$700B — Big Tech's 2026 AI infrastructure spend, and no end in sight
Microsoft, Meta, and Google all raise 2026 AI capex guidance. Hyperscaler total tops $700B. SoftBank adds plans for new US AI/robotics IPOs.
$700B
Last February, when Microsoft put out an $80B 2025 AI capex plan, Wall Street's first reaction was "too much." The capex-outrunning-revenue concerns were loud and constant.
Fifteen months later, $80B is the small number.
Microsoft, Meta, and Google have all raised 2026 AI capex guidance. Total hyperscaler spend tracks past $700B per Fortune's late-April aggregation. SoftBank added plans to take new US AI/robotics companies public.
Satya Nadella (Microsoft CEO): "We're past the point of debating compute scale. We're committing." That's the new tone.
Masayoshi Son (SoftBank CEO): "Compute is the new oil. We're building the refineries." Marketing aside, the underlying capital flows aren't an exaggeration.
The center of gravity in AI competition has moved from model to compute access. This is the clearest signal yet.
Who's spending — Microsoft, Meta, Google, SoftBank
| Company | 2024 capex | 2025 capex | 2026 guidance | YoY |
|---|---|---|---|---|
| Microsoft | $55B | $80B | $115B | +44% |
| Meta | $40B | $65B | $110B | +69% |
| Alphabet | $52B | $75B | $120B | +60% |
| Amazon (AWS) | $48B | $90B | $130B | +44% |
| Oracle | $20B | $35B | $55B | +57% |
| Top-5 total | $215B | $345B | $530B | +54% |
| Other + SoftBank direct | $75B | $120B | $170B | +42% |
| Hyperscaler total | $290B | $465B | $700B+ | +50% |
Microsoft 2026 capex of $115B grows 44% YoY against ~18% expected revenue growth. Short-term ROIC is under pressure. Guidance went up anyway because the constraint is compute supply against the revenue ceiling.
Meta is the most aggressive at +69%. Llama 5 training plus Meta AI infra. Mark Zuckerberg (Meta CEO) on the Q1 call: "underbuilding compute costs more than overbuilding."
Alphabet raises capex on Cloud + Gemini + YouTube multimodal integration, leaning into in-house TPU.
AWS rides Anthropic Claude hosting demand. Trainium 5GW deal in April 2025 — single-customer demand alone justified more capex.
Oracle gets a step from OpenAI's multi-cloud (announced late April) — alongside AWS and Google as a core infra partner.
Where it goes
| Item | Share | 2026 estimate |
|---|---|---|
| GPU/AI chips (NVIDIA·AMD·in-house) | 50% | $350B |
| Datacenter construction/land | 18% | $126B |
| Power infrastructure (substations, renewables) | 12% | $84B |
| Cooling | 8% | $56B |
| Networking | 7% | $49B |
| Software/operations | 5% | $35B |
GPU and AI chips at 50% — $350B. NVIDIA's 2026 datacenter revenue guidance of $250–280B captures 70–80% of that. AMD MI400, Google TPU v6, AWS Trainium 3 split the rest.
Power infrastructure at 12% — $84B. Each datacenter needs 1–5GW. New substations have 6–9 month backlogs in parts of the US. SMR contracts are accumulating — Microsoft, Amazon, and Meta combined for ~12GW of pre-contracted SMR capacity in 2025 alone.
Wins and losses
For hyperscalers it's a short-term ROIC vs long-term share trade. Capex outrunning revenue holds for 1–2 years; year three needs revenue acceleration to make the math work.
NVIDIA gets unprecedented revenue visibility. $250–280B datacenter guidance, 2026 EPS consensus tracking up. In-house silicon (TPU, Trainium, MI400) is the medium-term share-loss risk.
Power utilities pull a structural increase in US/EU datacenter demand. SMR, hydro, renewables get sustained investment.
Datacenter hubs in Texas, Virginia, Ohio see industrial real-estate appreciation — and growing local fights over power and water.
Investors get clearer exposure to NVIDIA, AMD, TSMC, Broadcom, Applied Materials, and SMR-related names than to the hyperscalers themselves.
Past cycles — capex booms
Dotcom telecom capex, 1997–2001. WorldCom, Global Crossing, Qwest sank $200B+ into fiber. Crash followed in 2001. The fiber became the foundation of 2000s internet.
iPhone/mobile capex, 2007–2014. AT&T and Verizon spent $300B+ on 4G. ROIC pressure, but mobile internet revenue justified it.
Cloud v1 capex, 2015–2020. AWS, Azure, GCP combined $400B+. Skeptics overruled by the cloud revenue explosion.
Self-driving capex, 2018–2023. Waymo, Cruise, Aurora ~$50B. Revenue never showed up. Cruise effectively wound down by GM in 2024. Failed cycle.
Pattern: capex justifies itself when revenue acceleration follows. AI capex sits between dotcom telecom and cloud v1. Which way it tips is the next 2–3 years.
Counter-moves
China — Alibaba, Tencent, Baidu — combined 2026 capex around $80B. 1/9 the US figure, routed around export controls via Huawei Ascend and Cambricon.
Europe — Mistral, Aleph Alpha — can't compete on capital. EU is talking €30B sovereign AI fund — about 1/20 of US Big Tech.
Korea/Japan — Samsung, SK Hynix, NTT — focus on memory and infra supply. Stronger position as suppliers than as frontier-model builders.
Neoclouds — CoreWeave, Lambda, Crusoe — benefit from rental demand. Direct hyperscaler datacenter buildout pressures their mid-term spread.
Skeptics, by name
Aswath Damodaran (NYU professor, valuation specialist) — once capex exceeds 30% of revenue, dotcom-pattern risk rises. Some 2026 guidances cross that line.
Jim Chanos (Kynikos Associates) is publicly increasing shorts. Scenario: 2027 capex peak, 2028 correction.
Both grant AI revenue growth. Doubts focus on capex payback speed and short-term ROIC.
Stakes
- Wins: NVIDIA, AMD — record datacenter revenue. Texas/Virginia/Ohio datacenter hubs — real estate, jobs. SMR/renewable power — sustained investment lift.
- Loses: Big Tech short-term ROIC under pressure. Local communities — power/water conflicts. Climate observers — rising datacenter carbon footprint concerns.
- Watching: SEC — capex accounting guidance. Korean/Japanese memory supply chain — HBM demand. EU sovereign AI fund — capital-gap response.
What changes
Devs: GPU availability and pricing slowly improve, but H100/H200/B200 still go to Big Tech first. Smaller SaaS rely on mid-tier GPUs and spot instances.
Founders: niche opportunities in supplier/tooling/middleware — GPU efficiency SaaS, model-cost tracking, capex accounting tools.
Investors: AI infrastructure is a near-term theme, but watch 2027–2028 for cycle peak. NVIDIA, semi supply chain, power infra are the core exposures.
Consumers: minimal direct effect short-term, but datacenter neighbors will feel the power/water draw. In Korea, KT/SK/Naver Cloud capex acceleration is positive for industrial output and jobs.
3-Line Summary
- Big Tech 2026 AI capex tops $700B combined, +50% YoY.
- GPU and power infrastructure dominate; NVIDIA guidance steps up.
- Compute-as-power thesis hardens — short-term ROIC vs long-term share.
Sources
출처
AI 트렌드를 앞서가세요
매일 아침, 엄선된 AI 뉴스를 받아보세요. 스팸 없음. 언제든 구독 취소.