Explainer · Infrastructure

The data-centre financing stack, explained

Who pays for the AI build-out at each stage, and why private credit sits in the middle of it

Written 3 September 2026 · All explainers

How to read the markers. [D] — verified from a source retrieved for this note, listed under Sources. [I] — industry-understood; not formally disclosed by a named party. [S] — our own inference or judgment. Every figure is either sourced or marked as an estimate.

The short version

A data centre is paid for in two halves. The building, its power supply and its cooling are a long-lived shell that a tenant leases for a decade or more. The servers inside are equipment that goes stale in about five years. Lenders finance the shell like infrastructure and the chips like machinery, and nearly every headline about AI financing is one of those two halves being paid for at one moment in the building’s life.

That life has four stages: site and power, construction, lease-up (the tenant signs and moves in) and stabilised operation. Each has a natural payer: the hyperscalers’ own cash and bonds throughout, infrastructure equity first, construction loans and private credit for the build, joint ventures and sale-leasebacks where an outside owner pays for a campus the tenant controls. Securitisation — asset-backed securities (ABS) and commercial mortgage-backed securities (CMBS) — arrives last, refinancing buildings that already have tenants and rent.

The scale is why every channel runs at once. Amazon, Alphabet, Microsoft and Meta have guided to roughly $730bn of capital spending for 2026 between them (our sum; the four define “capex” slightly differently). Morgan Stanley puts global data-centre investment through 2028 at about $2.9tn, about $1.5tn of it from outside the hyperscalers’ own cash. [D — see Sources]

Two risks sit badly with this chain: the chips lose value far faster than the building that houses them, and the rent on most of these buildings comes from the same handful of tenants. The shell financing looks like investment-grade infrastructure. The compute financing does not.

1. What a data centre is as a financial asset

Strip a data centre to its lease and it is a power contract with a roof. The landlord builds the shell, core, base power and cooling. The tenant brings the servers. Rent is quoted per kilowatt of reserved power per month: CBRE’s North American average asking rate for a 250–500 kW requirement was $195.94 at the end of 2025. Build-to-suit hyperscale leases run 15 years or longer. [D — Build, CBRE, DLA Piper]

Two things follow for anyone lending against it. The landlord’s asset is long-lived. RBC Capital Markets describes data centres as assets “with 50+ year useful lifespans, typically operating under 5-15 year lease agreements”. Microsoft has just lengthened the accounting life of its data-centre and office buildings from 15 to 25 years. [D]

The tenant’s asset inside it is short-lived. Meta depreciates most servers over 5.5 years. Amazon cut a subset of its servers from six years to five in January 2025, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning”. Microsoft’s CFO said in July 2026 that about two-thirds of its quarterly capex went on “short-lived assets, primarily CPUs and GPUs”. [D]

So the same campus is two credits. The shell is underwritten on a long lease to a large tenant, like a toll road. The chips are underwritten on a customer contract and a resale value, like aircraft — except that nobody yet knows what a five-year-old AI accelerator sells for. [S]

2. The size of the bill

The four largest builders publish their own numbers each quarter. The table keeps each figure with its period and wording, because the four do not define “capex” the same way. [D]

Company 2026 guidance (latest) Earlier 2026 guidance Most recent actual Stated on Source
Amazon “approximately $220 billion in cash CapEx” ~$200bn (Feb 2026) 2025: $128.3bn, net purchases of property and equipment 30 Jul 2026 Q2-2026 call; Q4-2025 release
Alphabet $195–205bn $180–190bn (Apr); $175–185bn (Feb) H1 2026: $80.6bn 22 Jul 2026 Q2-2026 release and call
Microsoft “approximately $175 billion” for calendar 2026, incl. finance leases higher, before a lease-accounting change Apr–Jun 2026: $41bn incl. finance leases 29 Jul 2026 FY26-Q4 call
Meta $130–145bn, incl. finance-lease principal $125–145bn (Apr); $115–135bn (Jan) 2025: $72.22bn 29 Jul 2026 Q2-2026 call; Q4-2025 release

Guidance has only moved up through 2026. The midpoints sum to roughly $730bn, against $72bn (Meta) and $128bn (Amazon) spent in 2025. Microsoft’s figure fell only because more of its future leases will be operating leases, which do not count as capex; its “underlying investment plans remain unchanged”. [D; the $730bn sum is S]

