NORBERT BIEDRZYCKI

We may have fundamentally misunderstood the economics of AI, because it won't be free

AI won't be free

For the past two years, we’ve heard almost exclusively one narrative: AI will make knowledge work cheaper, faster, and practically unlimited in cost.

But early signals are starting to emerge that reality might look very different.

According to recent reports, Microsoft began restricting internal access to Anthropic’s Claude Code because per-token costs started spiraling out of control. In some cases, a single engineer’s monthly AI usage was reportedly reaching $1,500–$2,000. Similar stories are beginning to surface across other large tech organizations. Companies are implementing AI spending caps because budgets can disappear far faster than anyone anticipated.

And I think this is an incredibly important moment.

For a long time, many people assumed AI would become something like the internet: cheap, ubiquitous, and practically free at the marginal level. The problem is that AI models don’t work like ordinary software. Every query generates real infrastructure costs: GPUs, energy, inference, memory, data transfer, and increasingly complex operations being executed by the models.

What’s striking is that even as the cost per token falls, total consumption is growing even faster.

Models are getting larger. Context windows are expanding to hundreds of thousands of tokens. AI agents are executing increasingly complex tasks. Companies are moving from occasional AI use to building entire operational processes around it.

And that’s exactly where a fundamental economic problem might emerge.

Maybe AI won’t become the “nearly free commodity” so many people assume today. Maybe it will become a new infrastructure layer of the economy, something closer to cloud computing or electricity for the digital world. Incredibly powerful, but expensive.

A new kind of divide

This could also lead to a new kind of divide. Not necessarily between people who can or can’t “afford” AI. Open source will likely ensure that baseline models remain broadly accessible.

The real divide may run somewhere else entirely: between organizations that learn to use AI effectively and economically, and those that burn through enormous amounts of capital without proportional gains in productivity.

In practice, this might mean that skills like AI cost management, model usage optimization, agent workflow design, and “AI FinOps” will become just as critical as cloud computing competencies once were.

There’s one more consequence I think is underappreciated. The most advanced AI systems may remain extremely expensive. Not simple chatbots, but autonomous agents, reasoning systems, deep research platforms, multimodal systems running 24/7.

And in that world, competitive advantage won’t come from “the company uses AI.” It will come from one organization being able to do the work of ten others through AI, and doing it economically.

This might be where a new form of economic and productivity inequality is being born. Not between those who have access to AI and those who don’t. But between those who can scale their intelligence through AI, and those still operating in the old model of the world.

Read next

  1. AI won’t run out of customers. It will run out of power and time to repay its debt

    Many people read the sell-off in data center stocks as the first crack in the AI bubble. I look at the same numbers and see something else. The demand is there, and it is substantial. The trouble lies elsewhere: in the construction itself, in access

  2. The AI bubble is the wrong questionLinkedIn

    Microsoft has just pulled out of a deal to build a data centre in Loughton, near London. Not for lack of demand, but for lack of power: the grid connection there won't be ready until the 2030s. And it's not the only case.

  3. We replaced AI subscriptions with four senior developers. It was the best AI decision we made.LinkedIn

    AI was supposed to reduce costs, increase productivity, and allow smaller teams to do more. So we did what most tech companies are doing right now. We bought more tools, added more subscriptions, and started "scaling AI." At first, it worked.