Stanford CS153 Frontier Systems | The Discipline of Delivering Value per Gigawatt

Measure daily active users per gigawatt, not gigawatts alone

With infrastructure costing $40–50 billion per gigawatt, the critical metric is business value—such as daily active users or revenue—delivered per dollar spent, rather than raw capacity deployed.

Treat idle capacity as a major outage

Google considers sub-96% node allocation a major outage because unused capacity is wasted capital; the focus must be on 'goodput' and ensuring every accelerator delivers value rather than sitting idle waiting for data or orchestration.

Flops without bandwidth are wasted money

Citing Amdahl's 1967 law (1 MIPS requires 1 MB/s I/O), Vahdat stresses that modern infrastructure must balance compute with HBM bandwidth, SRAM, and network capacity to prevent starvation and achieve true utilization.

Sparse models break hardware balance assumptions

The shift to mixture-of-experts architectures demands more memory bandwidth relative to compute than current hardware provides, explaining why some clusters operate at only 11% Model FLOPs Utilization (MFU).

Frontier labs choose capacity over five-nines uptime

Unlike enterprise services requiring 99.999% reliability (30 seconds/year downtime), AI training workloads now prioritize throughput, accepting 99.9% reliability (3.65 days/year) in exchange for doubling usable capacity.

Synchronous training breaks loose coupling

Traditional distributed systems assumed fungible nodes where any rack could fail unnoticed, but synchronous AI training makes every accelerator special—if one fails, the entire job stops, requiring new failure management approaches.