AI cost attribution is moving to the infrastructure layer, because that's becoming the only way the spend stays credible

The question

Spending on AI is incredible. Last month, I wondered how many CIOs will be able to prove the value of AI spending in six to twelve months. Now we are seeing the search for spend credibility find its footing.

The missing connection

The first piece is by Sima Nadler and Alex Meijer for the Cloud Native Computing Foundation. They argue that platform teams can see the GPU bill and the token count, but not the connection between them.

The CNCF post, dated August 5, 2026 in its link, is about OpenCost 1.121.0 and first-of-a-kind Kubernetes inference cost tracking.

This becomes a problem as executives ask ROI questions and find that nobody on the team can answer with numbers.

Attribution under the app

Mota's piece comes on the heels of the CNCF one. In a recent CIO opinion piece, Eduardo Mota makes the case that cloud-style tagging can name a server's owner but not whose AI dollar it was, which is why the bill shows up too late to stay credible.

The model he offers moves that attribution under the app: a kernel-level sensor maps GPU, CPU, and every outbound model call back to the workload that caused it, in real time, with nothing planned in advance.

Its link titles it the case and model for real-time AI cost visibility at the infrastructure layer.

Whose dollar it was

The trend is taking shape as a move to attribute cost at the infrastructure layer. Then spend can stay credible long enough to survive the next budget.

Of course, these efforts will not prove the value of AI investments. They just keep the question from dying on the invoice. Six to twelve months from now, the CIOs who can answer will be the ones who can show whose dollar it was.