Tag
Briefings and ideas about GPUs from AI podcast discussions and the research they cite, with links to the original sources.
Erik Bernhardsson, Modal's CEO, says the company keeps renting GPU capacity on three-year deals. He worries less about rising prices than about a collapse that would leave buyers having paid too much just as resale markets dry up.
Modal CEO Erik Bernhardsson says GPUs are mostly sold as fixed multiyear blocks, while real demand rises and falls. After a $750 million raise, he argues that pooling thousands of customers lets Modal resell that capacity in flexible slices.
Investor Chamath Palihapitiya argued that AGI exists but can't spread until power and data centers are built, and predicted a compute shortage within three or four years. Alex Wissner-Gross countered that demand and smaller models matter too.
Evan Conrad, CEO of SF Compute, says delivering traded GPU time made his firm do a cloud's work, yet it owns no chips because that would let it manipulate prices. He argues NVIDIA and big clouds couldn't run such a market neutrally.
SF Compute CEO Evan Conrad says only about 5% of his company's capacity is on its order book, but customers recoup $25 per $100 on average by reselling. In his account, interruption terms, lead time and expected demand shape GPU prices.
SF Compute CEO Evan Conrad argues that GPU time is too varied to trade as one uniform commodity. He says cash-settled contracts and price indexes should be built on markets that actually deliver compute, or they risk manipulation and poor hedges.
AI safety researcher Jeffrey Ladish proposed that the government ask AI labs to shift computing power from training new models to serving customers. He argued that constituents' calls to Congress can create the pressure to act.
Positron CTO Thomas Sohmers said AI agents are multiplying demand for memory. He claimed Positron's first-generation product sustained 93% of theoretical memory bandwidth, against an average of 30–40% for NVIDIA GPUs when decoding with transformer models. Its next chip, Asimov, is planned for 2027.
On Moonshots, Emad Mostaque argued governments should buy AI compute for universities and open science. Alexander Wissner-Gross said private labs and markets should decide. Dave Blundin described being outbid for a 72-GPU system.
Yash Patil, whose company Applied Compute sells post-training, says teams should first improve prompts, context and tools, then retrain weights for their biggest workloads, tasks that change over time and judgment-heavy tasks.
Yash Patil, CEO of Applied Compute, argues that the biggest AI workloads are the most valuable both to custom-train and to serve, which he says gives a training provider an edge in inference. He calls serving the easier job.
Alex Wissner-Gross said Sonnet 5.5's 70% Terminal-Bench score is no better value per attempt than Opus 5.5; Anthropic's charts put the savings at lower effort settings. He said Gemini 4 Argon stands out mainly on avoiding hallucinations.