AI infrastructure spending has reached unprecedented levels. Here’s where the money is going.
FY2026 CapEx Comparison
| Company | CapEx | YoY Change | Focus |
|---|---|---|---|
| Microsoft | $120B+ | +60% | Azure AI, OpenAI infrastructure |
| Meta | $72.2B | +84% | AI training, Reality Labs |
| $75B (est) | +50% | TPU clusters, Gemini training | |
| Amazon | $85B (est) | +45% | AWS expansion, Trainium |
Total Big 4: $350B+ in single year
Infrastructure Commitments
| Company | Commitment | Timeframe |
|---|---|---|
| Microsoft | $625B backlog | Multi-year |
| OpenAI (Stargate) | $500B | Through 2029 |
| Meta | $200B+ total | Through 2028 |
| Oracle (OpenAI) | $300B | 5-year |
Energy Commitments
| Company | Power Strategy | Capacity |
|---|---|---|
| Meta | Nuclear secured | 6.6 GW |
| Microsoft | Three Mile Island + renewables | ~2 GW |
| SMR + renewables | ~1.5 GW | |
| Amazon | Nuclear exploration | TBD |
Custom Silicon Investment
| Company | Chip | Status | Purpose |
|---|---|---|---|
| TPU v5 | Production | Gemini training | |
| Amazon | Trainium2 | Production | AWS AI workloads |
| Meta | MTIA | Production | Inference optimization |
| Microsoft | Maia 200 | Production | Azure AI acceleration |
The Pattern
Every hyperscaler is racing to:
- Secure power (energy is the new bottleneck)
- Build custom silicon (reduce NVIDIA dependency)
- Lock in compute commitments (guarantee capacity)
Data compiled from Microsoft’s Frontier AI Dilemma and The Re-Engineering of Meta on The Business Engineer.
More Resources
