The AI bill is coming

One of the best conversations from our last CTO Circle gathering in Madrid was about what happens when AI stops being subsidized.

For the last few years, a lot of AI adoption has been hidden behind flat subscriptions, VC-funded infrastructure and very generous pricing. That phase is ending. The moment AI stops looking like €20/user/month and starts looking like actual infrastructure usage, finance teams will wake up very fast.

Most AI workloads do not need frontier models. That is why inference and AI cloud companies are so hot right now: Baseten, Fireworks, Together, Groq, Cerebras, Replicate, Modal, Nebius, DeepInfra, CoreWeave. They are all attacking different parts of the same problem: how to run AI workloads faster, cheaper and with more control.

When AI stops being subsidized, teams will need to be much more intentional about which models they use, where they run them, and how much each workload should cost. A smaller fine-tuned model can sometimes work better than a frontier model.

Classification, extraction, summarization, routing, internal search, first-pass review, and support triage are valuable workloads, but they should not all have the same cost profile.