Baker vs Marcus: Do Fat Model Margins Help or Hurt Neoclouds?

On Sept 21, investor Gavin Baker and Gary Marcus engaged in a multi-round exchange on X over how the LLM profit landscape affects neoclouds. The core disagreement: does fat model-layer profit help or hurt upstream compute suppliers? The two also cleared up a misreading over an "AI single-buyer monopoly" framing, with Baker apologizing for putting words in Marcus's mouth.

Confirmed

• Gary Marcus argued that the more profitable LLM providers are overall, the better neoclouds do; if LLM businesses earn poor returns on capital, neoclouds inevitably suffer. He drew an analogy to the decline of the US steel industry, noting high-cost suppliers in steel and oil eventually exited.
• Gavin Baker rejected the analogy that steel and oil consumption rose after cars proliferated, arguing the better question is what happened to steel and oil consumption after mass car production. He stressed that steel and oil are commodities, but the inputs to producing tokens are not all commodities.
• Baker's core thesis: the higher the model-layer gross margin, the more CoreWeave, Oracle, and Nvidia get hurt. He gave the example that if a single lab captured 99.999999% of profits, neoclouds would not be better off.
• Marcus clarified he never defended monopsony, and that his core argument since August 2023 has been that OpenAI will not become a monopolist because labs are all essentially using the same recipe.
• Baker apologized for the earlier "AI single buyer/monopoly" remark, saying it was only a thought experiment and not meant to speak for Marcus.

Why it matters

• The debate goes to the heart of value distribution across the AI stack: even in an AI boom, how profits split between the model layer and the compute layer determines how much CoreWeave, Oracle, and Nvidia actually benefit—rejecting a simple rising-tide logic.
• The opposing steel-and-oil analogies offer two rival frameworks for judging whether AI compute becomes commoditized and whether neoclouds can sustain high returns, directly relevant to AI infrastructure investor expectations.

2026-09-21 ~ 2026-09-21 · 5 related posts

• [source] Why model-layer margins should be negatively correlated with CoreWeave, Oracle and Nvidia — GavinSBaker · 2026-09-21
• [source] Gary Marcus invokes steel-mill decline to argue neoclouds hinge on LLM profitability — GaryMarcus · 2026-09-21
• Gavin Baker Clashes with Gary Marcus: Model-Layer Monopoly Would Crush Neoclouds — GavinSBaker · 2026-09-21
• [source] Gary Marcus clashes with Gavin Baker: OpenAI won't be a monopoly because everyone shares the same recipe — GaryMarcus · 2026-09-21

1 near-duplicate retellings: GavinSBaker