OpenRouter's Head of Insights On Anthropic IPO & AI's Moat Problem | Peter Walker
1 hr 2 min
Peter Walker, Head of Insights at OpenRouter, joins Sam Jacobs, AJ Bruno, and Asad Zaman with a view few people have: 210 trillion tokens a week flowing across 600 AI models, from Anthropic and OpenAI to dozens of open weight inference providers. Topics include Peter's move from Carta to OpenRouter just 10 days before Stripe agreed to acquire it, what a token actually is and why cached tokens now shape AI pricing, and why CFOs struggle to tie fast-rising AI spend to ROI. Plus, whether AI labs can hold on to enterprise customers when switching costs are close to zero, what investors are really betting on with the Anthropic IPO, and whether Peter would buy it on day one. Key Takeaways: - AI labs have almost nothing locking customers in. "I've never seen a part of tech where two things are true. One, there's very, very, very, very, very small switching costs. Almost none," said Peter. The second is the clock: "75% of the tokens on OpenRouter go to models released in the last 60 days. It's planned obsolescence." Even a lab on top knows that no matter how good a model is, "it will be out of date" within months, so it has to keep winning the next release. - Most companies are paying for more intelligence than they need. Peter hears it from CFOs all the time: "my whole company is using Opus and Fable and I know they don't need it." For a task like reading a PDF, "if I find a model that reads the PDF 98% as well as Opus, but costs me literally 10% as much, I'm gonna use that model." Public benchmarks only go so far. OpenRouter's stance: "usage is the benchmark. You have to be doing this on your own company usage to understand what models work for you." - AI spend is heading for a showdown with payroll. Token prices have fallen, but usage has grown much faster, and Peter expects 2027 budgets to feel it: "if your AI usage bill grew 500% this year, thinking that it's gonna grow another 500% is really scary." His view: "at some point, you cannot continually increase AI spen...