What is the compute challenge of building a proactive ambient agent compared to a code generation or chat product?

With a coding product, the user prompts it, it runs something, comes back, and then you exchange more messages. There is a lot of background work but it is fundamentally interactive and triggered. Instinct is different because it can wake up and sleep at any moment throughout the day based on its own judgment. It might wake up at 6 AM because it knows you wake at 7 AM and scan everything to prepare for your day. It might go dormant at 10 AM and wake back up at 4 PM because something requires attention. That kind of proactive ambient behavior generates orders of magnitude more token flow than any current AI product. Then layer on 10 percent day over day growth and the fact that compute has a lead time of several months. If you buy 10 times your current compute today, you will consume that in a couple of weeks. If you buy ahead too aggressively and growth slows even slightly, you are wrong by 3 or 4 times on your capital allocation. That is the problem I spend about 40 percent of my time on.