How much do big tech companies spend on AI in 2026?

As of 4 October 2026, company figures give Alphabet $195–205 billion, Amazon about $220 billion, Microsoft about $175 billion for calendar 2026 and Meta $130–145 billion. Our sum is $720–745 billion for the four. Definitions differ, since Amazon states cash capex while Meta and Microsoft include finance leases, so treat the total as an estimate.

Is there an AI bubble in 2026?

The evidence supports both views. Demand outruns supply at the big clouds and lab revenue is rising fast, yet financing is shifting to debt, backlogs depend on a few cash-burning labs, and the Bank of England flagged circular arrangements. No source can settle it today. Watch pricing power, financing mix and how long hardware stays useful.

How much revenue does AI need to justify the spending?

Bain estimated in 2025 that AI compute needs about $2 trillion of annual revenue by 2030 and would fall $800 billion short. Our sketch for 2026 spending alone, assuming six-year chips and a 10% return, gives roughly $164 billion a year in carrying cost, or about $330 billion of revenue at a 50% margin.

Will the AI capex boom raise or lower AI prices for my business?

Both can happen. Prices for a fixed capability keep falling, but top-tier models list at $10/$50 per million tokens, memory costs are rising and capacity is scarce. Budget a range, re-price quarterly and keep a tested second provider, because planned price changes have been announced and cancelled.

How do I avoid AI vendor lock-in?

Keep an evaluation set so you can test any model, route calls through a gateway or abstraction layer, avoid long prepayments you cannot verify, and keep prompts and tool definitions portable. A second provider tested quarterly turns a vendor outage or price rise into a configuration change.

AI Capex Boom: What Big Tech Spends and What It Must Earn

AI capex at four big tech firms is heading for about $720–745 billion in 2026. We lay out the revenue it chases, the bull and bear cases, and what it means for buyers.

By Alexandru Secureanu4 October 20266 min read

AI capex, the money big tech spends on data centres, chips and power, is heading for roughly $720–745 billion in 2026 across Alphabet, Amazon, Microsoft and Meta, by our sum of the companies' own figures as of 4 October 2026. It chases a revenue base that is growing fast but is still much smaller, so whether this is an AI bubble in 2026 depends on whether usage and prices keep paying for assets that wear out in about six years.

How much is AI capex in 2026?

Capex means capital expenditure: spending on assets that last more than a year. The table uses each company's latest earnings release or call, read on 4 October 2026.

Company	2026 capex, company figure	Latest demand signal	Contracted backlog
Alphabet (Q2, 22 July)	$195–205B guidance, raised from $180–190B. Q2: $44.9B, up 100%	Google Cloud $24.8B, up 82%	$514B
Amazon (Q2 release, call)	About $220B cash capex, raised from about $200B because of memory costs. Q2: $53.1B	AWS $42.2B, up 37%, $169B run rate	$496B
Microsoft (FY26 Q4 call)	About $175B for calendar 2026, including finance leases. Q4: $41B	Azure passed $100B for the fiscal year, up 43% in Q4	$678B
Meta (Q2 release)	$130–145B. Q2: $31.08B	Revenue $60.8B, up 28% (no cloud segment)	not applicable
Oracle (Q4 FY26, 10 June)	Free cash flow of −$23.7B in FY26	Cloud $9.9B, up 47%	$638B

Adding the four capex figures gives $195 + 220 + 175 + 130 = $720B at the low end and $205 + 220 + 175 + 145 = $745B at the high end. Definitions differ: Amazon states cash capex, while Meta and Microsoft include finance leases. Microsoft also changed an accounting estimate, extending datacenter lives from 15 to 25 years, which moves more leases out of reported capex, so its CFO said the shift adjusts the figure to about $175B while the spending plan is unchanged. The IEA reported in April that capex of five large technology companies passed $400B in 2025 and would rise a further 75% in 2026.

What revenue is AI capex chasing?

Disclosures are partial and mixed with non-AI revenue. Anthropic told investors its annualised revenue run rate reached $65B at the end of July, about seven times a year earlier, with preliminary Q2 revenue of $11.5B, and CNBC reported OpenAI's run rate at $40B. A run rate is a recent period annualised, not recognised or audited revenue. Reuters, citing the Financial Times, reported that OpenAI projects revenue rising from $36B in 2026 to $350B in 2030, about $856B of compute and infrastructure spending by 2030 and negative free cash flow of $278B over 2026–2030. Microsoft reports over 30 million paid Microsoft 365 Copilot seats, and Amazon says its AI and chips businesses each passed a $25B run rate.

Here is a carrying-cost sketch, an illustration and not a forecast. Take $730B of capex, the midpoint. Assume two thirds goes to servers and chips, close to the share Microsoft reports for short-lived assets, depreciated over six years, and one third to buildings and power, depreciated over 25 years. Annual depreciation is (730 × 2/3) ÷ 6 = $81B plus (730 × 1/3) ÷ 25 = $10B, about $91B. A 10% required return on $730B adds $73B, so one year of capex costs roughly $164B a year to carry. At a 50% margin on new revenue, that needs about $330B of revenue, for 2026 spending alone. The two lab run rates sum to $105B. Cutting the server life to four years adds about $41B of yearly depreciation. Capex also serves non-AI cloud and advertising, so the gap is overstated, and backlogs overstate certainty in the other direction.

Backlogs help only partly. The four companies with backlogs report $678B, $514B, $496B and $638B, about $2.3 trillion, but backlog is contracted, not recognised, and part comes from labs funded by investors. Microsoft says commercial backlog rose 84%, or 25% excluding OpenAI, with a weighted average duration of 2.3 years and roughly 30% recognised within 12 months. Oracle says $75B of its large AI contracts are prepaid or customer-supplied hardware, raised $43B of debt and $5B of equity in FY26, and expects about $40B more in FY27.

