Am I wrong? AI costs for SMEs are going to climb steeply in the next two years, before plummeting enormously.

Let me walk you through my thinking.

The Climb

Right now, AI Agents are approaching human reliability for many roles. Customer service, data analysis, first-draft writing, scheduling, research – the list grows monthly.

Since the core providers are running at a loss, this will naturally mean that your AI Agent price (the “token tax”) will invariably approach the cost of an actual employee in those roles.

Put another way: if I can replace my $3,000/month employee with AI, then I will happily pay $3,000/month for the AI given its extra benefits – it doesn’t take holidays, doesn’t call in sick, and scales instantly. Unless competition pushes prices lower, this is where the sales teams at OpenAI, Anthropic, and Google will push us all.

Your $200/month AI investment today will inexorably bridge the gap to $3,000 or more. You might need multiple AI agents, or more technical expertise that takes the bill to $5,000+.

And you will be happy paying this for your business. Because the value is there.

The Fall

But once a technical level is reached, we don’t need ever-better models.

Think about it: I use Claude Sonnet for most of my basic work, with Opus when something more complex is required. Fable is great, but I rarely need it. Model upgrades behind the scenes can actually cause problems – consistency in our employees is an asset. You don’t want your fantastic accounts person showing up in a super chatty mood one day, unexpectedly, especially if their maths reasoning suffers too!

So how soon until the average SME can afford their own AI server? A GPU (maybe a previous generation, via a data centre that upgraded) and an open-weight model that can deliver the work?

There will always be a use case for the latest model. But why pay the token tax when, for a small capital outlay, you can bring consistency in-house with low variable costs?

The $20,000 Box

A $20,000 or $30,000 box in the corner of your office – running last year’s model, on last year’s hardware – promises the same output for your standard operations, with no monthly token spend.

For a business paying $3,000/month in AI agent fees, that box pays for itself within a year. And then it’s essentially free – barring electricity and the occasional hardware refresh.

Who does that benefit? How do the loss-making giants respond? What happens to the startups building exclusively on API access?

I think the implications are significant – for AI companies, for SMEs making investment decisions today, and for the broader economy.

But I’m curious: am I wrong?