The dominant story of the AI buildout is concentration: hundreds of billions of dollars poured into a handful of vast, power-hungry data centers. This week two very different companies sketched the opposite idea, that a lot of AI work will happen not in one giant campus but scattered across thousands of small ones. It is still a minority bet, but it is starting to look like a deliberate one.
Tesla supplied the most eye-catching version. On 18 June it filed a US trademark for "MEGAPOD," described as modular data center hardware for artificial intelligence computing. The filing builds on Elon Musk's earlier claim that Tesla could run AI workloads on the spare power at its Supercharger network, which he pegged at roughly seven gigawatts of available capacity, and on idle hardware in parked cars. The pitch is that the expensive parts of a data center, grid hookups, land, cooling, are things Tesla's charging network already owns. Treat this as ambition rather than product: it rests on a trademark and a founder's statements, not a deployment, and the orchestration, security and grid questions are real.
The less glamorous version may be the more telling one. Amit Shah, chief executive of the startup InstaLILY, argues that the next divide in enterprise software will be between companies that rent intelligence from hyperscalers and those that own it. His "Small Data Center" runs a full reasoning and governance stack close to where work happens, on factory floors and in warehouses, rather than in a browser tab. He claims it cut a logistics customer's routing times from 15 minutes to three. His reasoning is plain: industrial operations run under tight latency, patchy connectivity and relentless cost pressure, and manufacturers will not hand critical decisions to a system they cannot audit or govern.
The two stories share a thesis worth taking seriously. Earlier waves of distributed computing moved files or transactions around a network. This one moves intelligence, and Shah's sharper point is that intelligence compounds at the edge: every exception and workflow feeds a private model that gets more capable over time. BitTorrent never got smarter the more people used it.
There is a reason the hyperscalers are not leading this charge. Their economics reward centralized consumption, and distributed inference quietly undercuts a roadmap built on ever-larger central training runs. They are not ignoring edge compute so much as moving carefully, because cannibalizing their core business is uncomfortable. That leaves the opening to outsiders: a carmaker with a spare power network and a startup selling to factories. Neither may win. But when the most-discussed constraint on AI is electricity and the most-discussed solution is yet another multi-gigawatt campus, it is worth noticing who is betting the answer points the other way.