Preparing for the extraction phase of AI
The free ride is ending. If you have been using AI tools heavily over the past two years, you have benefited from one of the most aggressive subsidy campaigns in the history of consumer technology. Companies were burning capital to get you hooked, to gather data, to win market share. Inference was expensive, and they were eating the cost. That era is winding down.
What is coming next is what Tim Wu would call the extraction phase: the period when AI companies shift from acquiring users to monetizing them. We have seen this pattern before. When streaming services launched, the first few years were a golden age — cheap subscriptions, massive content libraries, no ads. Then came the price hikes, the password-sharing crackdowns, the ad-supported tiers. The economics always catches up.
Think of it like moving into a newly built neighbourhood. For the first year or two, the developer keeps rents artificially low to fill the units. Once occupancy is high enough, the market rate kicks in. You did not sign up for the market rate, but here you are.
The signals are already visible. Free tiers are shrinking. Usage limits are getting tighter. Subscription prices are creeping up. Some perks that launched with fanfare have quietly disappeared. The companies are not doing this out of malice — inference costs are substantial, and the era of loss-leader AI was never going to last indefinitely.
The practical question is how users should adapt to this new extraction phase. Several approaches are worth considering.
The first is that a multi-tool approach will serve you better than loyalty to a single platform. The AI landscape is competitive enough that playing providers off each other is a legitimate strategy. If one raises prices, there are alternatives. Staying informed about what each tool actually costs, and what you are actually getting for it, matters more now than it did when everything was effectively free.
The second is that local models warrant closer scrutiny. Running a capable AI model on your own hardware has gone from a hobbyist experiment to a practical option for many use cases. It is not right for everything, but for tasks where privacy, cost, or availability matter, it is worth understanding what is possible.
The third is to think in tiers rather than defaults. Not every task requires the most powerful and expensive model. A sophisticated reasoning task and a quick summarisation job do not need the same tool. Getting intentional about matching the job to the appropriate tier is a habit that will save money and probably improve results.
None of this is cause for alarm. The tools are useful, and most of them will remain accessible at a reasonable price. But the dynamic is shifting, and the people who adapt early — who treat AI access as something to manage rather than take for granted — will be better positioned than those caught off guard.
The subsidized era taught us what AI could do. The extraction phase will teach us what it is actually worth to us.
What would change about how you use AI today if you knew the price was going to double next year?
Take care,
Emanuel