The AI cost blind spot: why your fastest-growing bill never hits a cloud invoice

Ask a finance team where their AI spend is and you'll usually get a shrug and a link to a dashboard that shows compute, storage, and network — but not tokens. That's not an oversight. It's structural.

Why classic FinOps can't see it

Three things conspire to keep AI spend off the radar:

• It bills outside the cloud. Direct API spend on OpenAI or Anthropic never appears on your AWS or Azure invoice. Your cost dashboard ingests the cloud bill — and the cloud bill doesn't know your model calls exist.
• Keys go untagged. A model deployment or API key rarely carries the cost-allocation tags your infrastructure does, so even when spend is visible, it can't be attributed to a team or feature.
• Deployments look free until the invoice lands. A provisioned model deployment shows no meter until traffic flows — and by then the number is already large.

Token consumption is growing 20× toward 2030 while the per-token price floor has stopped falling. AI bills go up, not away.

What seeing it actually looks like

Reading AI spend means reading each provider's admin API the way you read a cloud: enumerate the model deployments and API keys, pull token usage per model, and price it against real input/output rates — because an input-heavy agent workload priced on a blended rate can be overstated two or three times over.

Once you can see it, the waste is obvious and specific: idle model deployments still holding provisioned capacity, dormant API keys that are pure attack surface, prompt-caching opportunities hiding in a 100-to-1 input/output ratio, and traffic parked on a premium model that a cheaper tier would serve.

None of that is visible on a cloud invoice. All of it is measurable — if you look in the right place.