Skip to content
Where Industry Data Becomes Insight.
Visuals

AI Capex Comparison 2026: How Eight AI Infrastructure Leaders Are Spending

AI infrastructure spending is rising across the industry, but the money is not flowing toward the same business model. The differences in silicon, customers, financing and monetization offer a useful map of where each company expects AI value to accrue.

September 7, 2026Hyunjun Seo | Editor S
Eight AI capital streams diverging into custom silicon, cloud infrastructure, external compute, enterprise software and internal AI applications.

AI capital expenditure is usually presented as a leaderboard: who is spending the most, how quickly the number is rising, and whether free cash flow can absorb it. That comparison is useful, but it misses a more consequential distinction. The major AI infrastructure builders are not using capital to build the same business.

Microsoft is combining third-party accelerators with its own silicon and an enterprise distribution layer. Alphabet is using custom silicon across both internal workloads and Google Cloud. Amazon is turning Trainium into a merchant-cloud product as well as a cost-control mechanism. Meta still has a much more first-party destination for its compute. Oracle, CoreWeave and Nebius are increasingly building capacity against external contracts. SpaceX now sits somewhere between an integrated AI company and a compute supplier.

The direction of AI capex may therefore tell us more about the next phase of AI than the headline amount.

Comparison of AI capital flows across Microsoft, Alphabet, Amazon, Meta, Oracle, CoreWeave, Nebius and SpaceX
Where the AI Capex Goes. The same infrastructure spending can lead to very different combinations of silicon ownership, customer exposure and monetization. Source: Sector Foundry analysis; underlying data: company filings and earnings disclosures through September 4, 2026.

The capex map matters more than the leaderboard

The latest spending numbers are already extremely large, but they are not directly comparable. Microsoft reported $41 billion of capital expenditure in its fiscal fourth quarter, with roughly two-thirds directed toward shorter-lived assets such as CPUs and GPUs. Alphabet purchased $44.9 billion of property and equipment in the June quarter. Amazon spent $54.2 billion on property and equipment, while Meta reported $31.1 billion of capex including finance-lease principal payments.

Those accounting differences make a simple ranking less useful than it appears. The more informative question is what the resulting capacity is designed to do.

Microsoft, Alphabet and Amazon are moving toward versions of a full-stack model. All three can combine infrastructure with proprietary silicon, models, cloud services and existing software or distribution. Amazon’s Trainium commitments from Anthropic and OpenAI are particularly important: custom silicon is no longer only an internal cost-saving tool. It is becoming something customers can commit to at multi-gigawatt scale.

Meta represents a different economic case. Its infrastructure can generate returns without becoming a large merchant cloud. Better models, recommendations, advertising systems and consumer agents can monetize compute inside an existing application base. If that model works, some of AI’s largest returns may never appear as cloud revenue.

Strategic map comparing internal versus external AI monetization and silicon control

Four Strategic Endgames. AI infrastructure strategies differ along two important dimensions: who consumes the compute and how much of the underlying stack the company controls. Positions are qualitative, not quantitative scores. Source: Sector Foundry analysis.

The new clouds are making compute itself the product

Oracle, CoreWeave and Nebius point toward another outcome: compute remains a valuable independent layer even if models and applications become more competitive.

CoreWeave ended June with roughly $104 billion of revenue backlog and 1.5 GW of active power. Oracle reported $638 billion of remaining performance obligations, with much of its recent increase tied to large AI infrastructure contracts. Nebius is also expanding against contracted demand while using customer prepayments to fund part of the required hardware.

This matters because the financing model is beginning to diverge along with the technology model. The first generation of hyperscale AI infrastructure was largely financed through the balance sheets of highly profitable technology companies. The next generation increasingly includes contracted cash flows, customer prepayments, supplied GPUs and project-style financing.

Comparison of balance-sheet, contract-backed and customer-funded AI infrastructure financing models

Who Pays for the Next Megawatt? AI infrastructure is moving from a predominantly corporate-balance-sheet model toward a broader mix of contract-backed and customer-funded capacity. Source: Sector Foundry analysis; company disclosures.

Capex is also a forecast

These capital maps imply several different beliefs about where AI is heading.

More proprietary silicon at Microsoft, Alphabet and Amazon would suggest that cost per unit of useful inference is becoming as strategically important as frontier-model performance. Continued expansion at CoreWeave, Oracle and Nebius would suggest that scarce compute can remain a standalone profit pool rather than being fully absorbed into vertically integrated platforms. Meta’s model tests whether the largest value pool sits further downstream, inside advertising, recommendations and consumer applications.

SpaceX adds a fourth possibility. Its AI segment spent $15.8 billion on capital expenditure in the June quarter alone while generating revenue from both AI products and external cloud services. The same infrastructure can therefore support proprietary models and be rented when external demand offers attractive economics.

None of these outcomes is fixed. Today’s strategies are being shaped by accelerator scarcity, power constraints and unusually strong demand for frontier compute. As supply normalizes, the relative attractiveness of owning chips, renting capacity and funding infrastructure against long-term contracts can change.

That is why the useful AI-capex indicator is not simply the total number. Watch what each company chooses to own, who ultimately consumes the capacity, who finances it, and where the return is expected to appear. Those decisions are increasingly a map of what each company believes the AI economy will become.

Sources Used

Research Cut-off: 2026-09-04 16:42 KST

Public primary sources

  • Microsoft FY2026 Q4 earnings call — July 29, 2026. Microsoft reported $41B of quarterly Capex, roughly two-thirds in short-lived CPU/GPU assets, while expanding Maia alongside Nvidia and AMD infrastructure.
  • Alphabet Q2 2026 results / SEC filing — July 22, 2026. Q2 purchases of property and equipment were $44.924B, versus $22.446B a year earlier.
  • Amazon Q2 2026 results — July 30, 2026. Q2 property-and-equipment purchases reached $54.208B; Amazon said the increase was primarily driven by AI investment. Trainium has multi-year, multi-GW commitments from Anthropic and OpenAI.
  • Meta Q2 2026 results — July 29, 2026. Q2 Capex including finance-lease principal was $31.08B, with 2026 guidance of $130–145B.
  • Oracle FY2026 Q4 results — June 10, 2026. RPO reached $638B. Oracle disclosed $75B of prepaid or customer-supplied hardware associated with large AI contracts, materially changing its capital-funding model.
  • CoreWeave Q2 2026 results — August 11, 2026. Revenue backlog was approximately $104B, active power reached 1.5GW and contracted power about 3.7GW.
  • Nebius Q2 2026 shareholder letter — August 12, 2026. Q2 Capex was approximately $5.7B; 70% of deals included prepayments covering 50–60% of related Capex.
  • SpaceX Q2 2026 Form 10-Q. AI-segment Capex was $15.828B in Q2 and $23.551B for the first half. The company also disclosed external cloud-service revenue alongside subscriptions, APIs and other AI products.

Archives

Categories