← Back to feed

The hyperscalers are spending like utilities and generating free cash flow like startups. That gap is now large enough to matter.

Amazon, Alphabet, Microsoft, Meta, and Oracle are on track to spend somewhere between $700 billion and $900 billion on capital expenditure in 2026, up 36% over 2025 according to CreditSights estimates. About 75% of that — roughly $450 billion — is directly tied to AI infrastructure: GPU clusters, custom accelerators, data centers, and the power and cooling to run them.

The revenue to justify that spending does not exist yet.

The Free Cash Flow Numbers

The clearest signal is in projected free cash flow, not in capex totals.

Alphabet generated $73.3 billion in free cash flow in 2025. With $175-190 billion in 2026 capex guidance — and cloud backlog that doubled quarter-over-quarter to $462 billion — its projected 2026 FCF is approximately $8 billion. An 89% decline in a single year. The backlog is real; the cash flow is not there yet.

Amazon generated $38 billion in free cash flow in 2025. Its 2026 capex guidance is $200 billion — the largest absolute commitment in the group, more than doubling its 2025 outlay. At that pace, Amazon’s 2026 free cash flow is projected negative, a position the company has not been in during the AWS era.

Microsoft is the most resilient: $74 billion in FCF in 2025, projecting approximately $53 billion in 2026 against $190 billion in capex commitments. Still declining, but not collapsing.

Meta generated $52 billion in FCF in 2025 and projects $20-30 billion in 2026 against $115-145 billion in capex — a company spending 45-50% of its expected revenue on capital infrastructure. The ratio belongs to power utilities and industrial manufacturers, not technology companies.

The Revenue Gap

Sequoia Capital’s David Cahn calculated in 2024 that there was a $200 billion annual revenue gap between AI infrastructure spending and actual AI ecosystem sales. His updated analysis puts that figure at $600 billion.

The calculation is direct. Take Nvidia’s run-rate revenue forecast, multiply by approximately 2x to capture the full cost of AI data centers (GPUs are roughly half the total cost of ownership; energy, buildings, and cooling make up the rest), multiply by 2x again to account for a 50% gross margin requirement for the end-user of compute. The result is the annual revenue that the AI ecosystem needs to generate. The current run rate falls far short.

Bain Capital puts the target even higher. Its analysis of hyperscaler data center capacity under construction concludes that the infrastructure being built today will need to generate $2 trillion in annual revenue by 2030 to justify its cost. Current annual AI-related revenue across the ecosystem runs at $300-400 billion. Bridging that gap requires a roughly 5x revenue increase in four years — while the infrastructure keeps being built.

The Debt Machine

The free cash flow decline is being bridged with bond issuance. Hyperscalers issued approximately $121 billion in bonds in 2025, roughly four times the five-year average. Meta completed a $30 billion offering; Oracle a $25 billion deal in early 2026. Oracle’s 5-year credit default swap has tripled since September, and Barclays has flagged a scenario where Oracle exhausts cash by November 2026 absent additional financing.

Morgan Stanley projects the sector may need $1.5 trillion in additional debt over coming years to sustain the buildout.

This is not 1990s telecom speculation — the hyperscalers have balance sheets and contracted demand that those companies lacked, and the revenue is growing fast. Anthropic’s annualized revenue is reportedly approaching $45 billion; that cash flows back into Alphabet and Amazon infrastructure commitments. The demand is real.

But the timeline question is live. Goldman Sachs projects meaningful AI productivity gains starting in 2027, not 2026. The infrastructure is being built to serve demand that the models and the enterprise AI market still need to unlock. The gap between committed capital and realized revenue will widen through the year before it narrows.

What Investors Are Watching

The market has already begun to differentiate. Alphabet and Amazon are being rewarded where cloud revenue provides visible return on investment. Meta, where the ROI is internal and harder to quantify, is being penalized. Those judgements will continue to be refined as 2026 and 2027 earnings reveal how closely the revenue trajectory tracks the capex commitment.

The hyperscalers are roughly 45-55% through the accelerated compute cycle on a dollar basis as of year-end 2026, by one estimate using verified SEC filings and NVIDIA quarterly results. Peak capex is most likely 2027-2028 at $700-800 billion, followed by normalization. The cycle is not reversing. The question is whether the revenue arrives in time to catch the cash.