This note grew out of Otto's ongoing work on hyperscaler-backed financing and the scale of contracted AI capacity. An exchange on X about the aggregate gigawatts supported by the largest hyperscalers was the right catalyst to formalize our internal audit trail and publish it.
Credit where it is due: the exchange was with Shanu Mathew (@ShanuMathew93). Shanu is one of the better follows for the power-market mechanics behind AI infrastructure, particularly data-center load, transmission, grid reliability and the distinction between announced capacity and what can actually be energized.
What follows separates the parts that are disclosed, the parts that must be estimated and the points where apparently precise numbers stop being comparable.
Otto's current working estimate is approximately 45 GW of planned or under-development data-center capacity economically supported by Alphabet, Amazon, Meta, Microsoft and Oracle. A defensible range is 40–50 GW. This is not 45 GW already energized, and it is not 45 GW of formally disclosed corporate guarantees.
First, define what is being measured
Count a project once when a big-five balance sheet or binding contract is economically supporting the data-center capacity.
- Include owned builds, signed leases, joint ventures and intermediary projects with a binding hyperscaler contract or guarantee.
- Exclude optional future phases, speculative campus maximums, power projects without a committed data-center load and projects that have only been announced.
- Do not add the developer's project to the hyperscaler's total a second time.
- Keep planned, under-construction, commissioned and utilized capacity in separate buckets.
The financial commitment ledger
The latest available filings disclose approximately $831.0 billion of leases that have not yet commenced across the five companies. This is the clearest public evidence that the buildout is contractually real. It is not a pure data-center total and it is not corporate debt today.
| Company | Uncommenced leases | As of | What the filing says |
|---|---|---|---|
| Oracle | $260.0B | May 31, 2026 | Substantially all related to data-center arrangements; generally expected to commence through fiscal 2028. |
| Microsoft | $196.6B | Mar. 31, 2026 | Additional leases primarily for data centers that had not yet commenced. |
| Meta | $182.88B | Mar. 31, 2026 | Data centers, colocation and certain network infrastructure; commencements extend through 2036. |
| Amazon | $106.347B | Mar. 31, 2026 | Company-wide leases not yet commenced. Amazon does not identify the entire amount as data-center-only, so this is the least clean line in the total. |
| Alphabet | $85.2B | Jun. 30, 2026 | Primarily data centers, expected to commence between 2026 and 2031. A separate $5.8B short-term lease was also disclosed and is excluded here. |
| Total | $831.027B | Latest respective filings | Gross undiscounted future lease payments. Different asset mixes and lease terms prevent direct dollar-for-dollar comparison. |
Important: $831 billion is a gross contractual-payment ledger, not a present-value liability, not a data-center construction budget and not a clean measure of critical IT load. Oracle, Microsoft, Meta and Alphabet describe their figures as substantially, primarily or specifically related to data-center infrastructure. Amazon's table is broader.
Start with the reported 34+ GW floor
A June 2026 summary of Jefferies Digital Infrastructure research reported more than 34 GW under development across four operators: Oracle at 13.2 GW, Meta at 9.3 GW, Microsoft at 8.1 GW and Google at 2.7 GW. The rounded company figures add to 33.3 GW; the source describes the unrounded aggregate as more than 34 GW.
| Operator | Reported capacity under development | Included in narrow floor |
|---|---|---|
| Oracle | 13.2 GW | Yes |
| Meta | 9.3 GW | Yes |
| Microsoft | 8.1 GW | Yes |
| 2.7 GW | Yes | |
| Reported aggregate | 34+ GW | AWS excluded |
AWS is the missing large operator. Amazon says AWS added 3.9 GW of power capacity in 2025, expects to double total power capacity by the end of 2027 and is monetizing capacity as it is installed. Otto therefore uses a deliberately broad 6–12 GW AWS and coverage adjustment, rather than pretending Amazon has disclosed one directly comparable pipeline number. The adjustment does not assume that all AWS capacity is externally financed or guaranteed. It only brings the excluded operator back into the broad economic-support estimate.
The 6–12 GW adjustment is an Otto estimate, not a company disclosure. The low end is intentionally conservative relative to AWS's 3.9 GW addition in 2025 and stated plan to double total capacity by end-2027. The high end allows for AWS's larger installed base and reporting-scope differences without counting optional campus phases.
Use Alphabet's disclosed backstop book as a rough unit-cost proxy
Alphabet disclosed $43.8 billion of credit-derivative backstops at June 30, 2026. Reporting by The Information says Alphabet had backstopped leases on about 2.4 GW across roughly ten projects.