Industry-wide estimates are larger. Morgan Stanley, relayed by the Structured Finance Association (SFA) in July 2026, estimates about $2.9tn of global data-centre investment through 2028 including land, buildings, power and hardware, with roughly $1.4tn funded from hyperscaler cash flows and $1.5tn to be financed externally. The IEA projects data-centre electricity use more than doubling, from 415 TWh in 2024 to about 945 TWh in 2030, and warns that about 20% of planned projects could be delayed by grid-connection queues. [D] The “so what”: the hyperscalers can fund perhaps half of this themselves; the other half is the market every other channel competes for.

3. The lifecycle, and who pays at each stage

The SFA’s July 2026 primer states the sequence plainly: “Development typically begins with construction loans, project finance, private credit or corporate debt, whose investors assume development risks such as construction, equipment procurement and power availability. Once a facility is complete, leased and generating predictable cash flows, it may refinance through ABS or CMBS”. [D]

Stage 1 — site and power. Land, grid connection, permits and, increasingly, on-site generation. This is equity money: developers, infrastructure funds and the hyperscalers themselves. Power is the gating item. Brookfield’s partnership with Bloom Energy, expanded from $5bn to $25bn in June 2026, finances fuel-cell generation on site so that campuses do not wait years for a grid connection; Brookfield’s $100bn AI-infrastructure fund (November 2025) sits behind it. [D — Bloom Energy release]

Stage 2 — construction. Construction loans from banks, project finance and private credit carry the build, secured on the land, the contracts and the sponsor’s equity. In Bisnow’s August 2026 tally of $1.3tn of US data-centre debt, banks still lead, but PIMCO ($23bn), Blue Owl ($11bn) and Blackstone Credit ($8bn) are among the largest lenders. [D — trade source]

Stage 3 — lease-up. The tenant moves in, starts paying rent and installs its own servers, paid for from its own capex or, for the newer AI companies, from chip-backed loans (section 4). For a build-to-suit campus this stage is short: the lease was signed before construction began. [I]

Stage 4 — stabilised. With rent flowing and a long lease in place, the owner refinances into the cheapest money available: ABS, CMBS or, for the largest single-tenant campuses, a bond sold against the lease. [D — SFA]

Who pays at each stage of a data centre's life Four stages run left to right: site and power, construction, lease-up, stabilised operation. Beneath them five funding channels are drawn as bars covering the stages where each is active. The tenant's own cash flow and corporate bonds cover all four stages. Developer and infrastructure equity covers site and construction. Construction loans and private credit cover construction and lease-up. Joint ventures and sale-leasebacks cover site through lease-up. ABS and CMBS cover only the stabilised stage. The lifecycle · which channel pays where 1 · Site and power land · grid queue · generation 2 · Construction shell · power · cooling 3 · Lease-up tenant moves in, adds servers 4 · Stabilised long lease · rent flowing Tenant cash flow · bonds self-build campuses, and the servers in every campus · funds every stage Infrastructure equity developers · infra funds land, grid connection, on-site power, sponsor equity Construction debt banks · private credit carries the build until rent starts JV · sale-leaseback Hyperion-type structures outside owner funds the shell; tenant leases it back and guarantees residual value · bonds sold before completion ABS · CMBS securitisation refinances leased, stabilised assets only Equity first · debt in the middle · securitisation last · the tenant throughout
Figure 1. The four stages of a data centre's life, left to right, with the funding channels active at each. The tenant's own cash and bonds run through every stage; equity goes in first; construction debt and private credit carry the middle; securitisation arrives only once the asset is leased and stabilised. Stage boundaries are schematic.

