Austin Lyons

Had a nice dinner in Maui chatting with @clattner_llvm and @samcharrington. Chris has a huge vision for Mojo and Max and Qualcomm. Quite bullish.

Agree — gigawatt deployments mean co-optimizing EVERYTHING for the workload at hand https://t.co/nMv80sSMQk

World’s most important assets are more valuable than traditional financial model depreciation expects. Ampere and Hopper would sit close to a 6 year depreciation curve (but a poor fit), and Blackwell is just off the charts! https://t.co/uYU21Y9UoC

I joined David Goldman on @CelestaCapital's TechSurge podcast to talk about the race to build the next trillion-dollar AI chip company. The main idea we dig into is that no one has yet brought a chip to market designed specifically for LLMs. Blackwell is a GPU that morphed toward the workload. Nobody has yet stood up a gigawatt of clean-sheet LLM silicon, and that gap is where the next trillion-dollar chip company comes from. We cover: • Why AI infrastructure now sells as full systems, and how Nvidia got to rack scale first • Prefill vs. decode, and why the SRAM bets Groq and Cerebras made years ago suddenly paid off • How neoclouds built $125B+ of public-market value while most investors missed them • Circular financing, and why I land on the demand being real • My four conditions for the next trillion-dollar chip company • Clean-sheet architectures: @OpenAI's Jalapeno, @TensordyneInc's log math, @Etched's low-voltage inference • Why enterprise and on-prem inference could end up bigger than the cloud Full episode: https://t.co/3J52KIygMo

I joined David Goldman on @CelestaCapital's TechSurge podcast to talk about the race to build the next trillion-dollar AI chip company. The main idea we dig into is that no one has yet brought a chip to market designed specifically for LLMs. Blackwell is a GPU that morphed toward the workload. Nobody has yet stood up a gigawatt of clean-sheet LLM silicon, and that gap is where the next trillion-dollar chip company comes from. We cover: • Why AI infrastructure now sells as full systems, and how Nvidia got to rack scale first • Prefill vs. decode, and why the SRAM bets Groq and Cerebras made years ago suddenly paid off • How neoclouds built $125B+ of public-market value while most investors missed them • Circular financing, and why I land on the demand being real • My four conditions for the next trillion-dollar chip company • Clean-sheet architectures: @OpenAI's Jalapeno, @TensordyneInc's log math, @Etched's low-voltage inference • Why enterprise and on-prem inference could end up bigger than the cloud Full episode: https://t.co/3J52KIygMo

btw what WILL be compute-negative for training? AI datacenter buildout backlash. That's what should really concern the semis industry... the pacing will come from the fears of our neighbors, not from the fears of SF researchers https://t.co/r2YdwxI1lo

More recently $AMZN warrants from $QCOM follow the same logic. QCOM is new to datacenter. Warrants can help. Can also be thought of as a discount. https://t.co/unCdWRilrU

High-NA halves the exposure field, cutting the maximum die size per shot in half. For now, stitch two half fields into one die on today's 6" masks. Long term, 6x12 inch masks are the answer to recover a full die with no seam. Of course that requires ecosystem alignment, which $ASML and $INTC are spearheading.

Good point from @vikramskr. Yes, NVHBM opens up partnership opportunities for Nvidia. But more simply, for Nvidia’s own use case, owning the base die is in its best interest: “So far, NVIDIA has had to live with the base die design choices made by memory companies, some of which are highly questionable as we discussed earlier. NVHBM is NVIDIA’s way of taking destiny into their own hands.”

South Korea trying to be an AI diffusion leader 👀 WSJ: “South Korea plans to give its entire population free access to generative AI services, the first major state-led offering treating the technology akin to a public utility.” “The AI tools will be linked directly to government systems and allow South Koreans to book doctor’s appointments, receive help with apartment hunting or get tax advice. Small businesses will be able to tabulate their taxes and check their eligibility for government-support programs, and parents will receive suggested educational content for their children.” “Around a quarter of South Koreans pay for generative AI, according to a government survey earlier this year. In contrast, about 2% of Americans pay, according to a PNC Bank survey this year.” @jukan05 @jjding99