Hyperscaler Capex Hits $780B as a16z Bets on AI Revenue Upside

The bull case comes with a weak spot

Andreessen Horowitz's latest State of Markets briefing makes a confident argument that the AI boom rests on earnings, not inflated multiples. Inside the same case is a statistic that undercuts it. Growth partners Sarah Wang, David George, Alex Immerman and Santiago Rodriguez presented a deck projecting that Alphabet, Amazon, Meta, Microsoft and Oracle will spend about $780 billion on capital expenditures in 2026. That is nearly double the $416 billion they spent in 2025, and the firm expects spending above $1 trillion a year from 2027 on.21 The same presentation found that 69% of S&P 500 companies have live AI deployments but only 2% track a success metric over time.21

Wang's headline claim is that AI's revenue potential is "unbounded to the upside," while cost reduction is a "latent" opportunity that will always be available.21 That framing is the most useful idea in the deck for enterprise buyers. It also exposes the main risk to the whole trade. Infrastructure is being priced as if companies will find large new revenue lines. Most companies that run AI today cannot yet show, with ongoing measurement, that it pays off at all.

What a16z argued

The panel built its case on market structure. The briefing notes that the market has risen about 90% in the four years since the cycle began, roughly 17% a year, while trading multiples have fallen about 20% over the same period.21 Immerman contrasted today's S&P 500, priced at less than 20 times earnings with 15% earnings growth, with 2021 and 2000, when growth and multiples moved apart.21 Wang added that high-tech equipment, software and R&D now make up roughly 55% of US capital spending.24

On demand, the panel cited a combined cloud backlog of about $1.7 trillion at Microsoft, Google and Amazon. It also said consensus forecasts have free cash flow recovering from 2028, after being squeezed during the buildout.22 Rodriguez said the faster spending comes from agents running parallel, long-running tasks that need far more compute than occasional chatbot queries. He also said hyperscalers that "blinked" on capex have regretted it.21 Some data center materials reportedly can't be obtained until 2028.21

For enterprises, the panel offered ROI examples: Chime's 50% cumulative reduction in cost-to-serve, Shopify's 8% improvement in onboarding, and ServiceNow's $1 billion in AI annual contract value.21 Wang said her team asks portfolio companies such as Stripe a simple question: are new AI dollars going into customer-facing products or into cost optimization?22

The $780B figure depends on who is counted

Other estimates broadly match a16z's capex number, but the totals depend on which companies are included and how spending is counted. Vanguard research, as reported by The Motley Fool, also puts the five-company total near $780 billion, up 87% year over year, with more than $1 trillion a year from 2027 through 2030.4 Bain's September technology report uses about the same $780 billion figure.6

Counts limited to the four largest hyperscalers come in lower. Forbes cites $700 billion to $725 billion for Amazon, Microsoft, Alphabet and Meta.3 CRV puts the total at about $725 billion.8 One tracker reaches roughly $735 billion using company guidance: about $220 billion for Amazon, $195–205 billion for Google, about $175 billion for Microsoft and $135–145 billion for Meta.1 CRV notes that no hyperscaler reports AI capex as a separate line item, so any total is an estimate built from other disclosures.8

The higher estimates go further. Morgan Stanley raised its five-company projection to $805 billion.10 Bank of America puts 2026 near $860 billion, heading toward $1.2 trillion in 2027.7 Every source agrees on the direction, which is close to doubling year over year. The precise total depends on the counting method more than on any change in the facts.

