Big Tech's $690B AI Infrastructure Spending Spooks Analysts
$660–690B in 2026 AI infrastructure capex from four hyperscalers has analysts warning free cash flow could collapse before the spending converts to revenue.
The four hyperscalers — Alphabet, Amazon, Meta, and Microsoft — along with Oracle are collectively on course to spend between $660 billion and $690 billion on AI compute, data centres, and networking infrastructure in 2026, according to analysis published this week by the Economist, Futurum Research, and CNBC. That figure is nearly double the combined capital expenditure of the same companies in 2025 and dwarfs the current revenue base of the AI software and API economy it is supposed to serve.
The Economist published a detailed analysis on May 13 framing the spending as a structural cash flow problem rather than a growth investment. Free cash flow at Alphabet could fall to roughly $8.2 billion this year — down nearly 90 percent from recent levels — while Amazon is projected to run a cash flow deficit of between $17 billion and $28 billion as data centre construction and GPU procurement outpaces revenue generation.
The Numbers in Perspective
Futurum Research's breakdown shows Amazon leading the group at a projected $200 billion, with the majority going to AI-oriented data centre capacity. Alphabet follows at $175 to $185 billion, a figure the company itself revised upward multiple times during the first quarter, driven by a 55 percent surge in its cloud backlog. Microsoft is tracking above $120 billion — it spent $37.5 billion in its most recent single quarter alone — and Meta has committed between $115 billion and $135 billion, including a one-gigawatt data centre facility in Ohio and an expansion in Louisiana. Oracle, historically a software company rather than an infrastructure builder, has announced $50 billion in capital investment, a 136 percent increase over 2025.
The revenue side of the equation is where analysts grow uncomfortable. OpenAI's $20 billion in annualised recurring revenue — the most frequently cited evidence that AI software spending is accelerating — represents roughly three percent of the projected 2026 hyperscaler capex total. The combined pure-play AI vendor revenue for the year is unlikely to exceed $35 billion. Against $660 to $690 billion in infrastructure investment, that ratio implies a capital formation cycle that will take years to balance.
Capex-to-Revenue Ratios at Historic Extremes
What has changed is the ratio of capital expenditure to revenue. Meta's capex is running at approximately 54 percent of revenue; Microsoft at 47 percent; and Alphabet at 46 percent. These figures are materially higher than any prior period in the companies' histories during their growth phases, and they are being sustained at a time when the companies are mature, not startup, businesses.
The Economist analysis notes that the depreciation charges associated with today's spending will spread over the useful life of the assets purchased, typically three to seven years for data centre infrastructure. That means the headwind from 2026's capex will not fully materialise until 2029 or 2030, a timing that conveniently sits beyond the planning horizons that most corporate forecasters are asked to address. Analysts at multiple investment banks have flagged that investors are effectively being asked to underwrite a speculative infrastructure build that will be tested by adoption curves no one can predict with confidence.
Power as the Binding Constraint
One recurring finding in the analysis is that the primary constraint on AI infrastructure expansion is not capital or chips — it is electrical power. Microsoft has acknowledged an $80 billion unfulfilled Azure backlog that stems largely from power availability rather than insufficient demand. Data centres require not just electricity but a reliable, low-latency supply that is difficult to bring online quickly in most jurisdictions.
The Economist estimates that global data centre electricity consumption will double between 2022 and 2026, a trajectory that has prompted hyperscalers to pursue nuclear power agreements, offshore data centre proposals, and long-term contracts with utilities that bypass normal grid procurement processes. The Stargate Project — the OpenAI, SoftBank, Oracle, and MGX partnership targeting $500 billion in AI infrastructure investment by 2029 — has power acquisition at the centre of its site selection strategy.
Who Actually Benefits
One question the analysis raises without fully answering is who benefits from the spending wave. Chip manufacturers — Nvidia above all, but also AMD, Broadcom, and now Cerebras after its successful May 14 IPO — are the most obvious near-term winners. Data centre construction firms, power infrastructure providers, and cooling system suppliers are also capturing significant value.
The AI model providers and application developers sitting downstream of the infrastructure layer are running on subsidised compute at prices that reflect cloud provider ambitions rather than actual cost recovery. At some point, Futurum Research notes, pricing signals will need to reflect the real cost of the infrastructure being consumed — a normalisation that would compress AI application economics and force a reckoning with unit economics that current revenue multiples implicitly defer.
Whether that reckoning comes in 2027, 2028, or later will depend on how quickly enterprise AI adoption translates into contracted, recurring revenue at the application layer. The infrastructure is being built on the assumption that it will. The jury, as with the Musk trial currently underway in Oakland, is still out.