bottleneck — a problem to be waited out or engineered around. For an incumbent that already holds interconnected capacity, firm generation, or a permitted site, it is a moat — a multi-year barrier that competitors cannot cross at any price. The durable value in this whole build-out clusters around the second group. Own the scarcity, not the queue.
The Investment Map — and the Tail Risk
If the scarce input is power and the connective capacity to deliver it, the beneficiaries are the picks-and-shovels of the electrical system, not the model-builders:
But this is where discipline earns its keep, because the consensus is now very crowded — and 2026 brought the first real cracks in it.
The single discipline that separates durable exposure from stranded exposure is the contract. Assets with committed, investment-grade offtake — regulated power, contracted hyperscale capacity, take-or-pay firm generation — survive even a sharp demand disappointment. Merchant compute and speculative capacity built on a forecast do not. The AI build-out is real; the question is never whether to have exposure, but which cash flows are contracted. That is the line between owning a tollbooth and owning a bet.
AI turned a decade of flat electricity demand vertical, and in doing so revealed that the true scarce inputs in computing are physical: power, land, water and the connective capacity of the grid. The chips were never the constraint. The tollbooth is the connection to the electrical system, and it is the thing worth owning.
Read every deal in this build-out through the same lens: where is the power, is the offtake contracted, and is the scarcity a moat or a queue? The bull case — two demand shocks stacking on one physical base — is strong. The tail risk — a debt-funded, concentrated bet on demand that may arrive late — is real. Both are true at once, and the contract is what tells them apart.
The smiths who prospered were not the ones with the finest hammers. They were the ones who had secured the fire — who owned the heat itself, while their rivals stood in line at a forge they did not control.
Leave a Reply