IDC forecasts worldwide AI spending of $301 billion in 2026. The investment is real, the productivity gains are real, and the strategic urgency is justified. What is also real — and increasingly visible to the CFOs whose budgets fund it — is that the cost of running AI in production is now growing faster than the cost of buying it.
The first lap of enterprise AI was defined by three conditions running together: urgency to adopt, real productivity gains, and relative affordability. The second lap — already beginning — will be different. Affordability is ending, urgency persists, and the productivity gains are now embedded in operations enterprises cannot easily wind down.
"AI is not free. The companies that internalize that fact first — and engineer accordingly — will define the next decade of enterprise AI."
Behind this reckoning is a structural dependency on subsidized foundation model pricing forming inside enterprise operations. Current per-token prices are calibrated for adoption velocity, not for sustainable margin. As workflows become load-bearing for production operations, switching costs rise sharply.
The companies best positioned for the second wave are not the foundation model labs and not the end-user enterprises trying to build AI capability internally. They are a third category — intermediaries whose proprietary engineering encodes the structured work that can be made deterministic, and invokes large language models where their judgment adds value. This paper proposes a name for that category: Deterministic AI Infrastructure.