AI Layer Value Accrual & Strategy
The through-line of the week: fragmentation and competition at the model layer is good for everyone who is not at the model layer, and open-source taking share transfers margin from model labs to infrastructure suppliers and application builders. The pacing, open-source shift, and duopoly-risk threads all fed this one thesis from different angles. [1]
Pacing thesis — more compute, lower model margins: Opened the week (Sep 14) arguing that frontier labs 'pacing' the frontier will do so by spending more time and more compute on alignment, monitoring, and evals — meaning pacing is actually bullish for AI infrastructure demand, not bearish [2]. The frontier labs that pace 'likely spend slightly more money on compute at the cost of lower margins' [2]. Reinforced this Sep 16 with quotes from OpenAI CFO Sarah Friar ('so much opportunity to drive growth that I am still highly focused on getting more compute to keep that flywheel going') and VP of Compute Strategy Sachin Katti ('we'll need even more compute to make sure future models are more safe and aligned') [3]. His take: 'If you thought pacing was negative for AI infrastructure demand, think again. Almost as bullish as open-weight AI taking share but not quite.' [3] Added a strategic angle: the alignment spending may also reduce contingent liabilities by showing a 'duty of care' — calling it 'an important and responsible step before going public' [2]. Also noted in passing that OpenAI is discounting [4] — consistent with the margin compression theme.
Open-source share shift — margin transfer to infra and app layers: Quoted @rauchg's data showing open models at 78.4% of token volume on Vercel AI Gateway, with more dollars now spent on open models than OpenAI [5]. His framework: 'Open models taking share shift $ margin from the model layer to the infra and app layers' — positive for the AI infra trade [5]. This thesis drove his biggest exchange of the week — a 20-reply, 21k-view debate with Gary Marcus ([1]). Marcus challenged that OpenAI/Anthropic losing share to open models would hurt infra demand and asked what happens to Coreweave, Oracle, Nvidia 'when generative AI becomes a utility with near-zero margin' [1]. @gavinsbaker pushed back with a steel-and-oil analogy: 'what happened to the consumption of steel and oil after the automobile entered mass production?' — commodities, yes, but 'not all of the inputs into producing tokens are commodities' [1]. The core argument: 'The valuations of Coreweave, Oracle and Nvidia should logically be negatively correlated with margins at the model layer especially at an extreme' — because higher model-layer margins on top of token cost means fewer tokens produced, meaning less infra revenue [1]. Inverted it for Marcus: it would 'obviously be negative for everyone in AI infrastructure' if one lab were a monopsony buyer of compute, and 'terrible for the world if they were a monopoly seller of AI' [1]. Fragmentation at the model layer is good for everyone who is not at the model layer — 'This is all really basic stuff' [1]. Sketched an extreme endpoint: a world where firms like Latham & Watkins run fine-tuned open-source models on their own GPUs and Nvidia funds the open-source training runs — 'it is possible for there to be zero profits at the model layer and yet high returns for AI infrastructure suppliers' [1]. Doubled down when Marcus kept pushing, then disengaged: 'I genuinely do not think you understood what I was trying to say' [6] and apologized for any offense before wishing Marcus well [1].
Duopoly risk — OpenAI/Anthropic convergence would hurt infra and apps: Stated plainly Sep 19 that 'OpenAI and Anthropic trending towards a duopoly would be negative for AI infrastructure demand and a disaster for the application layer' [7]. Reinforced this Sep 16 in the pacing thread: 'A scenario where OpenAI and Anthropic are a duopoly would be terrible for compute demand regardless of whether they accelerated or paced. And open source losing would also be terrible for compute demand' [3]. This is the flip side of the open-source thesis — without fragmentation, the model layer captures too much power and starves everyone else. The duopoly concern surfaced again in the Marcus exchange as a thought experiment: a single lab with 99.999999% margins 'would be a disaster for the neoclouds and everyone who supplied that single lab' [1].
other
Outside of AI analysis, he mused that 'we need to bring back paper' for note-taking, recalling a 2010 meeting where Staples' CFO was pleased people still used paper — a moment he'd found worrying at the time [8]. The rest of this category is one-liners, pleasantries, and disengagements, including politely cutting off a back-and-forth with Gary Marcus [6].
Also this week
Market Strategy & Valuation (~5%): Shared Openrouter data showing extraordinary share gains for OpenAI vs. Anthropic over the prior two months — OpenAI rising from 20% to 50% share while Anthropic fell from 80% to 50% since June [9]. This 4.4k-view, 10-reply post was his second-most-discussed of the window and provided the empirical backdrop for the broader margin-transfer thesis.
Top conversations
20 replies · 21.5k views@gavinsbaker's 20-reply, 21k-view debate with Gary Marcus over whether open models taking share is positive for AI infrastructure. Marcus argued it would hurt infra demand and lead to near-zero margins; @gavinsbaker countered that model-layer margin compression transfers value to infrastructure suppliers, and that a monopsony at the model layer would be far worse for everyone downstream. [1]
10 replies · 4.4k viewsA standalone post sharing Openrouter data showing OpenAI's share rising from 20% to 50% vs. Anthropic's decline from 80% to 50% since June, drawing 10 replies. @gavinsbaker framed it as evidence of extraordinary share gains for OpenAI. [9]
Calls
- ▲ LONGCoreweave, Oracle, NvidiaValuations of Coreweave, Oracle, and Nvidia should be negatively correlated with margins at the model layer, especially at an extreme — higher model-layer margins mean fewer tokens produced and less infra revenue. [1]
