The AI build's constraint has moved into the supply chain

An aluminum order opened the week in digital infrastructure, which is not where the constraint has been thought to live. Rio Tinto and Prysmian will put inert-anode aluminum into an Ohio data center under an arrangement that also lists Amazon and Wesco, and the shape of that arrangement carries further than the metal: a buyer of finished compute has gone straight to the producers, past the distribution tier that has stood between them.

The comparison it invites is to power procurement, where a buyer with enough load contracts with a producer rather than a utility, and that playbook has been run hard at the power end. Microsoft's appeal window has become a fight over whether connection dollars land on a tariff schedule or in a bilateral agreement, which is a dispute about payment mechanics and not about whether the connection bills get paid.

Metal has no equivalent of an interconnection tariff, so terms get built deal by deal, which is why the counterparty is a smelter rather than a utility. Specifying inert-anode metal goes further still: the buyer is naming a production route rather than accepting a spot price, and that suggests an appetite for the supply chain's emissions profile as well as its delivery schedule.

Tonnage is the least of it. One contract supplying one Ohio hall will not reprice aluminum. The template is the asset, and it is worth more to the companies that can copy it than to the companies that signed it.

What the arrangement does not carry is the thing that would make it a benchmark—a volume, a price and a delivery date—and early templates rarely do. The contracts that follow will show whether direct input is cheaper than the channel it replaced, and until one of them publishes terms, the trade is a direction rather than a number.

For an allocator, the significant thing is that the chain has closed into one counterparty: the buyer of the hall and the buyer of the metal are the same balance sheet with the same delivery interest. That changes the diligence: a tenant with metal, cable and power contracted in its own name is a different credit from a tenant waiting on the equipment channel's promises, and the difference shows up first in who sets the energization date.

Suppliers on the other end of such a contract gain something the equipment tier cannot offer: a named customer with a dated need. If the model spreads, the intermediary tier's share of the order book narrows, and the margin that used to sit in the middle of the chain has to land somewhere, whether on the supplier's side, in the hyperscaler's cost base, or in the price of a hall that can promise a date.

Direct-input contracting is the year's real procurement story, more consequential than any single hall announced alongside it. The arithmetic is blunt: halls get announced faster than the inputs that fill them get made, so the customer holding metal, cable and power is the only kind whose delivery date is not somebody else's promise.

There is a second-order effect worth naming: when the inputs that govern schedules sit outside the buildings, the economics of the buildings stop moving with rent and start moving with procurement. A hall that cannot energize because a transformer is late carries its capital without the revenue meant to amortize it, and the tenant's credit does not capture that.

Hitachi Energy's $528 million plant at Gallman, South Carolina will more than double the company's United States capacity for small and medium transformers, with production beginning in 2029—too late to clear any of the 2026 through 2028 order book. Small and medium transformers are the category a project must have settled before it can energize anything at all.

Read the plant as a statement about the base it doubles: when one facility more than doubles a national category's capacity, the category was thin, and thinness is what turns an ordinary component into a schedule-setting one. The capital is aimed at that thinness, but a four-year lag means the relief arrives after the current round of halls has already been delivered or delayed on its own terms.

For an investor the read is narrower than it looks: transformer exposure in this cycle is exposure to lead time, and lead time cannot be bought in size. That pushes capital toward the trades that skip the wait—refits of sites that already hold a connection, permitted land, anything that shortens the distance between a queue position and an energized hall. Announced greenfield megawatts that would land after the equipment does are the weaker side of that trade, and the gap should widen as 2029 approaches.

PGIM's value-add fund made the cleanest version of that trade in Munich, buying raw land in 2024 and selling it two years later as an entitled 30-megawatt data center site. Nothing was built; a permission changed hands, and permission is the part of merchant development whose return has least to do with the price of steel, copper or a transformer slot. Permitting, interconnection and land assembly produce no rent while they consume years of work, and they are activities a fund can complete without owning a single server.

Nabiax's €800 million deal for the 100-megawatt Alcalá hall in Madrid is the contrast: roughly $9.2 million a megawatt, and no tenant named in the announcement. Asterion Industrial Partners appears alongside Nabiax in the party list. A megawatt price with no tenant behind it is a price for a queue position and a shell, which makes it the noisier of the two benchmarks, because the buyer is underwriting Spain's grid queue as much as the building.

Anyone marking Munich entitlements off the Madrid number is marking a permit with a megawatt multiple. Entitlements and queue-priced halls are different assets, and the entitlement side is the better place to be through 2028, since it is the one that gets paid for a completed process rather than for a place in line.

Two more trades from the same week show developers buying time instead of capacity. OData's $630 million retrofit of two operating Latin American data centers turns grid-constrained power into AI capacity, the fastest conversion available when a new connection would take years, and it leaves Aligned Data Centers with a regional reference customer for the liquid cooling system it aims to sell across the region. Retrofits occupy an unusual position in this market: the power already exists, the connection is already made, and what remains is equipment and labor rather than a queue.

Brazil's ReData law suspends import and consumption taxes for five years in exchange for reserved domestic capacity, renewable electricity and a water ceiling. The water limit shapes the engineering well before it shapes the tax line, so sovereigns are writing the physical terms of the build and paying for them with tax relief rather than with capital. That is the same trade in another currency: a state reserving capacity, electricity and water is underwriting the physical build and taking payment in industrial policy, and the developer's binding constraint moves off the tax line and onto an engineering specification.

Then there are the people, the one input on this list that a check cannot clear: more than two-thirds of data center developers and operators report staffing below what their halls require, according to our survey, and the scarcest skills are precisely the ones that convert a queue position into a running megawatt. Labor is also the only input here without a forward price: metal can be contracted years out and land can be optioned, but a crew has to be trained, and training runs on a slower clock than capex.

The shift has an obvious consequence for underwriting: diligence used to run through the tenant's covenant and the market's rent, and now it has to run through the supplier list, the delivery dates on the long-lead items and the operating crew the developer has actually hired. Those are harder things to schedule into a data room, and they are where the next round of project wins and losses will be decided.

None of this argues against the Madrid hall or the South Carolina plant; both are honest prices for scarcity, and scarcity is real. It argues about which input you hold while the queue clears, because metal, gear, entitlement, water and crew now set the delivery schedule, and the same investors are pricing finished megawatts off the same spreadsheet. The next test is the first direct-input contract that publishes a volume and a delivery date. When that number lands, the template stops being a template and becomes a benchmark every equipment supplier has to answer.