There is a difference I find fundamental, and one that many companies have not yet internalized... the difference between a chatbot and an agent. A chatbot is simple... you ask a question, it answers, you pay one token in and one token out, it is manageable. An agent is completely different... it chains dozens, hundreds of requests to accomplish a task, whether that is coding an application, writing a report from scattered notes, or analyzing a dataset and producing recommendations. And every step, every intermediate request, is a billed API call.

And well, it turns out Google now processes 3.2 million billion tokens per month... that is seven times more than a year ago, and three hundred and thirty times more than two years ago. At OpenAI, daily interactions with Codex have been multiplied by 15 since January 2026. 179 customer companies have exceeded 1,000 billion tokens generated in total, and we are still only at the first deployments of agents in the enterprise.

×100
Request multiplier per agent
An agent chains 10, 50, 100 requests for a single task. Every step is billed in the cloud. In local, the marginal cost is zero.

So now, do the math with me... on a cloud, every step of the agent is a billed request. If you have an agent that does 100 steps to accomplish a task, that is 100 times the cost of a chatbot for the same task. And that means invoices are going to explode, because agents are exactly what everyone wants to deploy now... it is the product of the year 2026, and it is also the product that will kill budgets.

The zero marginal cost: the local advantage

And there, something rather remarkable happens with local AI. On a local server, an agent that does 100 steps costs nothing more than electricity. The model runs on your machine, every iteration is local computation, there is no API, no meter running, no invoice arriving at the end of the month... the marginal cost is strictly zero.

€0
Marginal cost of a local agent
100 steps on your server = 0 token billed. No API, no meter. The more the agent iterates, the more the gap widens with the cloud.

The more the agent iterates, the more the gap widens between cloud and local. A cloud agent at 0.01 euro per token, 100 steps, that is 1 euro per task... the same agent on your server, that is 0. And you multiply that by the number of agents, by the number of employees, by the number of tasks per day. It is mathematical, and that is why I say agents make local AI economically obvious.

LeCun: current agents do not plan, they imitate

Yann LeCun, in his interview with Le Figaro in June 2026, said something about agents that is worth pausing on. He says that the agentic systems offered today by Google or OpenAI do not have the ability to predict the consequences of their actions... and well, they are trained to imitate humans, and imitation, according to LeCun, imitation remains a reproduction of observed behaviors, and really not intelligence in the sense LeCun means it.

And that ties into something I notice in the field... the agents we deploy today are heavy consumers of tokens because they proceed by trial and error, by successive iterations, a bit like someone groping in the dark. And every fumble, in the cloud, is billed. In local, you fumble for free, if I can put it that way... the model tests, gets it wrong, starts over, adjusts, all of that on your machine, without it costing a single cent more.

Cloud for the prototype, local for production

Now I am not saying you should switch everything to local overnight, that would be absurd. The cloud has its place, especially for prototyping, for proof of concept, for tasks that require the latest frontier of the model. And well, local has its place for large-scale agent production, for recurring, internal uses that will consume thousands of tokens per day and per employee.

If you are thinking about deploying AI agents in your company, and well, maybe now is the time to do the math between cloud and local before signing... and I can help you do it.