Uploaded June 2026 | Updated September 2026, 3 weeks ago
The Physics of Thought and the Architecture of Intelligence
Amin Vahdat, Google
We are in a time of massive transition across technology and society. This talk explores key lessons and insights that we can learn from history as we stand at the precipice of a new epoch of computing, the era of intelligence. To prepare for this future, we first start with a historical view for parallels to inform both what we work on but also how we do so. Building on this historical perspective, we delve into the architecture of intelligence, showing that efficient scaling of compute is central to advances in the field. This scaling will come from advances in many fields, though we focus on the central role networked systems play in scaling intelligence, from TPU on-chip networks to rack-scale serving, to ML training supercomputers, to distributed control and transport over Gigawatt-scale regional computing hubs. We conclude the talk with research opportunities for the future, again drawing lessons from history to ensure that we as technologists also focus on addressing the costs of scaling intelligence, from sustainability to safety to security to societal policy.
View the full NSDI '26 program at usenix.org/conference/nsdi26/technical-sessions
The Physics of Thought and the Architecture of Intelligence
Amin Vahdat, Google
We are in a time of massive transition across technology and society. This talk explores key lessons and insights that we can learn from history as we stand at the precipice of a new epoch of computing, the era of intelligence. To prepare for this future, we first start with a historical view for parallels to inform both what we work on but also how we do so. Building on this historical perspective, we delve into the architecture of intelligence, showing that efficient scaling of compute is central to advances in the field. This scaling will come from advances in many fields, though we focus on the central role networked systems play in scaling intelligence, from TPU on-chip networks to rack-scale serving, to ML training supercomputers, to distributed control and transport over Gigawatt-scale regional computing hubs. We conclude the talk with research opportunities for the future, again drawing lessons from history to ensure that we as technologists also focus on addressing the costs of scaling intelligence, from sustainability to safety to security to societal policy.
View the full NSDI '26 program at usenix.org/conference/nsdi26/technical-sessions