$831.027B ÷ $18.25B per GW = 45.5 GW
| Assumed dollars per supported GW | Implied capacity from $831.027B | Interpretation |
|---|---|---|
| $20.8B/GW | 40.0 GW | Higher-cost or longer-duration structures |
| $18.25B/GW | 45.5 GW | Alphabet backstop proxy |
| $16.6B/GW | 50.1 GW | Lower-cost or shorter-duration structures |
This is a cross-check, not proof. A maximum credit-derivative exposure is not the same accounting object as gross future lease payments. Lease duration, rent escalators, land, shell, power, cooling, network content and residual-value support vary project by project. The calculation is useful because it lands near the bottom-up estimate, not because every dollar is comparable.
The credible conclusion is not “45 GW is disclosed.” It is “34+ GW is the narrow observable floor, two imperfect methods converge in the mid-40s, and 40–50 GW is the honest working range.”
How we controlled for double counting
What is known, inferred and still open
From capacity to support to investment consequence
The causal chain matters more than the accounting label. Hyperscaler support lowers financing risk for the project and increases the probability that planned capacity becomes real orders for the physical stack.
That is a demand and financing signal, not a blanket buy signal. The strongest beneficiaries are the scarce inputs required before capacity can be energized. Developers and lenders can still be exposed to construction delays, tenant concentration, lease-duration mismatches, technology obsolescence and residual values.
What would change the estimate
Direct company GW disclosures
AWS publishing its current and contracted power capacity in the same scope as the Jefferies four-company ledger would narrow the range immediately.
Project cancellations or retrades
Terminated leases, reduced phases, delayed power delivery or a tenant moving capacity elsewhere would lower the estimate.
New binding guarantees
Finalization of Alphabet's additional estimated $24.1 billion of future backstops, or comparable disclosures by peers, would raise the supported-capacity ledger if the underlying projects are not already counted.
Commencement and commissioning
As projects begin, uncommenced commitments move into lease liabilities. That changes the accounting bucket, not necessarily the economic exposure. Commissioned MW must be tracked separately from contracted MW.
Scope-clean filings
A data-center-only split for Amazon's $106.347 billion commitment line would improve the financial cross-check and could move either bound.
Otto's verdict
Use approximately 45 GW as a planning estimate, not a reported fact. Treat 40–50 GW as the working range, 34+ GW as the narrow observable floor and 2.4 GW as the only publicly mapped Alphabet lease-backstop figure.
The investment conclusion is that hyperscaler support is making a very large pipeline of AI capacity financeable and pulling demand forward across the physical stack. It does not remove execution or residual-value risk.
The analytical edge is not manufacturing a more precise number. It is keeping legal guarantees, future lease payments, under-development capacity and energized capacity in the correct buckets.
Follow the work as it develops
Some of our best research begins with an exchange on X, where a useful claim or disagreement deserves a closer look. Follow Otto for real-time analysis and the detailed follow-through when a conversation needs more room.
Sources and audit trail
- Alphabet Q2 2026 Form 10-Q: $85.2B uncommenced leases; separate $5.8B short-term lease; $7.6B financial guarantees; $43.8B credit derivatives; estimated $24.1B future backstops subject to final terms.
- Amazon Q1 2026 Form 10-Q: $106.347B leases not yet commenced and the composition of broader contractual commitments.
- Meta Q1 2026 Form 10-Q: $182.88B uncommenced leases for data centers, colocation and certain network infrastructure.
- Microsoft fiscal Q3 2026 Form 10-Q: $196.6B additional leases, primarily for data centers, not yet commenced.
- Oracle fiscal 2026 Form 10-K: approximately $260B additional lease commitments, substantially all related to data-center arrangements.
- The Information, How Google Is Using Wall Street Financing Techniques to Expand Chip Sales: reported Alphabet backstops on about 2.4 GW across roughly ten projects.
- GridReadiness summary of Jefferies Digital Infrastructure research, June 2026: 34+ GW under development across Oracle, Meta, Microsoft and Google. Otto did not have the underlying Jefferies note and therefore identifies this as a reported secondary summary.
- Amazon CEO Andy Jassy's 2025 shareholder letter: AWS added 3.9 GW in 2025 and expects to double total power capacity by the end of 2027.
- Aviva Investors, Bond Voyage: independent context on approximately $800B of committed but uncommenced off-balance-sheet leases across the largest hyperscalers.
Methodology date: July 26, 2026. Dollar figures are gross undiscounted future commitments unless otherwise stated. Company filing dates differ. The range will be revised when more comparable project-level power and contract data become public.