What Meta actually signed

Meta and BlackRock announced on 28 July that they had formed a venture to develop and operate a one gigawatt data centre campus in El Paso, Texas, at a development cost of about 14 billion dollars, roughly 13 billion euros or 11 billion pounds. Funds managed by BlackRock hold 80 percent of the venture. Meta holds 20 percent and will be the campus's sole initial occupant, providing construction, administrative and property management services.

The terms matter more than the headline number. Meta contributes land and construction-in-progress worth about 2.3 billion dollars. BlackRock contributes about 4.9 billion dollars in cash, part of it funded from a 12.5 billion dollar debt financing, through Global Infrastructure Partners and HPS Investment Partners. Meta then receives a one-time distribution of about 1 billion dollars to bring the two stakes into line with the 80/20 split.

Then come the two clauses almost nobody quoted. Meta's leases run a four-year initial term with four options to extend, a potential 20 years in total. And Meta provides residual value guarantees with an aggregate threshold of about 13 billion dollars, decreasing over time, payable as the shortfall between the asset's fair value and that threshold within the first 16 years of the lease term.

Nine percent of the cost, ninety-three percent of the guarantee

Read the two numbers together. Meta puts in 2.3 billion dollars of assets and takes 1 billion back in cash, so its net commitment at closing is about 1.3 billion dollars, near 1.2 billion euros. That is roughly 9 percent of the 14 billion the campus will cost to build.

Against that, the residual value guarantee threshold is about 13 billion dollars, close to 93 percent of the development cost. Meta's contingent exposure is therefore around ten times the net value it actually committed at closing.

This is not a criticism of the structure. It is a description of what a very large buyer can do that you cannot: separate the cash it ties up from the capacity it controls, and pay for that separation with a guarantee rather than an upfront cheque. Credit analysts have already made the narrower point that the bondholders are in substance lending to Meta. The part nobody has written is what the same document says about the contracts sold downstream.

The first decision lands in 2032

Capacity begins coming online in 2028. A four-year initial lease from that point puts Meta's first genuine renewal decision around 2032. The extension options run to a potential 2048. The guarantee runs 16 years and decreases along the way, so the protection the lenders hold decays before the lease options are exhausted.

Now hold that against the contract in front of you. Committed-spend agreements for AI compute in Europe are typically written for one to three years with a minimum commitment, and increasingly with take-or-pay language that bills the floor whether or not you consume it. There is no equivalent of the four-year break and no equivalent of a decaying obligation.

The asymmetry is the point. The party best placed to forecast demand for AI compute wrote itself a short term and repeated options. The party least able to forecast it, the buyer, is asked to commit for a fixed period at a fixed floor.

None of this gigawatt is being built in Europe

A gigawatt is being financed, and none of it sits in Europe. El Paso is in the Texas market, where a large load can be connected on a timetable European and British operators do not have. That is not a footnote to the deal, it is a large part of why the money went there.

ENTSO-E, the association of Europe's transmission system operators, published an assessment on 30 April 2026 concluding that European data centre electricity demand is expected to grow by more than 50 percent between 2025 and 2030, and that concentrated deployment in already constrained regions is straining transmission capacity. The binding constraint on this side of the Atlantic is grid connection.

For a business in London or Manchester the consequence is practical rather than patriotic. Capacity that comes online in Texas in 2028 is capacity you rent across an ocean, priced in dollars, under economics set by that build. If a workload has to sit inside the UK or the EU for legal reasons, that is a separate and more expensive conversation, and it needs to start before 2028 rather than during it.

Ask for the structure the vendor wrote for itself

Ask for the shape, not just the price. A short initial term, extension options priced at signature, and no obligation that outlives the term you actually committed to.

Three questions do most of the work. What is the shortest initial term you will write? What does an extension cost if I take it, and is that price fixed now? And what do I owe if I choose not to extend?

The third question is the one that finds the residual. In a property lease it is called a guarantee. In a compute contract it usually appears as an unused-commitment charge, an early-termination fee, or a clawback of the discount you were given for committing in the first place. It is the same instrument wearing different clothes.

Vendors will tell you that long commitments are simply how infrastructure gets financed. That claim is now testable against a public document. On 28 July, one of the largest buyers of AI infrastructure in the world financed 14 billion dollars of it on a four-year lease.