Amazon CEO Andy Jassy on the $220 Billion AI Buildout
Andy Jassy argues that AI demand is so large that Amazon's $220 billion capex still leaves demand unmet, and that companies must rebuild their organizations and products around AI.
Watch on YouTube ↗Charlie's summary
- ▶ 1:31
A rushed buildout became a blessing: it forced Amazon to regionalize its flat national fulfillment network
Andy says Amazon doubled its fulfillment network in about 18 months and built a UPS-sized last-mile network in about 20. Costs to serve suffered. That pain led Amazon to re-examine its “most closely held beliefs” and regionalize the US network. Items sat closer to customers, which cut costs and sped delivery and unleashed demand. Amazon chose speed for customers and then fixed the architecture after.
- ▶ 12:47
GenAI is growing several hundred times faster than early AWS, and capex of $220 billion still leaves demand unmet
Andy says AWS was a $60 million run rate business after three years. Its AI business is over $25 billion after three years of this wave. He says the risk is bounded: data centers are built about two years ahead, but servers and networking are bought only months ahead, with good demand visibility. 2026 is sold out, 2027 largely sold out, and a good chunk of 2028 too.
- ▶ 16:21
Data center opposition rests on myths, and the US cannot afford not to build
Andy compares the buildout to the 1950s highways and bridges, and warns that China would set the norms otherwise. He says golf courses use 200 times more water than Amazon's data centers, backup generators run about 10 hours a year, and grid underinvestment is the real cause of energy rate rises. He adds that Amazon has paid about $3 billion in tax in St. Joseph County, Indiana.
- ▶ 21:28
AI demand is a barbell, and the middle, enterprise production workloads, will be the largest segment
Andy says one end is AI labs and a few runaway apps consuming huge compute. The other end is enterprises using AI for cost avoidance, such as customer service and fraud. The middle is existing production workloads, which mostly don't use inference pervasively yet. Within five to seven years he expects most enterprise applications to use inference widely.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.