
The math is brutal: if you spend $200M training a model but $2B annually running it at scale, your competitive advantage shifts to whoever can reduce that $2B – not whoever trained the model.
The Economics
Model Training: $100M-$500M one-time cost, 6-18 month lifespan, low defensibility (models leak, open-source catches up)
Inference Infrastructure: Billions in sustained capex, 5-10 year amortization, high defensibility (physical assets, long-term contracts)
The competitive advantage has moved from models to compute economics. Infrastructure advantages compound while model advantages decay.
Amazon’s Scale
– Project Rainier: 500,000 Trainium2 chips in one cluster
– Trainium2: Multi-billion dollar business, grew 150% QoQ
– AWS: 3.8 gigawatts of power capacity added in 12 months
– Enterprise: 90% of Fortune 500 already on AWS
Key Takeaway
As Five Defensible Moats shows, infrastructure is hardest to replicate. In every gold rush, the money was made selling picks and shovels. Amazon cornered the AI shovel market.
Source: Amazon’s AI Superstructure on The Business Engineer









