As a founder, one of the most visible costs in the business is no longer software licences. It is AI subscriptions.
I am not particularly surprised by that. The cloud model was built around human interaction, not always-on agentic work. Once you start running agents, you start burning tokens at a rate that a consumer plan was never designed for.
That is a real problem. It is also a useful reminder that AI is not just a software purchase. It is an infrastructure decision.
And once you get past the hype, the practical question becomes simple: where should your core AI workload live?
Cloud pricing was never designed for constant use
At first, the cloud AI story looked easy. A cheap monthly plan. A few hundred tokens here and there. A bit of coding. A bit of writing. A bit of analysis.
Then the usage becomes real. The workflow turns into a loop. The agent reads, retrieves, reasons, checks, retries, and does it again. One small task turns into many model calls. A developer consuming “a bit of AI” suddenly looks a lot more like a GPU farm in human form.
The providers responded, understandably, by tightening limits, increasing pricing, and pushing heavier usage toward enterprise plans. That is not a surprise. It is a business model doing what business models do.
What matters is that it changes the economics for small teams and individuals. The “cheap AI” story stops being cheap when it becomes real work.
Local AI is becoming practical rather than ideological
The good news is that local models are now good enough for a significant slice of real work.
Quantised models are much more capable than they were even a year ago. Hardware costs are still real, but the break-even point is falling. If a team is running agentic workflows frequently, local infrastructure starts to make sense far sooner than it used to.
Once the hardware is there, the marginal cost of the next token is usually much lower than a cloud API charge that can climb with every loop and retry. This matters more than most people admit. At a certain point, the issue is not just “is it cheaper?” It is “can I predict the cost?”
The sensible architecture is hybrid
For most people, the answer is not “everything local” or “everything cloud.” It is a hybrid model.
Keep the routine, frequent, sensitive, and latency-sensitive work local. That includes agent loops, retrieval, repetitive refactoring, summarisation, and the tasks you want under your control. Use the cloud for the peaks: the occasional frontier tasks, the experiments, the hard reasoning problems, or the work that really needs a bigger model.
That is not a pure ideology. It is simple operating logic.
The bottom line
If a team is running AI around the clock through cloud APIs, it is not just using a service. It is accepting a cost model it does not control.
Local-first AI is becoming the practical answer for founders, developers, and small organisations because it brings the work home, keeps the cost curve more predictable, and reduces dependence on cloud pricing that is being shaped by the people selling the service.