Stanford CS153 Frontier Systems | The Discipline of Delivering Value per Gigawatt
- Source
- youtube.com
- Author
- Stanford Online
- Date
Context
In a Stanford CS153 lecture, Amin Vahdat, introduced as the head of internal infrastructure at Google, argues that gigawatts of capacity is the wrong headline number. A gigawatt costs roughly $40 to $50 billion by figures discussed in the room, but what matters, Vahdat says, is the value delivered per dollar, such as users served. If chips fail or cannot be scheduled, useful , which Vahdat calls goodput, falls well below what was paid for, and idle capacity is a bug.
runs synchronously across many chips, so one failed node can stall the whole job. Historically services aimed for five nines of availability, which requires duplicate power feeds that leave about half the capacity unused. Vahdat says frontier labs would now often take double the capacity with a few days of downtime a year, while noting that this is not true of everyone. Vahdat makes related goodput points in a separate Sequoia interview.
- throughput — How much model output a provider can generate over time, commonly measured in output tokens per second.
- Training — The process that adjusts a model using examples and a measure of how well it performs.
- Model — A computational system that uses learned patterns to produce an output from information supplied to it.
videoStanford Webinar - Building Trustworthy AI: Navigating Security in the Agentic Era
videoStanford CS547 HCI Seminar | Autumn 2025 | HCI Systems in the Age of AI Code Generation
videoStanford CS25: Transformers United V6 I Advancing Science and Medicine with Collaborative AI Agents
videoStanford CS153 Frontier Systems | Andreas Blattmann from Black Forest Labs on Visual Intelligence
Checking sign-in…
Loading comments…


