Where will Intelligence be allowed to live?

AI’s Second Act, Part Two: the politics of placement

In my last article, I argued that AI’s second act is a distribution problem: the frontier is no longer who trains the most capable model, but who gets intelligence to where decisions, workflows, and users actually are. If training is the power plant, inference is the grid.

But the grid analogy has a flaw worth sitting with. The electricity grid expanded because everyone in the value chain benefited from broader distribution - generators, utilities, appliance makers, and households all won as power reached further. AI’s grid may not work that way. Here, distributing intelligence creates winners and losers. And the losers are some of the most powerful incumbents in the industry.

That changes the question. The interesting debate ahead is not ‘can’ we push intelligence closer to the user. Technically, we already can. The debate is whether we will be ‘allowed’ to.

Where the Analogy Breaks

Electricity became transformative when it could be delivered safely and locally, on each user’s own terms. AI has the same potential arc. But electricity never threatened the power plant’s business model. Distributed AI does - and two forces are already pushing back.

Centralized compute economics. As inference begins to overtake training as the dominant workload, much of the infrastructure originally built for training will be repurposed to serve it. That repurposing hides an architectural mismatch. Training infrastructure is optimized for raw compute density (FLOPS). Localized, distributed inference is governed by something almost entirely different: memory bandwidth and the physics of moving data. If inference migrates toward enterprise environments, edge devices, and sovereign infrastructure, the centralized high-FLOPS bet becomes harder to justify both financially and structurally. The incumbents who placed that bet have every incentive to keep the workload where their hardware already sits.

Behavioral data. For ad-driven platforms, the location of AI is not an engineering question; rather, it is a revenue question. An assistant that personalizes on-device never sends behavioral data back to the platform. That is excellent for the user and a direct threat to any business whose model depends on understanding user intent at scale. The most valuable byproduct of centralized inference is not the answer returned to the user; it is the data exhaust the platform keeps.

Both forces create the same structural incentive: a pull toward keeping intelligence centralized - even as the technical and customer case for distributing it grows stronger.

The Technical Case Is Already Settled

From almost every user and enterprise perspective, local or edge inference wins on the merits:

• Faster: no round-trip to a distant server
• More private: sensitive data never leaves the trust boundary
• More resilient: it keeps working when connectivity does not
• More controllable: enterprises and governments can enforce their own policies directly

These are not abstract preferences. A bank cannot route fraud detection through a shared external cloud. A hospital cannot expose patient workflows to a third-party provider. A defense agency cannot depend on foreign AI infrastructure. In these environments, distributed AI is already the obvious direction. The real concern is access: whether they can deploy it on terms they control.

The technical and institutional logic points one direction: intelligence should live where decisions are made.

The Business Logic Points the Other Way

Centralized AI also delivers real value that’s easy to undervalue in this debate: managed reliability, rapid iteration, security at scale, and a single place to enforce policy. But it concentrates leverage too - metered usage, interaction data, model-access management, and durable customer relationships. Centralization tilts intelligence toward something you subscribe to rather than something you fully own.

“Where is it most efficient to run inference?” is only the surface question. Beneath it sits a more political one: who benefits when intelligence becomes local, and who benefits when it stays centralized?

The answer maps cleanly onto the players. Users and enterprises benefit from local AI through privacy, speed, and autonomy. Governments benefit through sovereignty. Cloud providers, ad platforms, and model-API vendors benefit from centralization. The next decade of AI infrastructure will be shaped as much by that tension as by any single technical breakthrough.

Utility or Platform?

Electricity became a utility - distributed, standardized, invisible. You do not think about voltage when you turn on the air conditioning. You think about whether the room is comfortable.

AI could follow that path. Or it could become a platform - a controlled interface where a handful of companies own the choke points: the assistant, the API, the data pipeline, and the behavioral layer underneath. Utilities compete on reliability and price. Platforms compete on lock-in. The difference determines who captures the value AI creates.

The most likely outcome is hybrid. Frontier models stay centralized where scale genuinely justifies it. More inference moves local as privacy, latency, cost, and sovereignty demand it. The real infrastructure battle is over where that boundary settles - and that boundary will be drawn by incentives at least as much as by engineering.

What Builders Should Watch

Inference infrastructure is not just a technical market. It is a political-economy market. The companies that win will not merely make inference cheaper; they will navigate the tension between what users need and what existing business models can economically support.

The clearest openings are in the layers that make distributed intelligence governable:

• Hybrid routing - deciding intelligently, per request, when to use a local, private, or cloud model
• Enterprise control planes - governance, visibility, and cost management across distributed deployments
• Sovereign infrastructure - AI that runs inside national or organizational boundaries, including disconnected environments
• Privacy-preserving personalization - adapting to a user without centralizing their behavioral data

These look like plumbing problems. But infrastructure markets are built on plumbing - and underneath, every one of these is a data-movement problem. The builders who win the next decade will not just ship smart software. They will design the memory architectures and interconnect fabrics that let context flow across hybrid boundaries without collapsing operating margins.

The Uncomfortable Question

AI’s first act asked: how much intelligence can we train?

Its second act asked: how efficiently can we distribute it?

The next question is harder, because it is not technical at all: where will intelligence be allowed to live?

Technically, it can move closer to users. Economically, incumbents have reasons to resist. Politically, users, enterprises, and governments are starting to demand it anyway. The decisive battles ahead are not about whether local AI is possible - that is settled. They are about whether it is permitted to become the default.

And that war does not start in a boardroom or a regulator’s office. It starts somewhere far less visible - on the silicon die itself, where the industry is quietly flipping from the FLOPS race to the data-movement race. That is where we go next.