What does the bull case say?

Demand exceeds supply. Alphabet's CFO called it a supply-constrained environment for several quarters, Microsoft says demand exceeds available supply, and Amazon says it will not have enough capacity for all demand in 2026.
Revenue is accelerating. AWS grew at its fastest pace in 18 quarters, Google Cloud grew 82% and Anthropic's run rate rose about sevenfold in a year.
Usage is climbing. Alphabet says its models process 22 billion API tokens a minute, up from 16 billion a quarter earlier, and Microsoft counts 50 million GitHub Copilot users and 100,000 Foundry customers.
The assets pay back. Amazon's CEO outlined servers that break even in under three years, with five-to-six-year lives and mostly five-year contracts.
Cheaper tokens expand use. If volume outgrows price falls, spend rises, as we show in the economics of AI inference costs.

What does the bear case say?

Financing is shifting to debt. The Bank of England's Financial Policy Committee record cites Morgan Stanley's estimate of about $450B of AI-related debt issuance by early September, more than double all of 2025, JP Morgan's estimate of $4.1 trillion of debt-financed AI capex over 2026–2030, and $700B of data-centre capex financed by private credit in 2026–2028. It flags opacity and circular arrangements.
Concentration. Backlogs lean on a few labs that are themselves burning cash.
Thin cash flow. Meta's Q2 free cash flow was $784 million against $31.08 billion of capex, Oracle's FY26 free cash flow was negative $23.7 billion, and Amazon's trailing twelve-month free cash flow was an outflow of $7.6 billion.
The revenue gap. Bain's 2025 report estimated that AI compute needs $2 trillion of annual revenue by 2030, and that supply would fall $800B short even after AI savings.
Accounting and costs. Lives of buildings were stretched, two thirds of Microsoft's capex is short-lived, and memory and component prices pushed up the capex that Amazon and Microsoft report.

Do power and chips limit the build-out?

Yes, and the IEA quantifies it. Data-centre electricity use rose 17% in 2025 against 3% for global demand, from about 485 TWh to a projected 950 TWh in 2030, around 3% of global electricity, with AI-focused use tripling. The IEA adds that bottlenecks reduce the likelihood of more aggressive near-term scenarios. Physical limits slow the build and support prices while capacity is scarce, and they raise the cost of each unit.

Is it an AI bubble in 2026? Our read

The data fit both stories, so we look for evidence that separates them: whether price per unit of capability falls slower than volume grows, how much demand comes from outside the labs, how the financing mix moves, and whether hardware stays useful as long as depreciation assumes. Two cautions from statistics. Run rates and backlogs are not revenue, and disclosed metrics are the flattering ones. Alphabet's $112B Q2 profit came mostly from a roughly $99B gain on equity securities, so read operating lines. The Bank of England's committee adds that a reassessment of AI productivity expectations could hit asset valuations and sovereign debt markets, not only AI shares. Our view is that the technology is real and the financing is stretched in places, which can both be true.

What does the AI capex boom mean for small and mid-size companies?

Prices. Fixed-capability prices fall, yet the top tier rose, since Claude Fable 5.1 lists at $10/$50 per million tokens, twice Opus 5. Anthropic's pricing page shows a Sonnet 5 rise planned for 1 September 2026 that was cancelled. Budget a range and re-price quarterly.
Vendor stability. The same page lists Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5 as retired on the first-party API, with some still offered on partner clouds. CNBC's report says Anthropic disabled two models for about two weeks in June to comply with a government export-control directive. Any vendor can have an outage of this kind.
Lock-in. Keep an evaluation set, test a second provider, avoid prepaying beyond what you can verify, and prefer interfaces that let you switch.

How Danube Labs approaches this

Danube Labs builds My Claude Code, a gateway that routes coding agents across 57 model providers, which keeps switching cheap. For AI choices in your business, see our AI systems and business consulting services, or book a free intro call.

Frequently asked questions

How much do big tech companies spend on AI in 2026?: As of 4 October 2026, company figures give Alphabet $195–205 billion, Amazon about $220 billion, Microsoft about $175 billion for calendar 2026 and Meta $130–145 billion. Our sum is $720–745 billion for the four. Definitions differ, since Amazon states cash capex while Meta and Microsoft include finance leases, so treat the total as an estimate.
Is there an AI bubble in 2026?: The evidence supports both views. Demand outruns supply at the big clouds and lab revenue is rising fast, yet financing is shifting to debt, backlogs depend on a few cash-burning labs, and the Bank of England flagged circular arrangements. No source can settle it today. Watch pricing power, financing mix and how long hardware stays useful.
How much revenue does AI need to justify the spending?: Bain estimated in 2025 that AI compute needs about $2 trillion of annual revenue by 2030 and would fall $800 billion short. Our sketch for 2026 spending alone, assuming six-year chips and a 10% return, gives roughly $164 billion a year in carrying cost, or about $330 billion of revenue at a 50% margin.
Will the AI capex boom raise or lower AI prices for my business?: Both can happen. Prices for a fixed capability keep falling, but top-tier models list at $10/$50 per million tokens, memory costs are rising and capacity is scarce. Budget a range, re-price quarterly and keep a tested second provider, because planned price changes have been announced and cancelled.
How do I avoid AI vendor lock-in?: Keep an evaluation set so you can test any model, route calls through a gateway or abstraction layer, avoid long prepayments you cannot verify, and keep prompts and tool definitions portable. A second provider tested quarterly turns a vendor outage or price rise into a configuration change.