We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Chamath: GPU finance through an airline lens
Chamath Palihapitiya called Alex Song's post a "super interesting comparison of GPU finance and the AI buildout to airlines." He noted that Ken Griffin and Sridhar Ramaswamy have made the same comparison, and said both are "very, very smart so worth learning more." Song says he traded ABS (asset-backed securities) on Wall Street and has structured more than $2B of debt financings. His view is that a GPU neocloud "runs on airline economics." He also argues that the debt behind it looks like aircraft engine ABS. His parallels:
- Compute is perishable inventory, like an unsold seat.
- More than 85% of costs are fixed.
- A levered cluster breaks even at roughly 65–75% utilization.
- NVIDIA plays the role of Boeing and Airbus combined.
More detail is in Part 1 of his blog post. This follows yesterday's debate about debt versus equity funding for the AI build-out.
Research picks on efficiency and agents
- Sarah Guo called Cartesia's new H-Net scaling-law work "fundamental." She pointed to a "huge compute efficiency gain," which she says comes partly from learned tokenization. Albert Gu's post reports 3–5x efficiency gains in controlled experiments. It also says that at larger scale the model matches an open-source model trained on 11T tokens while using about 10x less data and compute.
- Eric Schmidt called Maithra Raghu's Samaya Sterling announcement "pretty extraordinary, and on an open model to boot." Sterling is a post-trained 30B mixture-of-experts model with 3B active parameters. Samaya says it scored 58.9% on FrontierFinance, against 55.9% for the best frontier model, at 11x lower cost.
- Sarah Tavel called Alex Tamkin's Anthropic research post "Provocative and rings true." The study was a randomized trial with 50 professional data analysts. It tested whether experts still make the key decisions when they delegate work to an agent.
Martin Casado's notes on VC mindsets
Martin Casado published "The Evolving Landscape of Venture Capital Mindsets." It started as notes for his team, and he shared it "for those who have an interest in venture." The essay describes four mindsets: financial, productive, frontier and speculative. It recommends that a team pick one core philosophy and apply it consistently. The alternative he warns against is pulling arguments from all four as a "grab bag" to justify personal bias. Casado describes himself as "strictly a productive investor" in early-stage infrastructure. This is his own work.
Building agents
Jason Lemkin recommended D. Ternyak's article on building Owner Agent. His takeaway: Owner ranked hundreds of thousands of old support tickets by how much support time each took. It reduced them to about 2,000 operations, which became both its roadmap and its tests. Lemkin's summary: "Your support queue is the spec."
Longtime favorites
- Keith Rabois, asked whether he is bullish on podcasts, praised Ben Thompson's Stratechery instead. He said he has read it "for like a decade plus" and called it "by far the most consistent, original, and useful thinking in tech."
- Tobi Lütke said on Alex Heath's podcast that he recently reread Neal Stephenson's Cryptonomicon after his son read it. He said "literally everything in the book just happened," including cryptocurrencies and fractional ownership. The episode is on YouTube.
- Shaan Puri cited the old book Your Word is Your Wand. He said he loves its idea that words are spells that only work on ourselves, such as swapping "have to" for "get to."