Two channels cut across the stages. The hyperscalers’ own cash and bonds fund the campuses they build themselves, from land to servers. And joint ventures with an outside owner are set up at stage 1 and financed before the building is finished: “corporate bonds have largely funded single-tenant AI data-center developments, often before construction is complete”. [D]

4. Off the balance sheet, but not out of control

The structure that drew most attention is Meta’s Hyperion campus in Louisiana. On 21 October 2025 Meta and funds managed by Blue Owl Capital formed a joint venture to develop and own it, committing about $27bn “for the buildings and long-lived power, cooling, and connectivity infrastructure”. Blue Owl’s funds own 80%, Meta 20%. Blue Owl put in about $7bn of cash; Meta took a one-time distribution of about $3bn and contributed the land and part-built assets. Meta manages construction and the property, leases the facilities on “a four-year initial term with options to extend”, and gives a residual value guarantee — a capped cash payment if the lease is not renewed — covering the first 16 years of operations. [D — Meta release; Kirkland release]

The debt behind Blue Owl’s share is the record. A bankruptcy-remote vehicle, Beignet Investor LLC, sold $27.3bn of senior secured amortising bonds on 16 October 2025: 6.581% coupon, final maturity May 2049, rated A+ by S&P, Morgan Stanley sole bookrunner, PIMCO the anchor. S&P’s reasoning, as IFR relays it, is that the contract passes “substantial credit risk to Meta during both construction and operation phases”. [D — IFR, Fortune, Bisnow]

Meta keeps control of the campus and the option to walk away after four years, yet the $27bn of buildings and the bonds sit on the joint venture’s books, not Meta’s. Investors accept this because the guarantee makes Meta the effective credit. The guarantee is the hinge: it turns a four-year lease into something a 23-year bond can be sold against. And note what the money bought — buildings, power, cooling and connectivity. The servers are not in the $27bn. [S — our reading; the facts are D]

Two comparables apply the same idea to the chips rather than the shell:

  • CoreWeave, an AI cloud provider, closed an $8.5bn delayed-draw term loan on 31 March 2026 secured on “HPC infrastructure and associated customer contracts”, rated A3 by Moody’s — the first investment-grade rating for a GPU-backed loan — with Blackstone as anchor investor. [D — CoreWeave release]
  • Anthropic is the beneficiary of a $35bn secured platform, reported by PitchBook in June 2026, in which Apollo and Blackstone lend against AI chips held in a vehicle that leases the hardware to the company. [D — PitchBook via Yahoo Finance]

In both cases the collateral is a contract plus hardware: one customer, and the thing in section 6 that nobody can yet value at year five.

5. Securitisation: the refinancing, not the funding

ABS and CMBS are the cheapest money in the stack, and the last to arrive. KBRA found in May 2025 that they “have generally been backed by built and stabilized cash-flowing assets with little to no remaining construction or lease-up risk”. They refinance the construction loan; they do not replace it. [D]

Measure Figure Period Source
US data-centre ABS + CMBS issued $48.69bn across 88 deals; 70.8% ABS 2018 to May 2025 KBRA
Annual issuance over $25bn, “exceeding the combined total of the previous three years” 2025 L&G
Split of 2025 issuance roughly 50/50 ABS and CMBS 2025 RBC Capital Markets
Outstanding $4bn → $61bn 2020 → YTD 2026 Barclays, via SFA
Ratings cap 70 tranches rated by S&P, none above A 2020–2025 GlobalCapital

Why it cannot fund the build: the rating depends on rent that exists, and the average deal (about $600m for ABS, $1.2bn for CMBS) is a fraction of one AI campus. Morgan Stanley reckons ABS and CMBS could meet about $150bn of the $1.5tn external need through 2028. And S&P’s reasons for rating no data-centre ABS above single-A — obsolescence, tenant concentration, refinancing risk — are the three things the earlier stages of the chain carry. [D — SFA, GlobalCapital]

6. What the chain prices badly

Obsolescence. The building lasts 25 years or more; the chips inside are written off in five or six. That is the hyperscalers’ own accounting, not a bear case — and the two largest moved in opposite directions in the same month, Amazon shortening a subset of servers to five years and Meta lengthening most of its own to 5.5. Both are estimates; neither has sold a fleet of five-year-old accelerators to test them. Google has “rolled out four generations of its TPUs in about three years”. [D]

