Google’s Internal Memo: “We Have No Moat, Neither Does OpenAI” – The Rise of Open‑Source AI

A Google internal memo, titled “We have no moat, OpenAI also has none,” was leaked and confirmed authentic by SemiAnalysis. The document argues that both Google and OpenAI cannot win the AI arms race because open‑source models are overtaking them.

The memo highlights several open‑source breakthroughs: LLaMA’s public leak, Vicuna achieving >90% of ChatGPT/Bard quality with a 13B model, Alpaca adding instruction tuning, and LoRA enabling cheap, fast fine‑tuning on consumer hardware.

A timeline from February to April 2023 shows a cascade of innovations—quantized LLaMA on CPUs, one‑hour multimodal training, RLHF‑level open‑source models, and community‑built ecosystems like GPT‑4All—demonstrating that high‑quality models can be built for under $300.

The analysis stresses that data quality, not sheer scale, drives progress, and that low‑rank adaptation (LoRA) dramatically reduces fine‑tuning costs, allowing rapid iteration on small models that can match large‑scale systems.

In response to this pressure, Google’s AI chief Jeff Dean announced a policy shift to delay sharing internal research, signaling a defensive stance aimed at protecting Google’s core search business and future AI products.

The memo concludes that open‑source is the true winner of the AI race; Google must collaborate with the community and reconsider its closed‑door approach, even if it means releasing smaller model weights.

References: SemiAnalysis article Washington Post report