Insight, Energy & Regional Development
A six-year, US$45 billion compute agreement includes about 460 MW of load at a single site. That is not a cloud contract, it is the procurement pattern of an aluminium smelter. There is sixty years of evidence about what anchor-load bargains do to the places that host them.
Anthropic has agreed to spend around US$45 billion over six years renting computing capacity from Nscale, using Nvidia's Vera Rubin systems, with the West Virginia capacity alone representing about 460 megawatts.
Read that as an energy contract rather than a technology one, because that is what it is. A defined block of firm power, at a specific site, committed over a period long enough to underwrite generation and transmission investment. This is not how software companies buy infrastructure. It is how aluminium smelters, chlor-alkali plants and electric-arc steel mills have bought it for seventy years.
Which means the relevant precedent is not the history of cloud computing. It is the history of electro-intensive industrial siting, and that history is unusually well documented.
A disclosure worth making up front. The counterparty in the deal above is the company that makes the model used to draft these articles. The analysis here is the treatment the Lab would apply to any counterparty and takes no position favourable to Anthropic; the piece is published only because the same discipline is being applied.
The pattern has a consistent shape. A very large, continuous, price-sensitive load seeks the cheapest firm power available and is willing to accept a remote location to get it. Because the load is large enough to justify new generation on its own, the anchor's commitment becomes the condition on which the infrastructure gets financed. The anchor therefore writes the terms.
Ghana provides the fullest case and it is worth setting out properly.
The Akosombo Dam was completed in 1965 with an initial capacity of around 912 megawatts, built to power Ghana's industrialisation. Its financing was closely tied to a commitment from the Volta Aluminium Company, a venture dominated by Kaiser Aluminum, whose smelter provided the anchor demand the international financiers required. Valco secured low-cost power for a period of thirty years, at rates below what Ghanaian households and other industries paid, alongside substantial tax concessions.
Four things followed over the subsequent decades.
The downstream industry did not arrive. The smelter processed imported bauxite rather than Ghana's own reserves, so the integrated aluminium industry the project was meant to create never materialised. Ghana supplied the electricity and imported the raw material.
The terms became a permanent political problem. The arrangement was renegotiated under the Rawlings government in 1985, and the power price remained contested for decades afterwards.
The load left. Valco closed completely in May 2003 over its inability to agree an electricity price, restarted in 2006, shut again between 2007 and 2011, and passed into government ownership.
The costs were not only financial. Filling the reservoir submerged around seven hundred villages and displaced roughly eighty thousand people, and altered the ecology of the Lower Volta and the livelihoods of hundreds of thousands.
The dam is still there. It is a genuine national asset and Ghana's electrification owes a great deal to it. But the specific bargain that financed it delivered considerably less than promised to the country and rather more to the counterparty, and the asymmetry was structural rather than accidental.
Raymond Vernon described the obsolescing bargain: a foreign investor holds maximum leverage before committing capital, and loses it once the plant is built and cannot be moved, at which point the host can revise the terms.
Anchor-load deals invert this, and the inversion is the thing worth understanding.
In an anchor-load bargain, the sunk, immobile asset belongs to the host. The generation plant, the substation, the transmission line and the water infrastructure are built for one customer, cannot be relocated, and have no alternative buyer at the scale required. The customer's asset, by comparison, is comparatively mobile and comparatively short-lived.
So leverage moves towards the anchor over time rather than away from it. At renegotiation the host is arguing from a position of a stranded asset and a debt schedule, and the anchor is arguing from a position of being able to leave.
Asset lives run on very different clocks, and the mismatch is what gives the counterparty its leverage at renegotiation.
Two of the five sit with the customer. The other three sit with the region and outlast the contract by decades. Everything blue is what the host is left holding.
A data centre is not a smelter and the analogy should not be pushed past what it carries.
Compute demand is growing rather than cyclical, so the risk of the load disappearing because of a commodity price collapse is different in kind from aluminium's. A six-year contract is not a thirty-year one, which cuts both ways: less lock-in for the host, and less certainty underwriting the infrastructure. Data centres can be repurposed to other tenants in a way a potline cannot. And in the West Virginia case the host is a wealthy jurisdiction with regulatory capacity, courts and a functioning tax system, which is not the situation Ghana was in three years after independence.
The reason the comparison still matters is that this procurement pattern is now being replicated in exactly the places where those protections are weakest. Governments across Southeast Asia, the Gulf, North Africa and sub-Saharan Africa are courting data centre investment with power guarantees, land and tax concessions, and are being advised to compete on electricity price. That is the Valco negotiation with different equipment.
Five things, none of them exotic, all of them easier to obtain before the announcement than after.
A price that indexes to system cost rather than a fixed rate.
The single most consequential term. A fixed price looks stable and transfers all the risk of rising system costs to the host over the life of the contract.
Terms that scale with local demand growth.
If domestic load grows and the anchor's block becomes a larger constraint on the system, the contract should recognise that. Ghana's did not.
Commitments on local access to compute, not local jobs.
Job numbers at a data centre are small and everybody knows it. The plausible development benefit is domestic firms and institutions being able to buy compute at a workable price, and that is a separate condition which is almost never contracted for. Absent it, the promised digital economy is the integrated aluminium industry that never got built.
Interconnection that serves the region, not only the campus.
A transmission upgrade specified for one customer at one voltage is an asset for one customer. Specified slightly differently, it is regional infrastructure. That decision is made once, at design.
A decommissioning and stranded-asset provision.
What happens to the substation, the line and the water allocation if the load leaves in year seven. Ghana never had this conversation and spent forty years having it afterwards.
Four hundred and sixty megawatts landing somewhere is a water question, a land question, a labour-market question and a tax-base question, and it is a question about what the place is for.
Regions that have hosted a single dominant industrial employer before have a specific and earned scepticism about the arrival of another one. In Appalachia that is not sentiment, it is a century of evidence about what happens when an anchor industry's economics change. The right response to that scepticism is not reassurance. It is contract terms that hold when the economics do change, and an honest account of what happens if they do.
What travels with a large industrial operator into a region includes its assumptions about who bears risk. This is a version of the same bargain we described in platform-anchored infrastructure, scaled up to a national grid and a state government.
The Lab works on this across energy access and AI and digital systems, and on what large deployments mean for the places that host them.
If you are negotiating an anchor-load agreement, the terms above are worth putting on the table while you still have the leverage to ask. Tell us what you need to know.
This is an independent insight piece by Transitions Lab. For the Lab's applied work, see AI & Digital Systems and Energy Access & Off-Grid Systems. See also The Mandate Is the Mine on the same single-counterparty structure where the counterparty is a European quota rather than a data-centre. To discuss a study, see Contact.