The building and the box: cost per megawatt and accounting life Left: a stacked bar of cost per megawatt, with shell and core at 11.3 million dollars (JLL 2026 forecast, global average) beneath tenant fit-out of up to 25 million dollars for AI. The fit-out is about 69 percent of the total. Right: horizontal bars of accounting life in years: 25 years for Microsoft's data-centre buildings from fiscal 2027, 6 years for Microsoft servers, 5.5 years for most Meta servers, 5 years for a subset of Amazon servers. Cost · $m per MW Where the money goes one megawatt of AI capacity · JLL global average, 2026 $25m tenant fit-out servers, accelerators, racks ≈ 69% · "up to" figure for AI $11.3m shell and core building, power, cooling · ≈ 31% The larger half is the short-lived half Accounting life · years How long each half is written off over 0 5 10 15 20 25 years · operators' own disclosures Buildings Microsoft data centres, from FY2027 25 Servers 6 · Microsoft 5.5 · Meta, most servers (from Jan 2025) 5 · Amazon, a subset (from Jan 2025) 25 years for the shell · 5 to 6 for what sits inside it
Figure 2. The building and the box. Left: JLL's global average shell-and-core construction cost of $11.3m per MW (2026 forecast) against its estimate of up to $25m per MW of tenant fit-out for AI. Right: accounting lives from the operators' own disclosures — 25 years for Microsoft's buildings from fiscal 2027, five to six years for the servers of Amazon, Meta and Microsoft. The cost split is a global average, not a specific project.

In an AI campus most of the money is in the short-lived half: JLL puts shell-and-core at $11.3m per MW and AI fit-out at up to $25m, and Alphabet says 60% of its Q2 2026 infrastructure spend went on servers, 40% on data centres and networking. The infrastructure-like financing covers the smaller, safer half. The larger half is financed on the assumption that a customer contract outlives the chips. [D for the figures; S for the framing]

Concentration. The tenants are the same names in every deal. GlobalCapital’s account of S&P’s stance notes that “similar subsets” of hyperscalers recur across data-centre ABS, so an investor buying several deals is buying the same credit several times. The chip-backed loans in section 4 are mostly to smaller AI companies, which the SFA notes “present a different credit profile” from investment-grade hyperscalers. Tenants have already “exercised termination rights or withdrawn from major lease commitments”, per DLA Piper. [D]

Power. The IEA’s 20% of projects at risk of delay falls on stage 1 and 2 lenders: a building without power produces no rent to refinance against. [D for the figure; S for the inference]

The balanced reading: a stabilised, leased shell with an investment-grade tenant is a good infrastructure credit, and the market prices it as one. The chips, the construction period and the smaller tenants carry the risk the chain has not yet taken through a recession. That is where private credit has chosen to sit.

Caveats

  • The 2026 capex figures are guidance, not results, and each company defines the total differently: Amazon quotes cash capex; Meta and Microsoft include finance-lease principal; Alphabet’s figure is purchases of property and equipment. The ~$730bn sum of midpoints is our arithmetic and should be read as an order of magnitude.
  • Alphabet’s full-year 2025 capex and Microsoft’s full fiscal-2026 total were not retrieved from a loadable source this session and are left out of the table.
  • Microsoft’s earlier calendar-2026 figure (reported as about $190bn in April 2026) could not be retrieved from a page that loaded; the table says only that the prior figure was higher.
  • The 2025 securitisation issuance figure differs by source and definition: L&G says “over $25 billion”; RBC says about $25bn globally; a KBRA-derived figure of $48.69bn is cumulative from 2018 to May 2025. The differences are recorded, not reconciled.
  • Hyperion bond terms (size, coupon, maturity, rating, PIMCO as anchor) come from IFR and Fortune, not from a Blue Owl or Meta primary document. Meta’s release confirms only that a portion of Blue Owl’s capital “will be funded by debt issued to PIMCO and select other bond investors through a private securities offering”.
  • The Bisnow $1.3tn lender tally comes from a third-party data platform and, by its own account, understates private credit.
  • The shell versus fit-out split in Figure 2 uses JLL global averages. A specific AI campus can sit far from those figures (Thunder Said Energy cites $40m per MW for AI-heavy sites with “over half” in GPUs).
  • No secondary-market price series for used AI accelerators was retrieved. The obsolescence risk is stated from accounting lives, not from observed resale values.
  • McKinsey’s widely quoted $6.7tn-by-2030 estimate did not load this session and is not cited.

Sources [D]

All retrieved and confirmed to load on 3 September 2026.

Hyperion joint venture and bond

Hyperscaler capex and useful lives

Demand, cost and rent

Lifecycle, private credit and securitisation

Compute-backed and power financings