Chargeback for AI platform teams.

AI is the fastest growing line in most cloud estates, and it is growing under a platform team that often cannot say which product caused the increase. Chargeback closes that gap. Here is how to build it without breaking the platform.

Why is AI chargeback different?

Ordinary cloud chargeback leans on resource tags: a VM, a database, or a bucket belongs to a team, and the tag carries the cost. AI breaks that model because the expensive resources are shared. A single pool of GPUs serves many teams, one managed inference endpoint answers requests from every product, and a provisioned throughput reservation is bought once and drawn down by everyone. The cloud bill faithfully reports the pool, the endpoint, and the reservation, but it has no idea that team A sent ten times the tokens of team B. The cost is real and the consumer is identifiable, but only inside the platform, not in the billing data. That is the gap chargeback has to bridge.

What usage signal should drive the allocation?

Pick the metered unit that most closely tracks the cost driver for each service. For token based inference through Bedrock, Azure OpenAI, or Vertex AI, input and output tokens per team are the natural signal, weighted by model because a frontier model and a small model cost very differently per token. For self hosted models on a shared GPU cluster, GPU seconds or GPU memory time per tenant maps better than tokens, because cost follows occupancy of the accelerator. For provisioned throughput or capacity reservations, allocate the committed cost in proportion to each team's measured consumption of that capacity. The platform already has to emit these signals to run; chargeback simply tags them with a tenant identifier and rolls them up.

Who pays for idle capacity?

This is the question that decides whether AI chargeback is fair. Reserved GPU capacity and provisioned throughput are bought ahead of demand, so there is always headroom no team consumed. If you spread that idle cost across the teams, you punish them for a sizing decision they did not make, and they will reject the bill. The correct treatment is to allocate only consumed capacity to products and hold the idle remainder as a platform cost, then surface that idle line prominently. That separation does two things at once: it gives teams a number they accept, and it puts visible pressure on the platform to right size its reservations, because the idle line is now its own cost to defend.

Showback first, then chargeback

Resist the urge to move money on day one. Showback, presenting each team its real AI cost without yet charging it, builds trust in the allocation and exposes the obvious waste before anyone has to defend a budget. Run showback until the unit economics are stable and teams agree the numbers are right, then switch to chargeback so the consuming team owns the line in its own budget. Chargeback is what changes architecture decisions, model choices, and retry logic, because the team now feels the cost directly, but it only works on an allocation teams already believe. Skipping showback turns the first chargeback into an argument about the data instead of a conversation about the spend.

How do you keep the allocation honest over time?

AI pricing and model mixes change fast, so the allocation needs a cadence, not a one time setup. Reconcile the allocated total against the actual cloud bill every cycle so rounding and untagged usage do not drift. Reweight model factors when providers change token pricing or you adopt new models. Track unit economics per team, cost per thousand requests or per active user, so a rising bill can be read as growth rather than waste, or waste rather than growth. And keep the idle capacity line in front of the platform every review, because right sizing reservations to a defensible forecast is where the largest AI savings usually sit.

Make AI spend a number every team owns

We design AI chargeback and showback that allocate GPU, token, and inference cost fairly across AWS, Azure, GCP, and OCI, with zero provider commissions and unit economics your board can read. Our guarantee: we reduce your cloud spend or we reimburse our service fee. Pricing is either a Fixed Fee scoped up front or Gainshare, a share of verified savings with no retainer and no risk.