The Power-Cost Map
$8 billion plan to stand up a gigawatt of computing capacity in southern India
For a while I have written about where data centers get built as a domestic question. Which town grants the permit, which grid has spare capacity, which state still has frictionless ground before the consent fights catch up. That framing was right, and it was incomplete, because it assumed the map ended at the border. An order placed this week says it does not.
AM Intelligence, an Indian AI-infrastructure company tied to the Greenko renewable-energy group, placed a binding order for 9,000 of Nvidia's next-generation Vera Rubin systems, part of a roughly $8 billion plan to stand up a gigawatt of computing capacity in southern India. Bloomberg reported the detail that reframes the whole thing. A United States customer has already purchased the first tranche of that capacity and intends to run it from India, over undersea cable, at a latency of around 300 milliseconds. The company's founder was direct about where the advantage comes from: not chips, which anyone can buy, but access to low-cost renewable power. In his words, energy prices are a large part of global token economics.
I will admit I was mildly surprised it landed in India. Then I remembered that the surprise only exists if you believe compute is a special kind of industry. It is not.
Compute is not special
Energy-intensive industries migrate to cheap power. This is one of the oldest patterns in heavy industry, and it is not subtle. Aluminum smelting, which is essentially a machine for turning electricity into metal, left its home markets and relocated to Iceland's geothermal, Quebec's and Norway's hydro, and the Gulf's cheap gas. Steel, ammonia, chlor-alkali, all followed the same gravity. When power is the dominant input cost, the factory moves to the power. It does not ask the power to come to it.
For two decades, one thing exempted computing from that rule: latency. A data center serving users had to sit near the users, because distance is delay and delay is a worse product. That constraint kept compute anchored to population centers and wealthy grids. The India order is a signal that the anchor is dissolving for a large and growing share of the workload. A US customer willing to accept 300 milliseconds to reach cheap Indian power has decided, for that workload, that the latency no longer matters more than the electricity bill. Once that trade tips, compute becomes as footloose as aluminum.
The three filters
So here is the framework I would use to predict where footloose compute lands. It is the intersection of three filters, and a site has to clear all three.
Filter one: cheap, reliable power. Not just cheap, but firm, and ideally stranded or surplus, hydro that spills, gas that would otherwise flare, renewable capacity that outruns local demand. The reservation price a data center can pay for power is higher than almost any prior industrial load, because the compute contract behind it is worth more per kilowatt-hour than the alternatives. That is why data centers are now outbidding bitcoin miners for the same megawatts. But the site still has to have the power in the first place.
Filter two: permission to receive advanced chips. This is not an economic filter. It is a political one. The most advanced accelerators move under export-control regimes, and a jurisdiction has to be cleared to receive them at scale. India qualifies. Much of the world does not, or does so only conditionally. This filter is why the power-cost map and the deployment map are not the same map, a point I will come back to, because it is the crux.
Filter three: a latency-tolerant workload. Not all compute can move. The workload itself has to be one where 300 milliseconds is irrelevant. That sorts the whole market, and the sort is the most important analytical move here.
India clears all three filters at once, which is precisely why it is an early destination rather than a surprising one.
The pattern, not the anecdote
One order is an anecdote. The pattern is what makes it a thesis, and the pattern is already visible from several directions.
In the Gulf, the argument is being made with gas. Microsoft has committed more than $15 billion to the UAE through the end of the decade, and a Microsoft–G42 expansion is bringing new capacity online through the Khazna platform. The Gulf is doing with cheap gas what Iceland did with geothermal.
In the United States, the clearest tell is the bitcoin miners. For years, miners were energy arbitrageurs, chasing the cheapest stranded or curtailed watt on the planet and converting it into a tradeable asset. Their mining chips are useless for AI, but everything around those chips, the energized sites, the signed power contracts, the grid interconnections, the cooling shells, is exactly what AI is short of. So the miners are converting. CoinShares projects that listed miners could draw as much as 70 percent of their revenue from AI by the end of 2026, up from roughly 30 percent at the start of the year, and deals like IREN's multibillion-dollar GPU agreement with Microsoft show the conversion is real, not rhetorical. The scarce thing was never the silicon. It was a place with power where the silicon could run.
