The State of Edge AI is a research report I wrote with three co-authors (M. Sheldon, S. Kim, and S. Shrivastava), published with Peri Labs in October 2024. Its argument was unfashionable at the time: the cloud-only model of AI, where every token round-trips to a hyperscaler data centre, runs into three walls at consumer scale, and the industry would be forced to move meaningful inference onto devices and edge infrastructure.
The three walls, as the report laid them out:
- Latency. Interactive AI needs round trips the speed of light does not grant to distant data centres. Anything conversational, wearable, or in-the-loop with the physical world wants local compute.
- Privacy. The most valuable context (messages, health, screen content) is exactly what users and regulators least want shipped to third-party clouds. Local inference dissolves the dilemma.
- Bandwidth and cost. Streaming raw context up and tokens down for billions of users is an economic and network non-starter; the marginal cloud token has a price, the marginal local token effectively does not.
What happened next
The report's core prediction was made on 14 October 2024. On 9 June 2025, Apple opened its on-device foundation model to developers, turning local, private AI from a demo into a platform feature with an API surface. The judgement that mattered was not that edge hardware would improve, which everyone believed, but that a first-party platform owner would make on-device intelligence the default developer substrate. That is now the shipped reality on more than a billion devices.
Read the report
Download The State of Edge AI (PDF) ↗ The report is indexed on Google Scholar, where its citations are tracked.
How it connects to the rest of my research
Edge AI is one leg of a broader thesis about where the economics of intelligence settle: computation migrates toward users, capital markets learn to finance the infrastructure, and agents transact on their own rails. The financing leg is The AiFi Thesis. The market-integrity leg, on what happens when trading venues meet manufactured activity, is The Economics of Wash Trading and the wider research hub.