Microsoft, Google, Meta, and Amazon are guiding to more than $380B in combined capital spending for 2026 — up from roughly $246B in 2025 — and the overwhelming majority of the increase is AI infrastructure.
That is the short answer. The longer answer is more interesting: four companies with very different business models have arrived at the same conclusion at the same time, and the reason they are all spending at once tells you more about the AI race than any single benchmark does.
AI hyperscaler capex: Microsoft, Google, Meta, and Amazon compared
For 2026, the four largest US hyperscalers are collectively guiding to north of $380B in capital expenditure, the bulk of it AI data centers, GPUs, and power. Amazon is the largest single spender at roughly $118B, Microsoft and Google sit near $90–100B each, and Meta — which has no public cloud to monetize — is still guiding to roughly $70–72B. Compared with about $246B in 2025, that is a 55%+ year-over-year jump concentrated almost entirely in AI.
| Company | 2025 Capex | 2026 Capex (guide) | YoY | Primary AI driver |
|---|---|---|---|---|
| Amazon | ~$83B | ~$118B | +42% | AWS / Trainium 2 / Anthropic |
| Microsoft | ~$64B | ~$95B | +48% | Azure / OpenAI / Maia |
| Google (Alphabet) | ~$52B | ~$93B | +79% | Cloud / TPU v6 / Gemini |
| Meta | ~$47B | ~$72B | +53% | Llama / ads ranking / MTIA |
| Combined | ~$246B | ~$378B+ | +55% | AI compute + power |
| Nvidia share of spend | ~$70B+ | ~$130B+ | +85% | GB200 / GB300 systems |
Figures are approximate full-year capex based on company guidance and analyst consensus as of mid-2026; Nvidia share is estimated GPU spend captured across all four buyers.
Why Microsoft, Google, Meta, and Amazon are all spending at once
The simplest explanation is the right one: each company fears being structurally short on compute more than it fears overspending. When Satya Nadella says Microsoft is "capacity constrained," he means Azure is turning away AI revenue it cannot serve. That is the worst position to be in — demand you cannot fill compounds into lost platform share. So all four are buying ahead of demand rather than behind it.
But the motivations differ underneath the shared number. Amazon, Microsoft, and Google are building rentable capacity — every GPU hour they install can be resold to enterprises through AWS, Azure, and Google Cloud. Their capex is, in theory, a cost of goods that throws off cloud margin. Meta is the outlier: it has no public cloud, so its $70B+ is pure internal investment in ad-ranking models and the Llama family. Meta cannot resell a single GPU, which is exactly why investors scrutinize its spend the hardest.
Compute is the binding constraint
Model quality and cloud revenue both scale with GPU supply, not headcount
First-mover lock-in
Enterprises that build on Azure or Bedrock now are expensive to migrate later
Power is the new bottleneck
Securing gigawatts of grid interconnect takes years, so buying early is buying optionality
Asymmetric career risk
No CEO wants to be the one who under-built and lost a platform decade
What $380B in AI capex actually buys
The headline number is abstract until you decompose it. Roughly half of hyperscaler capex is servers and chips. The other half is the physical plant — shells, land, cooling, networking, and increasingly the power infrastructure that has become the real constraint. A single large AI campus now runs $5–10B and can draw anywhere from 500MW to more than 1GW, enough to power a mid-sized city.
Compute & Silicon (~50%)
The single largest line item across all four buyers
In-house chips to cut Nvidia dependence and cost
Data Center Plant (~35%)
$5–10B per hyperscale AI campus
Liquid cooling is now standard for GB-class racks
Power & Grid (~15%)
Power is the binding constraint, not chips
The shift in the bottleneck is the most important story buried in these numbers. Two years ago the constraint was Nvidia allocation. Today it is power and grid interconnects — which is why all four hyperscalers are signing nuclear PPAs, restarting decommissioned plants, and building gas turbines on-site. You can track the broader picture on the AI Spending dashboard.
Does the hyperscaler AI capex actually pay off?
This is the question every investor is now asking on every earnings call, and the honest answer is: partially, and not yet at the scale the spending implies. AI-related cloud revenue is real — Azure, Google Cloud, and AWS each cite tens of billions in AI run-rate — but it has not caught up to the depreciation wave that $380B+ in annual capex creates. When you spend $100B on assets that depreciate over five to six years, you book roughly $17–20B in annual depreciation per company before you have earned a dollar back.
That is why free cash flow has compressed even as revenue grows. Meta's FCF tightened as capex outran operating cash. Microsoft's capex-to-operating-cash-flow ratio climbed past 50%. The market has tolerated this so far because cloud revenue is still accelerating and because the alternative — under-investing — looks worse. But the patience is conditional. The moment AI cloud growth decelerates while capex keeps climbing, the multiple compression will be brutal. Compare quarterly trends on the Big Tech Earnings dashboard.
The Bull Case
- ✓ Cloud AI run-rate growing 40%+ YoY across all three clouds
- ✓ Compute is genuinely capacity-constrained, not speculative
- ✓ Custom silicon cuts unit cost 30–50% vs. merchant GPUs
- ✓ Each $1 of capex locks in multi-year enterprise contracts
The Bear Case
- ✕ $380B+ in annual spend front-loads massive depreciation
- ✕ FCF compression already visible at Meta and Microsoft
- ✕ GPU refresh cycles may shorten useful life below 6 years
- ✕ A demand air-pocket would strand billions in capacity
What this means for founders and investors
For founders, the $380B is a tailwind and a warning. The tailwind: compute is getting cheaper per unit and more abundant, so AI-native products that were uneconomic in 2024 pencil out in 2026. The warning: the hyperscalers are vertically integrating up the stack — buying labs, building models, and bundling inference into their clouds. If your startup's only moat is access to a frontier model, you are renting your business from a company spending $100B to make that model a commodity.
For investors, the read-through is that the AI trade has bifurcated. The picks-and-shovels layer — Nvidia, custom silicon, power, cooling, data center REITs — has visible demand backed by signed capex. The application layer is where the returns ultimately have to show up, because that is where the $380B is supposed to monetize. If enterprise AI adoption stalls, the entire capex stack re-rates downward at once. That correlated risk is the thing the synchronized spending creates: when four companies build the same bet simultaneously, they also share the same downside.
The hyperscalers are not spending $380B because the ROI is proven.
They are spending it because being short on compute is the one mistake none of them can afford to make.
Track AI infrastructure spending on the AI Spending Dashboard and quarterly results on Big Tech Earnings at Value Add VC. Originally published in the Trace Cohen newsletter.