Future spending commitments attributed to Alphabet, Amazon, Microsoft, Meta, Oracle and Anthropic exceeded $4 trillion in a calculation published by El Economista on November 9, 2026. The tally combines purchases, leases and other obligations across different time horizons, including Anthropic’s conditional maximum expansion case.
Those time horizons matter. Some payments concern near-term purchases; others relate to infrastructure agreements stretching decades ahead. The economic question is whether the resulting capacity can generate enough revenue to cover its costs and deliver a return to the people funding it.
The six-company calculation
El Economista reported the following company amounts. Alphabet accounted for the largest figure, while Anthropic’s entry depended on exercising its highest expansion options.
| Company | Reported future commitments |
| Alphabet | $1,076.8 billion |
| Amazon | $431 billion |
| Anthropic | Up to $895 billion if the highest expansion options are exercised |
| Meta | $810 billion |
| Microsoft | $519 billion |
| Oracle | $322 billion |
The categories behind these figures vary. In its October 30, 2026 analysis, Bloomberg reported more than $3 trillion across the five companies in this table other than Anthropic. Its breakdown included $1.3 trillion in future lease commitments and $1.9 trillion in other purchase commitments.
Bloomberg noted that some commitments concerned activities beyond data centers. Amazon’s figures included planned content-license spending, for example, while Meta’s included some Reality Labs consumer hardware. Bloomberg’s review also excluded current leases and debt. Some Microsoft purchase and construction figures came from its fiscal-year report ending in June, whereas the other companies’ figures came from more recent disclosures.
The October and November calculations therefore cover different company sets and reporting cutoffs. Their headline totals describe broad collections of corporate obligations with different contents.
Why commitments extend beyond annual capex
Capital expenditure, usually shortened to capex, measures investment in assets such as buildings and equipment during a particular period. Future commitments describe spending obligations that can extend across many periods.
NeoTeo’s earlier report on planned 2026 capital expenditure focused on annual spending plans. Our earlier English coverage of cash flow and future commitments examined how that investment relates to cash generation. The newer six-company calculation brings together a longer stretch of contractual spending.
A lease whose term has yet to start creates a different payment schedule from an equipment purchase. Multi-year chip agreements and long-term energy contracts add further schedules. These arrangements can secure the resources needed to build and operate infrastructure while committing the buyer to payments well into the future.
Meta offers a concrete example. El Economista reported $810 billion in commitments, including more than $450 billion described as non-cancellable. Around half of Meta’s total related to data-center leases that had yet to commence, with terms ranging from more than a year to 30 years. That makes the duration and flexibility of the contracts central to understanding the exposure.
Project financing adds another layer: the funding structure sits alongside the purchase or lease agreement. ARPU’s earlier explanation examines project-financing structures and joint ventures as part of the infrastructure buildout, using animated accounting graphics to show how future obligations extend beyond headline capex.
Anthropic’s minimums and expansion options
Anthropic’s reported agreements illustrate how minimum purchases, estimates and expansion options can produce very different levels of eventual spending. According to El Economista, its cloud-provider agreements included:
- Google: a $114.5 billion minimum agreement covering April 2026 through July 2033, plus estimated additional spending of $125 billion–$150 billion on computing using Google’s TPU chips during 2028–2031.
- Amazon: a $110 billion minimum agreement covering May 2026 through April 2036.
- Microsoft: at least $44.9 billion covering November 2026 through May 2033.
The newspaper also reported approximately $161.2 billion in Broadcom equipment leases, excluding interest; up to $120.5 billion with SpaceX through 2029; and more than $20 billion in planned AMD spending subject to initial capacity. It listed computing-services contracts of $44.6 billion with Nscale and $41.3 billion with Lambda, alongside plans for up to another $150 billion from 2028.
The practical financial difference is flexibility. A minimum purchase establishes a spending floor. An expansion option makes additional spending conditional on taking that option. A non-cancellable obligation leaves less room to reduce payments if the business develops differently than expected.
What a hypothetical data center would have to earn
A more useful way to examine potential returns is to put a revenue requirement beside a construction budget. In an analysis published September 15, 2026, Boston Partners modeled a hypothetical one-gigawatt data center with an all-in construction and equipment cost of at least $50 billion.
Its illustration assumed equal funding from debt and equity. At the $50 billion starting cost, that means $25 billion borrowed and $25 billion supplied as equity capital. Return on equity measures annual net income relative to that equity investment.
The return test is the recurring revenue needed to cover operating costs and earn income on the capital invested.
Boston Partners compared two return targets while holding the $25 billion equity base fixed:
| Assumed annual return on equity | Equity investment | Required annual net income |
| 10% | $25 billion | $2.5 billion |
| 20% | $25 billion | $5 billion |
The arithmetic is straightforward: $25 billion × 10% equals $2.5 billion in annual net income. Raising the required return increases the earnings hurdle even when the facility and financing split stay the same.
Boston Partners estimated annual operating costs at just over $12 billion. Under its assumptions, it put the required annual gross revenue at $15.5 billion–$18.6 billion to meet the illustrated return targets. Revenue is the money brought in; net income is what remains after expenses. A facility must generate enough of the former to produce the latter.
This is a hypothetical project with specified costs, financing and return targets. Its value is the explicit connection between the capital invested and the annual business needed to support it. Readers approaching the buildout as an investment story can find broader context in NTV88’s finance and investing coverage.
The investment risk is a mismatch between contracted payments and the cash generated by the capacity they finance. Long-lived, non-cancellable commitments can leave less room to adjust spending if revenue disappoints. The AI-bubble debate turns on expectations about future demand and achievable returns, while these contracts determine how much spending companies have committed along the way.
Customer concentration adds another financial risk
The revenue side has its own concentration problem. In an October 21, 2026 CNBC report, Kate Rooney cited Ramp data showing that the top 1% of AI spenders in its data accounted for about 80% of AI spending. Those high spenders averaged around $10,000 a month, while median monthly spending was below $20.
These customer-level figures describe the distribution of spending on AI services. About 80% of the large spenders were venture-backed private companies, according to the figures Rooney presented.
For infrastructure financed against years of expected usage, the durability of those customers’ spending matters. Rooney warned that weakness among those customers could ripple through model developers, cloud providers and chip suppliers.