| url | https://www.semianalysis.com/p/google-we-have-no-moat-and-neither |
|---|---|
| raw | raw/highlights-google-no-moat.json |
A leaked internal Google document from a researcher who’d been watching the open-source LLM ecosystem closely. The headline claim was, in retrospect, both correct and completely wrong — depending on what you think a moat actually is.
The argument
The memo argued that open-source LLMs were closing the gap with proprietary models so fast that proprietary AI advantages were structurally temporary. LoRA-style fine-tuning lets anyone personalize a model on consumer hardware in hours. The “Stable Diffusion moment” for LLMs was already underway. Therefore: neither Google nor OpenAI had a defensible position; the value would accrue to whoever built the best ecosystem.
Why it was right
The capability gap did close, faster than almost anyone expected. By 2024 you could run a Llama 3 derivative on a laptop and get answers that were within shouting distance of the best frontier model. The premise — that model quality alone is not a moat — turned out to be exactly correct.
Why it was wrong
The memo conflated model moats with product moats. ChatGPT’s actual defensibility was never the model — it was the brand, the user habit, the verb test, and the years of accumulated conversation history that competitors literally cannot replicate (chatgpt-pmf). Open-source eroded the technology barrier and barely dented the distribution barrier. Two years after the memo, ChatGPT was still the verb, OpenAI was still the brand, and the open-source ecosystem was busy building chatbot wrappers that nobody used.
The lesson is the most important one in modern tech strategy: when capability commoditizes, distribution wins (distribution). And distribution is exactly the kind of moat the memo was implicitly dismissing.
Connections
- seven-powers — the memo only considered Cornered Resource (the model) as a moat. It missed Branding and Scale Economies entirely.
- chatgpt-pmf — the case study showing how OpenAI built a real moat on top of an erodable one.
- cheating-is-all-you-need — Yegge’s argument for data moats as the post-model defensibility.
- software-dev-costs-moats — same dynamic, broader: when the core technology is cheap, the surrounding ecosystem becomes the battleground.
- network-effects — OpenAI doesn’t have classic network effects. It has relationship effects instead, which the memo didn’t consider.