By late 2025, multiple companies focused on “scaffolding” around large language models—rather than on training base models themselves—have in fact become very large, commercially successful businesses, matching Chamath’s thesis.

Infrastructure / tooling layer

• Scale AI provides data-labeling, RLHF, and deployment tooling used by major foundation model labs. Meta’s 2025 deal valued Scale at about $29B, with expectations of $2B in 2025 revenue, making it a large and clearly successful infrastructure business built around (not as) LLMs. (ft.com)
• Pinecone, a vector database that serves as long‑term memory for LLM apps, raised a $100M Series B at a $750M valuation and is described as a “critical component” of the generative‑AI stack. (pinecone.io)
• Weights & Biases, a widely used AI developer platform for training and monitoring models, was acquired by Nvidia‑backed cloud provider CoreWeave in a deal reported around $1.7B, to deepen CoreWeave’s AI developer tooling ahead of an IPO targeted at $35B+ valuation. (reuters.com)
• LangChain, an open‑source framework for building AI agents and applications, raised $125M at a $1.25B valuation in 2025, reflecting substantial commercial value from pure “agentic” scaffolding software. (techcrunch.com)

Application / developer-product layer built on top of models

• Anysphere (Cursor), an AI-assisted IDE built on top of foundation models rather than training its own, reached about $500M in ARR and a $9.9B valuation by mid‑2025. (en.wikipedia.org)
• Cognition AI (Devin), an autonomous coding agent relying on underlying model infrastructure, hit a $10.2B valuation with ARR growing from $1M (Sep 2024) to $73M (Jun 2025). (techcrunch.com)
• Lovable, an AI software‑engineering platform that builds apps from natural‑language prompts using existing models, surpassed $100M ARR just eight months after launch and raised $200M at a $1.8B valuation in July 2025. (en.wikipedia.org)
• Perplexity AI, an LLM‑powered search and browser company, relies on models from others but provides the surrounding product and UX. By 2025 it was valued around $18–20B, with ARR approaching $200M, and signed a $400M search integration deal with Snap, underscoring the scale possible at the app layer. (cnbc.com)

Infrastructure around AI compute

• Crusoe, which builds AI‑specific data centers (compute “plumbing” for model and app providers), raised $1.4B at a valuation above $10B and is constructing multi‑gigawatt campuses for OpenAI and others—another example of non‑model infrastructure becoming a very large business. (ft.com)

Across these examples, companies that supply the non‑model scaffolding—vector databases, data/ML tooling, agent frameworks, AI‑enhanced IDEs, and AI data‑center infrastructure—have already achieved multi‑billion‑dollar valuations and, in many cases, hundreds of millions to billions in annual revenue. That is exactly the kind of “really good business” Chamath described. While the long‑term landscape can still change, the evidence by November 2025 shows that his prediction has, in substance, come true.