scripod.com
Google's AI Infrastructure Chief, Amin Vahdat, on the Physics & Economics of Frontier AI

Overview

Shownote

Highlights

Transcript

Chapters

Pins

Google's AI Infrastructure Chief, Amin Vahdat, on the Physics & Economics of Frontier AI

Training Data

18 HOURS AGO
Training Data

Training Data

18 HOURS AGO
OverviewShownoteHighlightsTranscriptChaptersPins

Shownote

At 100,000-accelerator scale, something fails multiple times an hour, which is why Google's Chief Technologist for AI Infrastructure Amin Vahdat thinks FLOPS is a vanity metric. The metric that matters is what he calls goodput: the useful work a workload a...

Highlights

Google AI infrastructure chief Amin Vahdat explains why frontier-scale computing is an exercise in workload forecasting, system reliability, power planning, and hardware–model co-design—not simply buying more FLOPs.
00:02
Specialization is a bet on the future
04:13
AI racks reshape the building itself
09:01
Goodput, not theoretical FLOPS, is what matters
16:35
The TPU began as a contrarian bet
23:37
Co-design unlocks huge optimization opportunities
29:05
Changing the chip architecture mid-flight
35:06
Agents remove the human rate limit
39:17
Optical switching leaves the bits untouched
42:59
Power is the fundamental constraint
49:51
Inference cannot rely on retired training clusters alone
1:02:44
The 2036 rack may consume multiple megawatts

Chapters

Introduction
00:00
What makes a data center an AI data center
01:47
Goodput, not FLOPS: holding yourself accountable when something fails every hour
05:30
Doubling token capacity every six months, and where the gains actually come from
11:52
The TPU bet: from a contrarian call in 2013 to splitting 8i and 8t
15:32
The case for and against co-design
23:30
Shoulder to shoulder with DeepMind: intercepting chips mid-flight
26:11
Long-horizon agents change the shape of the data center
34:16
Optical circuit switching and the state of networking
37:50
Power is the binding constraint: utilities, gigawatts, and sizing a data center
42:35
Training vs. serving clusters, seven-year-old TPUs, and open standards
49:23
Orbital data centers and the supercomputer of 2036
58:29

Transcript

Amin Vahdat: In hardware, again, as you know, there is this opportunity. Where the more you specialize to a particular workload, the less flexible it is, the faster, the more power efficient the hardware is going to be. So it is this art, and it's this pro...