Right now most compute is centralized compute. But Agentic AI is going to push more and more enterprise inference to the edge. $NBIS x $PLTR’s asset-light model, powered by modular deployments, is a bellwether for this shift: putting the right amount of compute closer to the customer. Enterprises are increasingly developing their own models and serving them. This is at the heart of Arkady and Karp's partnership, and we are going to see this accelerate in 2027. Revolut is doing this already on @nebiustf. And this past week, $ZETA announced their in-house Athena model built on Fireworks for bizint. These models don't just exist to answer one-off prompts. They are increasingly wrapped within agentic workflows and exposed through APIs into customer-facing products and internal enterprise systems. Agentic workflows are more latency sensitive, especially when they’re sequential. Long-running agents can chain dozens or even hundreds of inference or tool calls. Even staying within the same country can add ~10–30 ms of round-trip network latency per call. Across a 100-step sequential agent workflow, that alone can add ~1–3 seconds of pure network delay. This is before model inference, tool execution, database queries etc. This is where modular deployments start to make a lot of sense. You don’t need every enterprise workload sitting inside a 500MW AI factory. For the latency-sensitive tail of enterprise inference, you want smaller pools of compute deployed closer to where the data and workflows actually live. This part is more speculative on my part... I expect $NBIS × $PLTR to be the beginning of a specialized-model revolution. Enterprises will develop and serve models tailored to specific workflows, industries, and datasets. In this case on @nebiustf. They will then "airdrop" them modularly into in sovereign or customer-adjacent nodes, closer to where the data and decisions actually live.