AI Capex
The billions hyperscalers and AI labs spend each year on GPUs, datacenters, and training clusters · the #1 driver of AI spending narrative.
The billions hyperscalers and AI labs spend each year on GPUs, datacenters, and training clusters · the #1 driver of AI spending narrative.
Basic
Hyperscaler AI capex 2025: Microsoft $80B, Google $75B, Amazon $100B, Meta $60B, combined $315B+ · mostly going to NVIDIA GPUs and data center buildout. xAI Colossus and OpenAI Stargate add tens of billions more. This capex is the physical buildout behind every AI training run and serving capacity announcement.
Deep
Capex accounting: GPUs + data center shell + power infrastructure + networking + cooling. NVIDIA captures 70%+ of GPU capex via direct sales; the rest goes to AMD MI300X, Google TPU (internal), and custom chips (Trainium, Maia). Data center shell + power can be another 30% of total spend. The capex-to-revenue ratio across AI hyperscalers exceeded 50% in 2025 · historically extreme. This gap drives the AI Bubble Index component on BenchGecko.
Expert
Capex depreciation: GPUs typically 5-6 year useful life, 2-3 year rapid obsolescence, driving aggressive writedowns. 2026 estimated AI capex: $500B+ globally. ROI on AI capex is uncertain · training runs deliver capability but revenue scales depend on downstream demand. Historical parallels: telecom 1999-2001 ($500B capex, 90% stranded) vs railroads 1840s (bubble + productive infra). AI capex has strong enterprise demand but monetization unclear past 2027.
AI capex as % of revenue is at dotcom-era levels. Every earnings call is scrutinized for capex discipline.
Depending on why you're here
- ·GPU + datacenter + power + networking · full stack capex
- ·Depreciation 5-6 year, aggressive obsolescence of 2-3 years
- ·NVIDIA captures 70%+ of GPU capex
- ·Your pricing as a customer reflects capex amortization
- ·Capex decisions at hyperscalers drive your future serving costs
- ·Watch capex / revenue ratio as a leading indicator of pricing pressure
- ·Capex is the #1 question on every hyperscaler earnings call
- ·ROI uncertain past 2027 · telecom-era parallel real
- ·BenchGecko Bubble Index uses capex-to-revenue as 10% component
- ·The massive amount AI companies spend on computers
- ·Hundreds of billions per year
- ·Why "AI is expensive" · it actually is, at scale
AI capex at $500B+/year is either the railroads of our era or the Cisco gear of 1999. Nobody knows yet · but BenchGecko tracks it daily.
Understanding capex cycles tells you which AI companies are about to cut pricing (high depreciation pressure) vs hold firm (delayed capex).