What happens when AI has to pay for itself?

The financials that change the conversation

AI is no longer just a technology story. It has become one of the largest capital allocation stories in modern tech, and the economics beneath the narrative are becoming hard to ignore. OpenAI's financials tell the clearest version of this story: $13 billion in revenue in 2025, $38.5 billion in losses, and $17.2 billion paid to Microsoft for the compute to run its models, more than it earned. Anthropic has committed to $330 billion in infrastructure spending while needing $174 billion in annual revenue by 2029, a number it is nowhere near today. The point isn't that AI is failing. It's that the financial structure supporting the current leaders requires a pace of growth no technology company has ever had to sustain while carrying this level of infrastructure burden. That changes the risk calculation for everyone building on top of it.

For that we need to keep three things in mind:

1. The "AI gets cheaper" assumption may be backwards

The prevailing bet was that AI would get progressively cheaper- better models, lower inference costs, broader adoption. That assumption looks shakier now. Companies under this level of financial pressure don't respond by lowering prices. They recover costs. As token-based and usage-linked billing becomes standard, enterprises are finally seeing what AI actually costs once the subsidised experimentation phase ends.

The real risk isn't that the model stops working. It's that pricing becomes unstable while businesses are still building products and budgets around it.

2. ROI is no longer a soft conversation

Enterprise AI buying behaviour has shifted. For two years, the conversation was capability-led, what the model could write, summarise, or automate. Now the harder question is whether any of that produces measurable business value.

‍Ed Zitron's analysis captures where the market has landed: enterprises are running AI initiatives, but very few have achieved meaningful returns. The conclusion is blunt the technology worked. The value didn't arrive. CFOs who approved AI spend without scrutiny are now asking hard questions. Any product or vendor without a concrete answer to the ROI question is going to lose the conversation.

3. Where the shakeout happens

Some companies have built real value, proprietary workflows, embedded operational use cases, measurable productivity gains, genuine switching costs. Those businesses have a clear reason for customers to keep paying regardless of what happens at the infrastructure layer.

Others are exposed.

Any business that wrapped an API, added a lightweight interface, and called it a product, without a durable workflow, measurable outcome, or defensible position, is vulnerable on every front. Infrastructure pricing rises: margins collapse. Model quality equalises: differentiation disappears. Customers scrutinise ROI: there's no answer.

The next phase won't reward proximity to the model layer. It will reward businesses that built something durable around it.

Risks worth being honest about

• If your product only works because model usage is artificially cheap, that's a structural risk.
• If customers can't clearly tie your product to a business outcome, that's a retention risk.
• If your differentiation disappears the moment a competitor offers similar model quality, that's an existential risk.

Our take

The winners will be companies that treat AI as part of an operating model rather than a novelty layer,  building measurable outcomes, durable workflows, and clear economic logic from the start. AI isn't disappearing. But it is entering its accountability phase, where discipline matters more than hype, and value matters more than capability. That's the transition worth paying close attention to now.

Let's talk about where your AI investments are actually going →

‍