The Numbers, Plainly Stated
The 2026 guidance stack, from company earnings calls through mid-2026: Amazon around $200 billion, Alphabet up to ~$185–190 billion, Microsoft in a similar band on a calendar-year basis, Meta guiding $115–135 billion and revising upward, citing — among other things — surging memory-chip prices. Combined: roughly $725 billion, against ~$410 billion in 2025, which was itself a record that seemed absurd a year earlier.
The demand signal funding all this is equally concrete. Nvidia's most recent quarter (reported May 2026): $81.6 billion in revenue, up 85% year over year, with $75.2 billion from data centers alone, up 92% — and a market capitalization hovering around $5 trillion, the largest in the world. Jensen Huang told the 2026 GTC audience Nvidia expects "at least" $1 trillion in revenue from its Blackwell and Rubin generations through 2027, with Rubin shipping in the second half of 2026. On the buyer side, model-provider revenue is scaling too: Anthropic reported passing $30 billion in annualized revenue in April 2026 (up from ~$9 billion at the end of 2025), with OpenAI around $25–33 billion depending on the month and whose accounting you accept — the two companies publicly dispute each other's numbers, which tells you how competitive the scoreboard has become.
Run the CAGR: hyperscaler capex went from ~$150B (2023) to ~$230B (2024) to ~$410B (2025) to ~$725B (2026 guided). That is a three-year compound growth rate north of 65% per year — for context, no comparable capital line in corporate history has compounded at that rate at this absolute scale.
Where the Money Actually Goes
"AI capex" collapses several very different assets into one line item, and the distinction matters for judging the risk:
- Accelerators (roughly half or more). GPUs and custom silicon (TPUs, Trainium) — the shortest-lived asset, depreciated over ~5–6 years, with real economic-life debate: a 2022 GPU still earns, but at collapsing rents. This is where the "what if it's obsolete in three years" fear concentrates.
- Buildings, land, and power infrastructure. Multi-decade assets: shells, substations, cooling. If AI demand disappoints, these still hold value — nobody un-needs electricity and industrial land near fiber.
- Memory and networking. The 2026 surprise: high-bandwidth memory and DRAM entered a genuine shortage, with hyperscalers absorbing supply and consumer prices rising — Meta explicitly blamed memory costs for raising its capex guidance. The memory makers are running their own mini-supercycle.
- Power procurement. Not always capex-proper, but the binding constraint: every hyperscaler has now signed nuclear PPAs or SMR deals (9.8+ GW committed as of May 2026), and wholesale power prices near hyperscale clusters have spiked as much as 267% — the buildout is now gated by megawatts, not money.
Financing shifted too, and this is the structurally new part: Meta raised $30 billion in a single 2025 bond sale, Oracle is funding its OpenAI-anchored buildout with tens of billions in debt, and pure-play GPU clouds like CoreWeave carry ~$25 billion of borrowings against Nvidia collateral. The 2023–2024 phase was cash-flow-funded by the most profitable companies on earth; the 2025–2026 phase increasingly runs through credit markets, private credit, and SPVs — which is exactly the transition that historically converts a boom into a cycle.
The Bubble Arithmetic, Honestly
The bear case is simple division. Direct AI revenue at the model layer — generously, $60–80 billion annualized across the major labs in mid-2026 — sits an order of magnitude below annual infrastructure spend. Someone is either early or wrong. The dot-com rhyme writes itself: in 1999–2001, telecoms buried $100B+ of fiber on demand forecasts that were directionally right but a decade early, and the vendors financing their own customers (Lucent, Nortel) got destroyed when the music stopped. Circular financing exists today too: Nvidia invests in OpenAI, Anthropic, and the neoclouds that buy its chips; clouds book each other's capacity; backlog numbers double-count across the chain.
The bull case is that this time the demand is measured, not forecast. Google reported token processing growing ~50x year over year into 2025; every provider is capacity-constrained; enterprises queue for GPU allocations rather than the reverse. Utilization — the number that killed the fiber boom, which ran at ~2–5% lit capacity — is effectively saturated for AI compute. Hyperscaler cloud segments are growing 25–30%+ on enormous bases, and the capex is being funded primarily by ~$500B+ of combined annual operating cash flow, not primarily by leverage. Fiber's problem was no revenue; AI's problem is merely that revenue trails spend by a couple of years.
The honest synthesis: the aggregate bet is probably rational; the distribution of outcomes across participants is not. If frontier demand keeps compounding, $725B looks cheap. If it plateaus even briefly, the shortest-duration assets (accelerators bought with debt by thinly capitalized intermediaries) get repriced first and hardest, while the hyperscalers absorb the pain as a depreciation drag on otherwise healthy earnings. Watch depreciation-to-revenue ratios, debt-funded share of capex, and the gap between reported backlog and recognized revenue — those three series will tell you which scenario is unfolding well before headlines do.
What It Means If You're Not a Hyperscaler
- For builders: the paradox of the supercycle is that it keeps making your inputs cheaper — token prices fall as capacity lands. The strategic play is architecting to ride the cost curve (routing, caching, small-model-first) rather than locking into any single provider's economics.
- For companies buying cloud: AI capacity is displacing general-purpose refresh cycles in some regions; negotiate reserved capacity early, and expect power-constrained regions (Northern Virginia, Dublin) to carry premiums.
- For investors: the S&P 500 is now roughly one-third exposed to seven AI-linked companies, so "not investing in AI" is no longer a choice index holders get to make. The second-order plays of 2026 — power, memory, cooling, grid equipment — have outperformed the headline names, which is typical of infrastructure cycles as they mature.
- For everyone: electricity is the spillover. Data-center demand is repricing power in specific metros, and the political economy of "AI raised my utility bill" is becoming a real constraint on siting — one reason the buildout is going nuclear, literally.
Compounding at 65%+ cannot continue indefinitely — that is not pessimism, it is arithmetic. The question is whether it decelerates into a plateau (fine) or air-pockets (ugly). Model the scenarios yourself; the difference between 20% and 60% growth sustained for three years is the difference between a big industry and the largest capital formation event in human history.