Every chart about AI adoption looks like a hockey stick right now. Users are up, revenue is up, and the headlines are relentlessly bullish. But nobody is talking the actual cost of delivering all that AI.

Right now, AI is heavily subsidised. Anthropic, OpenAI, and the rest aren't building profitable businesses… They're spending investor money to make AI feel cheap.

Think about what it actually takes to run a large language model. Every query you send burns compute on chips rented from Amazon, Microsoft, or Google. Those chips are expensive. The models sitting on top of them were trained at a cost of billions of dollars, and that training doesn't stop. It's a recurring cost, not a one-time sunk expense. Every new model costs money to build and money to run.

Anthropic has estimated that a typical Claude Code user costs around $13 a day to serve. That's hundreds of dollars a month. Right now, most of those users are paying a flat $20 monthly subscription. The gap between what AI costs to deliver and what customers actually pay is enormous. Right now, venture capital and hyperscaler partnerships are quietly filling that gap.

When you hear that AI adoption is surging, a more accurate version of that sentence is: a lot of people are taking a very good deal that can't last.

The standard rebuttal here is Moore's Law: technology always gets cheaper over time, so why would AI be different? It's a fair question.

The problem is that AI training costs aren't behaving like historical technology costs. They're going up, not down. Every generation of model requires more compute, more data, and more human oversight than the last. This isn't a manufacturing scale problem that gets solved by building more fabs. The raw material (high-quality data, human feedback, and frontier compute) is scarce and getting more expensive.

More tellingly, companies that are using AI at real scale are getting a shock. Uber, ServiceNow, and Microsoft have all reportedly blown through their AI budgets in a fraction of the time they expected. When the invoice arrives at enterprise scale, the economics look very different from the startup demo.

There's a popular argument that AI is following the Uber playbook: lose money now, build the network, raise prices when you have scale and loyalty. VCs subsidised our rides for a decade. Eventually Uber turned profitable. Couldn't AI do the same?

It's a tempting analogy, but it’s important to remember Uber subsidised the trip. It made getting from A to B cheaper, and it charged per trip. The unit economics were lossy, but they were at least legible. You could see a path to profitability at scale with pricing adjustments.

The level of subsidisation we’re seeing with AI is as if Uber had paid for the cars, the petrol, the driver's salary, and the passenger's clothes… and then charged $20 a month for unlimited travel. There's no natural inflection point where scale makes the cost structure work.

To actually become profitable, Anthropic and OpenAI don't need to grow their way out of losses. They need to raise prices, a lot. The question for investors isn't if that happens, but when, and who's positioned for it when it does.

Pivoting to software:

For the past two years, the bear case for enterprise software companies has been: what if enterprises just build their own AI tools? Why pay for an expensive software subscription when you can vibe-code your own workflows internally?

That argument made sense when AI tokens were nearly free. It makes much less sense when they're not.

Adobe, Salesforce, ServiceNow; these companies are already pivoting from seat-based pricing to credit-based pricing. On the surface that looks like a defensive move. But think about what it actually means: they're building the cost of AI tokens directly into their pricing models. They're profitable businesses with pricing power and enterprise relationships. When token costs rise, they adjust their credits. The cost passes through to the end user.

The AI subsidy will end one day. We're approaching the point where it's cheaper for most businesses to partner with a mature software provider than to run their own AI infrastructure.

The companies that have been punished by the market's fear of cheap AI disruption may, paradoxically, be some of the best-positioned names for what comes next.

Given the current valuations, the market isn't pricing that in yet. That's worth paying attention to.

Upgrade to unlock full access to the RatedA Portfolio (with weekly updates on every buy/sell and rationale), and the complete database of companies + ratings.