Most enterprises running generative AI in production have created a new problem on their cloud bills. Azure OpenAI, AWS Bedrock and Google Vertex AI tend to report spend as one aggregated line, so you see the total but not which model, deployment or team produced it. As AI moves from pilot to production, that gap turns from an accounting nuisance into a budgeting risk.
The playbook that broke
Reserved capacity logic doesn’t survive contact with tokens
For more than a decade, cloud cost management ran on a stable assumption: a workload lived on an instance, the instance belonged to a team, and you bought the rate down with reserved instances or savings plans. Token billing breaks every link in that chain. There’s no instance to reserve, and cost depends on input, output and cached tokens, model choice and provisioned throughput, so a small change to a prompt or a model can swing spend a long way.
The visibility problem is worse than the pricing problem. Configuration happens at the deployment level while cost rolls up at the account level, which means the engineers who can fix the spend are the ones who can’t see it.
Reserved era
Rate optimization
You committed to instances and bought the rate down. The server was the unit, it belonged to one team, and a savings plan made it cheaper for a year. That’s predictable, and increasingly beside the point for inference.
Token era
Unit economics
There’s no instance to reserve. Cost moves with input, output and cached tokens and with model choice, so you manage it by measuring per model, per deployment and per team, then acting on what you find.
What PointFive shipped
A detection engine that routes its own fixes
On June 8 at FinOps X in San Diego, PointFive launched its AI Efficiency OS and a companion product, TokenShift, the same week it announced a $60 million Series B led by Accel. The aim is efficiency that follows waste from the cloud all the way to the coding agent, and in August the company joined the board of the Tokenomics Foundation.
Underneath is DeepWaste, a detection engine with more than 500 validated detections across cloud infrastructure, data platforms, GPU capacity and managed LLM services. For AI it maps token-level economics on Bedrock, Azure OpenAI and Vertex AI, GPU rightsizing and model mismatch, where a different model would deliver better unit economics for the same work.
PointFive also acts on what it finds. It opens the ticket in Jira or ServiceNow, alerts the owner in Slack and generates the fix in the IDE, and nothing changes until a person approves it. TokenShift brings the same accounting into the IDE, where coding agents now spend their tokens.
Read-only discovery
PointFive runs agentless, read-only discovery across your environment, and findings start as soon as the integration is live.
DeepWaste detection
The engine reads configuration, telemetry and code, then surfaces the AI waste that generic cost tools miss.
Agentic remediation
Each finding carries an owner and a savings estimate and goes to the people who can fix it, with nothing changing until someone approves.
Savings measured
Teams track realized savings against PointFive’s estimates, and the loop keeps running as models, prompts and capacity change.
Savings that nobody has the engineering capacity to ship never reach the ledger.
Build or partner
This one isn’t worth building yourself
Our advice here is consistent. Token FinOps is a real discipline, and for almost every enterprise it isn’t worth building in-house. Attribution is hard, the billing mechanics change every quarter, and teams that try usually end up with a brittle dashboard nobody trusts and a roadmap item that never ships.
The fair question is whether you already have enough inference and GPU spend, and enough engineering capacity to act on what a platform finds, to justify a dedicated tool now. When PointFive fits, we’ll say so, and when it’s still early for your environment, we’ll say that too.
