TechnicalNews

Jensen Huang hypes Lex to everyone

Lex Clips

Jensen Huang demonstrates Nvidia's latest Grace Blackwell server technology to Lex Friedman at Computex, highlighting improvements in liquid cooling, assembly time, and CPU performance. Huang explains how the business model has shifted from selling CPU cores to generating and monetizing AI tokens.

Summary

Jensen Huang enthusiastically shows off Nvidia's Grace Blackwell server to Lex Friedman during what appears to be a Computex event. The key hardware improvements include achieving 100% liquid cooling (up from 80% in the last generation) and reducing assembly time from 2 hours to just 5 minutes. Huang emphasizes that ease of construction correlates with reliability. The Grace Blackwell CPU delivers 1.8x performance improvement over x86 processors. Huang explains a fundamental shift in the semiconductor economics: previous hyperscale CPUs were designed around a rental model where value was measured in dollars-per-core, incentivizing maximum core counts. In contrast, AI workloads don't rent cores but instead generate tokens that are monetized. This changes the optimization target entirely—the goal is now to complete the inference loop as quickly as possible and generate tokens at maximum speed. Huang humorously notes that this business model is straightforward: the more customers buy, the more profit they generate. The conversation includes lighthearted moments about Lex's presence at the event, discussion of a lengthy podcast Lex hosts, and casual remarks about food and cake at the conference.

Key Insights

  • Jensen Huang argues that the shift from x86 CPU economics to AI workloads represents a fundamental business model change: hyperscale CPUs were optimized for rental with value measured in dollars-per-core, but AI economics optimize for token generation speed rather than core count.
  • Huang claims that in the new AI paradigm, the core objective is to traverse the inference loop as quickly as possible and generate tokens at maximum speed, making speed-to-token-generation the primary performance metric rather than raw core availability.
  • Grace Blackwell achieves 100% liquid cooling versus the previous 80% in the last generation Grace processor, and reduces assembly time from 2 hours to 5 minutes.
  • Huang states that things which are easy to build are also reliable, connecting ease of manufacturing with system reliability.
  • The Grace Blackwell CPU delivers 1.8x performance improvement over x86 hyperscale CPUs, addressing the new token-generation-focused workload requirements.

Topics

Grace Blackwell server architectureLiquid cooling technology improvementsCPU performance gains over x86AI token generation vs. CPU core rental economicsHardware assembly and reliability

Transcript

[0:02] Good to meet you. >> Look at this guy. He went He's just traveling through Asia. >> Does everybody know the famous Lex Friedman? >> IT. STOP IT. >> DAD, say I love you. Look at this. Look at this. No cables, no hoses. 100% liquid cooled. >> Yes. >> Last generation Grace Blackwall was only 80% liquid cooled. Now it's 100% cooled. From 2 hours to assemble, it's now 5 [0:33] minutes. Things that are easy to build are also reliable. >> Officially in full production, right? >> Yeah, full production now. Incredible, right? >> That's incredible. >> All this, everything. >> And the various CPU, you're full in on the CPU. I saw the 1.8x over x86.…

Full transcript available for MurmurCast members

Sign Up to Access

More from Lex Clips

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.