Big Tech's $725B AI Capex Supercycle: What Happens When Four Companies Spend This Much

$725 billion is what Microsoft, Google, Meta, and Amazon will spend combined on AI capex in 2026 — up 77% from $410 billion in 2025. That's the short answer. The longer answer is what happens to the rest of the economy when four companies move that much money at once.

I track capital allocation for a living, and I've never seen four companies commit this much capital to a single category simultaneously — not during the dot-com buildout, not during the cloud land grab of the 2010s. This isn't normal corporate spending anymore; it's a supercycle with its own gravity, pulling in power utilities, chip fabs, debt markets, and eventually the Fed.

Big tech AI capex supercycle: what's actually happening in 2026

The big tech AI capex supercycle refers to the coordinated, accelerating capital spending by Microsoft, Amazon, Alphabet, and Meta on AI infrastructure — GPU clusters, custom silicon, and data centers — which is set to hit roughly $725 billion combined in 2026, up 77% from about $410 billion in 2025 and nearly triple the $226 billion these four companies spent in 2024.

Individually: Amazon is guiding to roughly $200 billion, Microsoft to about $190 billion, Alphabet to $175–185 billion, and Meta to $115–135 billion for calendar 2026, based on Q1 2026 earnings calls. Add Oracle, increasingly treated as a fifth hyperscaler because of its OpenAI infrastructure commitments, and combined hyperscaler capex approaches $700–765 billion depending on whose model you use.

For context on how fast this moved: Microsoft alone spent about $80 billion in fiscal 2025 — at the time the largest single-year capex figure in corporate history. Its FY2026 guidance of $190 billion is more than double that, in a single follow-up year. Track the underlying company financials on our Big Tech Earnings dashboard.

2026 AI capex by company: the big tech supercycle scoreboard

Figures are 2025 actuals and 2026 company guidance blended from Microsoft, Amazon, Alphabet, and Meta Q1 2026 earnings calls, Nvidia FY2025/FY2027 Q1 filings, and reporting from Tom's Hardware, CNBC, and Fortune. Oracle and combined-total figures are third-party estimates and may differ from GAAP-reported capex.

Where the AI capex supercycle money actually goes

The overwhelming majority of this spending is GPU silicon and the data centers built to house it. Nvidia captures an estimated 41.5% of every dollar hyperscalers spend on AI hardware, a function of its 85–90% share of the data center GPU market. Nvidia's data center segment generated $115.2 billion in fiscal 2025 revenue and posted a record $75.2 billion in a single quarter in fiscal 2027 — up 92% year over year — with hyperscalers accounting for roughly half of that quarter's data center revenue.

The big tech AI capex supercycle's trajectory: 2023 to 2031

This isn't a one-year spike. Combined hyperscaler capex has roughly tripled since 2024, and Goldman Sachs' baseline model implies annual AI capex could still reach $1.6 trillion by 2031 even under conservative assumptions — up from an implied $765 billion figure for 2026 in some broader hyperscaler-plus-neocloud models. Whether that trajectory holds depends heavily on one contentious accounting assumption: how long a GPU actually lasts.

The depreciation debate inside the big tech AI capex supercycle

The single biggest risk flagged by skeptics isn't demand — it's accounting. Hyperscalers currently depreciate Nvidia-based servers over 5 to 6 years, a window that's been progressively extended from the 3–4 years used a decade ago. Investor Michael Burry argues the real economic life of this hardware is closer to 2–3 years, given Nvidia's accelerating release cadence: A100 (2020), H100 (2022), Blackwell (2024), Blackwell Ultra (2025), with Vera Rubin slated for 2026 and Vera Rubin Ultra for 2027.

Burry estimates roughly $176 billion in understated depreciation and overstated profits across the industry between 2026 and 2028. Goldman Sachs' own sensitivity analysis shows that shortening assumed GPU life from 5 years to 3 years would push cumulative 2026–2031 depreciation from about $3 trillion to nearly $4 trillion — a swing large enough to move reported earnings at every hyperscaler simultaneously. The Federal Reserve named AI infrastructure spending a top systemic risk in 2026, ranking it just behind geopolitical threats.

The AI capex supercycle is now a power problem, not a chip problem

Once you commit $725 billion to GPUs and buildings, the next bottleneck is electricity. US data center power demand is projected to climb from 31 gigawatts in 2025 to 41 gigawatts in 2026, and total combined US data center energy demand is expected to nearly double from 80 GW to 150 GW between 2025 and 2028, per Goldman Sachs and industry capacity tracking. Data centers already account for roughly 4.1% of US peak summer power demand in 2025, rising to an estimated 5.3% in 2026 — and some models put data centers at 6.7–12% of all US electricity consumption by 2028.

Grid capacity additions are scrambling to keep pace: realized capacity additions were 6.4 GW in 2024 and 8.5 GW in 2025, but scheduled additions jump to 13.6 GW in 2026 and a projected 36.3 GW in 2027. That's the real constraint on the AI capex supercycle right now — it's no longer whether Nvidia can ship enough GPUs, it's whether utilities in Virginia, Texas, and the Pacific Northwest can permit, finance, and build power generation fast enough to energize the data centers those GPUs sit in. Several hyperscalers have responded by signing direct power purchase agreements with nuclear, gas, and even restarted plants — multi-decade commitments that only make financial sense if AI compute demand keeps growing at something close to today's rate.

What happens when four companies spend this much at once

Four things happen simultaneously, and none of them are contained to tech. First, Nvidia's revenue concentration risk compounds: with roughly half of a $75 billion quarterly data center haul coming from a handful of hyperscaler customers, any one of the four slowing its buildout materially dents Nvidia's growth rate — which is exactly why Nvidia has started splitting its data center reporting into hyperscaler and non-hyperscaler ("ACIE") buckets to manage investor expectations.

Second, power becomes the constraint, not chips. Data center electricity demand tied to this buildout is straining regional grids from Virginia to Texas, pulling utilities into multi-decade capacity commitments they wouldn't otherwise make — a bet on AI demand persisting for far longer than any single GPU's depreciation schedule.

Third, financing structures get more exotic. As capex outpaces free cash flow at several of these companies, more of the buildout is shifting to off-balance-sheet vehicles, special-purpose financing, and circular customer-supplier-investor deals (Nvidia investing in customers who then buy Nvidia chips) that make the true leverage in the system harder to see from the outside.

Fourth, and most relevant if you invest in or build AI companies: valuation multiples across the AI stack are now implicitly pricing in this capex continuing at something close to current growth rates. If the depreciation assumptions crack even modestly, or growth decelerates from 77% toward something closer to 20–30%, every AI-adjacent valuation built on today's revenue trajectories gets repriced at once. Check current private-market pricing on our AI Valuations dashboard, and compare it against public SaaS multiples on SaaS Valuations.

The honest read, after two decades of watching capital cycles like this: none of the four companies can afford to blink first. Whoever pulls back capex earliest signals to the market that they believe their own AI product roadmap is weaker than a competitor's, which is a far more damaging signal to send than one more quarter of compressed free cash flow. That dynamic — a capex arms race where slowing down is read as a confession of weakness — is exactly what tends to push a supercycle past the point any individual participant would choose in isolation. It's also why I'd bet on the growth rate decelerating well before any of the four actually cuts nominal spend.