Who pays for it

The a16z deck spends less time on funding than other coverage does. Forbes reports that capex is projected to use up nearly all of the major hyperscalers' operating cash flow this year. It warns that companies once valued for being capital-light are turning into capital-intensive infrastructure operators, a shift that tends to lower valuation multiples unless returns prove exceptional.3 A Columbia working paper cited by TFTC finds that 2026 is the first year the five companies' combined capex exceeds their combined operating cash flow.7 PIMCO estimated AI capex would absorb about 94% of hyperscaler operating cash flow in 2026–2027, up from about 40% in 2023.5

Debt is filling the gap. Morgan Stanley estimates hyperscalers will borrow about $400 billion in 2026, double the 2025 amount.10 Goldman Sachs expects more than a third of 2027 capex to be funded with investment-grade bonds.5 a16z's own social account has noted that hyperscalers increasingly rely on debt for the buildout.28

The revenue needed to justify this spending is large. On one synthesis of Goldman's work, hyperscalers need roughly $300–308 billion a year in AI revenue just to break even. The cleanest evidence of cloud revenue above the pre-AI trend is about $70 billion annualized, roughly a quarter of that bar.6 Bain goes further: an AI market approaching $6 trillion a year by 2031 would be needed to support projected infrastructure spending.9

Enterprise adoption is wide but shallow

The 69%-versus-2% gap from a16z matches a pattern found in many independent surveys, although each measures "value" differently. A compilation of 2026 enterprise data shows 88% of organizations using AI in at least one function. In the same data, 37% attribute any EBIT impact to AI and about 6% qualify as high performers with significant impact. KPMG found only 7% of leaders reporting established ROI.12 WRITER's survey of 2,400 leaders found that 29% see significant ROI from generative AI and 23% from AI agents. Three-quarters admitted their AI strategy is "more for show" than real guidance.11

The McKinsey figure that matters most may be the 37% reporting EBIT impact, which barely changed from 2025 even as enterprise-wide scaling rose.16 More deployment is not yet producing measurably better results.

Spending is also concentrated. Yipit data cited by a16z shows the top 1% of AI users spending about eight times more than the top 10%. Within AI-native portfolio companies, top users spend $7,500–$9,000 a month, compared with $200–$400 for median users.21 Earlier a16z analysis of Ramp data found the top 10% of companies spending about 50 times more per employee on AI than the median.

Copilot: big seat counts, mixed results

Microsoft 365 Copilot is the clearest test of whether broad deployment produces returns. Microsoft says Copilot passed 30 million paid seats in its July 2026 quarter.16 Customer stories describe Bristol Myers Squibb reaching 96% adoption across 30,000 employees and Kohler reaching 98% adoption in six weeks.16 A Forrester study commissioned by Microsoft estimated a 116% ROI and nine hours saved per user per month for a composite organization.13

Independent data is less positive. A Recon Analytics survey of more than 150,000 enterprise users found that Copilot was the preferred tool only 8% of the time when ChatGPT and Gemini were also available.19 One February report said only about 3.3% of Microsoft 365 users who interact with Copilot Chat pay for it.20 The same coverage found that specialized tools such as Dragon Copilot in healthcare and phishing-triage security agents performed better than the general-purpose assistant.19 One buyer's guide says Fortune 500 adoption headlines show how far Copilot reaches, not how deeply it is used.15

This supports Wang's distinction. A general-purpose copilot is mostly a productivity and cost tool, which she calls the "latent" pool. The examples a16z highlights, Shopify's onboarding gain and ServiceNow's $1 billion in AI contract value, are tied to revenue or to the product itself.21

Our read

The a16z argument holds up on infrastructure. Backlogs, supply constraints and regular upward revisions to capex guidance all suggest demand is real for now.226 The weaker part is the link between that infrastructure and enterprise P&Ls. If only 2% of large companies measure AI's impact over time, the claim that revenue upside is "unbounded" is a hypothesis, not a finding. The panel's own synthesis says that if the gap does not close, the application-layer thesis weakens.22

For companies working on AI transformation, Wang's test is useful. Teams that put AI budgets into structural, customer-facing revenue opportunities, and track the results, will be the ones that justify hyperscaler spending. Teams that buy seats and report adoption rates will look like the 37% who see some EBIT impact, a share that has not changed in a year.16 With Wall Street increasingly expecting capex to peak in 20277, the next year or so of enterprise results will matter as much as the next round of capex guidance.