Here’s a number that should make you pause: 100 million words. That’s roughly 200 full-length novels, the entire Harry Potter series about 13 times over, or approximately every earnings call transcript from the S&P 500 for the past decade. And according to Anthropic CEO Dario Amodei, AI models can technically process all of it in a single context window right now.
The bottleneck isn’t capability. It’s speed. Inference, the process of actually generating responses from that mountain of text, remains the practical constraint keeping this from becoming a shipping product. But the underlying architecture? Amodei says it’s already there.
What Amodei actually said, and why it matters
In an interview from July 2025, Amodei made a claim that has continued to circulate across AI communities: a 100-million-word context window is technically feasible with current model architectures. The kicker is what happens inside that window. Models learn during inference without changing their weights.
This concept, known as in-context learning, isn’t new. Researchers have been exploring it since the early days of large language models. What is new is the scale Amodei is describing. The AI industry has moved from context windows of a few thousand tokens in 2020 to hundreds of thousands of tokens by 2023. A 100-million-word window would represent a leap of several orders of magnitude beyond what’s commercially available today.
