👀

3) Open-weight models are increasingly being post-trained here in America by companies like our portco @FireworksAI_HQ using an enterprise’s private data to create a highly optimized specialized model. This customized model can be combined with a frontier model behind a router to create frontier performance at 40% of the cost to the end customer while consuming the same amount of underlying compute. Cursor’s Composer, Cognition’s SWE-2, Harvey’s Tenet are all early examples of this. Eventually every large enterprise in the world will likely preserve data sovereignty and post-train their own model using their data. These models will be combined behind a router with 1 or more frontier models. This hybrid, multi-model future has been well articulated by Satya Nadella, Alex Karp and others. “Own your weights, control your future.”   It has started to happen: open-weight demand has gone vertical over the last few months in the Vercel and OpenRouter datasets. These specialized open-source models in this hybrid future are just compressing margins at the model layer. And this drives more demand for compute as the compute is effectively being marked up less.   Turning to this weekend’s specific controversy. My comments about no one taking unfunded commitments from OpenAI and Anthropic seriously over the weekend were about the $1 trillion plus of “MOUs” signed with various Asian supply chain and semiconductor companies in 2025. No one takes those seriously. Contractual commitments matter and are taken seriously. Per Anthropic’s leaked S-1, they have committed to spending roughly $11 billion per year with Amazon for the next decade, $15 billion per year with Google for the next 7 years and $5 billion per year with Azure for the next 6 years. These are all priced well below current spot prices for compute *and* current forward prices for compute. The combined $31 billion per year of cloud spending is quite low relative to their reported $70bish ARR given ~80% inference margins and will obviously be internally financed. @altcap doing my best to make you happy and lower the bar. SpaceX has a very different strategy: they sell compute at much higher prices in-line with spot. This may result in an 8-9 month payback which may be close to the financing terms Nvidia has given them. The $161 billion in commitments to Broadcom will require external financing from some combination of Broadcom and private credit, which is beginning to happen. While the headline number of $413 billion of real, actual non-cancellable contracts sounds sensational and scary for Anthropic, it is fairly reasonable when one considers the duration and Anthropic’s scale. We shall have to wait for OpenAI’s S-1 to perform a similar analysis for them.   Most of the actual contractual commitments that have been made by OpenAI and Anthropic to various hyperscalers and neoclouds are now well below the current spot price for compute.  This is why Sarah Friar noted that she could sell all of their contracted compute commitments up 3-5x from their contract value. She was speaking the literal truth. In a world where both OpenAI and Anthropic disappeared, their commitments would be quickly assumed at higher prices by a variety of players.  I think the hyperscalers deeply regret *every* contract they signed in 2024, 2025 and perhaps even the first half of this year. As these contracts roll-off and their compute is repriced to market, their revenue growth should steadily accelerate which will help with the funding concerns for AI infrastructure. Result is that I think we may be approaching the peak of external financing as a % of the AI infrastructure buildout (cue at least one widely anticipated $100b hyperscale equity raise and then can put a fork in it). More to say, but will save it for a later date.

2) A future where OpenAI and Anthropic are a dominant oligopoly at the model layer is negative for almost every company on earth except the two of them. Especially bad for the entire token supply chain. They would be a monopsony for hyperscalers, neoclouds, and datacenter semiconductors. They would have immense buying power and eventually vertically integrate. Would obviously be terrible for the application layer.  And most importantly terrible for humanity. This is a quasi-consensus opinion in Silicon Valley.   This is why fragmentation and competition at the model layer is good for both AI infrastructure and AI applications. The model layer has become dramatically more competitive than it was only 4 months ago with Grok Bot, Muse Wonder and most importantly, open-weight models. I do think people would be shocked by the growth that essentially every open-weight focused inference cloud has experienced since May. This is really good for AI infrastructure demand. Noteworthy that even as Anthropic lost share and OpenAI has likely held share, the combination of the two has continued to grow at a really high rate led by OpenAI over the last 3 months. As Eric Vishria noted, everyone can win. As an aside, Muse is going to show us a lot about what matters for share at the frontier: just model dominance on the Pareto frontier or product quality as measured by harness/model integration where I still think Muse Code is behind Claude Code, Grok Build and Codex.  Although it is looking increasingly likely that the model subsumes the harness in the next 18 months.   Software had effectively zero marginal costs. AI does not. And therefore open-source impacts traditional software and AI very differently. An Opus 5.5, GPT-6.1, Muse Wonder, GLM 5.3, Kimi K3 token all cost roughly the same amount of compute and energy to produce. Effective per token revenue to the semiconductor industry is *roughly* constant for all these models. However, they charge wildly different prices to end users. An Opus or a ChatGPT token might be marked up 300 to 400% from cost (a 75% to 80% inference gross margin) whereas an open-weight token might only be marked up 50-100% (30-50% inference gross margin).  So for the same $1 in true underlying token cost, an Opus or ChatGPT token generates $4 to $5 in revenue at the model layer where an open-weight token might only generate $2 in revenue at the model layer.  And open-weight models have less buying power than Anthropic and OpenAI so they should generally generate more revenue per token to the clouds and neoclouds than Anthropic and OpenAI.  This is why open-weight models having significant share is the best case scenario for AI infrastructure demand. Robust competition between 4-5 frontier labs that compresses margins is the next best scenario. And ironically, it seems like these cheaper open-weight tokens may be net positive for demand at the absolute frontier of intelligence as we move to agent swarms. Meaning something like “Fable 5.1 extra-high” is worth paying for in a use case where it is coordinating a swarm of 100 open-weight agents.   All of this is why compute prices have gone vertical as the model layer became increasingly competitive. Lower margins at the model layer mechanically, even if counter-intuitively, drive more demand for compute. Perhaps the more important dynamic is that as intelligence becomes more economically useful, the value of compute should rise. This was Dwarkesh’s point about an H100 eventually renting out for the annualized value of a software engineer. Seems improbable but I continue to be surprised by the inexorable rise in compute prices. Inference clouds focused on open-weights are allegedly bidding close to ~$100 per watt for Blackwell clusters, which is a massive premium to the $50 per watt Google/SpaceX deal that was widely thought to be at least a local, unsustainable peak in compute pricing. Prices have gone up since that deal per 3P data.   Continued below.

Martin Shkreli says he kind of hates Dario, calling him a fear-mongerer who saw Pandora's box at OpenAI and left to build his own. "I want to pay deference to him creating a trillion dollar company, but I kind of hate the guy. I think he's a fear-mongerer." "Imagine you and I started a business and I said, I'm going to show you Pandora's box. Here it is, and you say, 'wow, that's amazing.'" "And he said, 'thanks a lot, Martin. I'm going down the street. I'm going to start Pandora's box 2.'" "I would probably sue you for trade secrets or something. This is exactly what happened."

MARTIN SHKRELI ON AI DOOMERS: “You can turn off the computer. It's not that big of a deal.” “I don't think there'll be a single AI-relevant drug in the next 10 years. AI does not help drug discovery.”