Underneath all of it sits the demand curve. The International Energy Agency projects global data-center electricity consumption roughly doubling, from about 485 terawatt-hours in 2025 to around 950 by 2030, with AI the primary driver, and expects US data centers to account for nearly half of all growth in American electricity demand over that window. When a load grows that fast, it cannot be satisfied from the incumbent grids alone. It has to go find power, and it will go wherever the power is cheapest and firmest.
Footloose and sticky
Here is the distinction that turns this from an observation into a forecast. Compute is not one market. It divides cleanly into footloose and sticky.
The sticky part is real-time inference that serves an interactive user, where latency is the product and every millisecond degrades the experience. That compute stays near population, on premium grids, and it is not going anywhere. The footloose part is training runs, batch inference, and overnight agentic work, none of which cares about a few hundred milliseconds. That is the majority of the compute load by energy, and it is the part that migrates.
The implication runs against a comfortable assumption. US leadership in designing the chips is durable, protected by an ecosystem and a design moat that do not move. But US leadership in deploying them, in hosting the actual gigawatts, is not guaranteed by the same logic, because deployment obeys the power-cost map rather than the flag. The country can lead in whose chips run the world and still export a growing share of where they physically run. The number to watch is not chip-design share. It is deployment share, and it is the one that can quietly migrate.
The two maps pull apart
This is where the policy tension becomes the whole story. There are two maps, and they do not overlay.
The export-control map decides where advanced chips are allowed to go. The power-cost map decides where compute wants to go. For most of the last few years those maps roughly agreed, because the cheapest willing power and the cleared jurisdictions were often the same places. They are diverging now. The cheapest firm power is increasingly in places that are not automatic on the control map, and the cleared jurisdictions do not all have cheap power. The winners of the migration are the narrow set of places that clear both filters at once, cheap firm power and permission to receive the chips. That is a smaller and stranger map than either the energy analysts or the policy hawks are drawing on their own, and mapping that intersection is where the real forecasting edge is.
It also globalizes an idea I have written about before. The right siting strategy was never to win the consent fight. It was to build where you are wanted, where a community will fight to keep the facility rather than to block it. Taken across borders, the community that fights to keep the data center might be an Indian state with surplus renewable power and few operational constraints, or a Gulf emirate with gas and capital and a national AI strategy. The escape from the consent bottleneck was never just federal land in Ohio. It is any willing jurisdiction with power to spare.
The honest caveats
Three things keep this from being a straight line, and they matter.
First, cheap power is not the same as secure power. A site with low electricity costs today can still hold a weak position tomorrow, because a utility, a legislature, or a hyperscaler can reprice the next contract. Power tenure, not the spot price, is what actually anchors a facility, and tenure is exactly what a cross-border host may struggle to guarantee.
Second, a gigawatt on paper is not a gigawatt delivered. India's grid, for one, faces real questions about whether it can scale to serve this kind of load reliably, and announced capacity across this whole build has a mortality rate that the press-release numbers never show.
Third, the demand doing the migrating is largely debt-financed. The India buildout is funded by a mix of debt and equity, like most of the rest of it, which means its durability rests on capital staying cheap. The power-cost map tells you where compute wants to go. The cost-of-capital map tells you whether it can afford to get there.
None of that breaks the thesis. It bounds it. The direction is clear even if the timing is not: once energy dominates the cost of a token, the latency-tolerant majority of compute becomes footloose and flows toward the cheapest firm power it can legally reach, across borders, exactly as every energy-intensive industry before it has done.
You do not remove a bottleneck. You relocate it. And when the bottleneck is power, you do not relocate the constraint. You relocate the entire factory to the cheapest source of it, even if that source is an ocean away.