Skip to content

Open source · Sep 2026

An open-source AI spend ledger, checked on a real H100 before anyone quoted it

unalloc joins OpenCost allocations with LiteLLM, OpenAI and Anthropic bills and reports the AI spend nobody owns. Before trusting its numbers, I pushed six case studies through it, validated the metering results on a rented H100, and gated the paper's own build.

Headline

12–14 pts

disagreement between two metering rules on one tenant's H100 bill

9

defects fixed before release, each with a regression test

6,241

H100 requests in the validation run, 0 errors

$2.15

cost of the GPU run, 29 minutes with a verified teardown

$23,597

of a synthetic month sent to a Helm chart name by a fallback label

The problem

OpenCost knows what Kubernetes workloads cost. The OpenAI and Anthropic consoles know what API calls cost. A single AI feature uses both, so neither can say what the feature costs or how much of the month’s spend has no owner. unalloc joins those ledgers, with money as Decimal, on whichever label means “owner”, and reports the unowned share.

What I checked

  • The fetch path, not just parsing. Tests against a real local HTTP server check URLs, auth headers and pagination, which fixture tests can’t see. CI also installs the built wheel and runs the demo.
  • Real workloads. A vLLM-style serving simulator, a PyTorch transformer with a real KV cache, tensor- and pipeline-parallel inference, and the CLI against mock provider APIs.
  • Real hardware. vLLM 0.29.0 on one NVIDIA H100, with a one-hour budget cap, a free dry run first, a watchdog, captured evidence for every step, and deletion verified by resource ID.
  • The paper. A build that fails on Typst that compiles but renders wrong, and a disclosure of AI tool use.

What it found

  • Nine defects in the tool, including every live billing fetch failing and pagination reading only the first page.
  • Fallbacks can hide the problem. With owner labels on leader pods only, a distributed deployment’s month was 66% unowned. A name fallback cut that to 4.4% by sending $23,597 (61%) to a Helm chart name, so reports now show fallback-attributed spend separately.
  • Metering is a decision. On the H100, a token meter and a time-share meter disagreed by 11.7 to 13.7 points on the retrieval tenant’s share, while GPU utilization read 97–99% at every load. Neither rule is a ground truth, and the paper says so.

Limits

One GPU, one model and synthetic traffic, with illustrative prices. The dollar figures come from generated case-study months, not a real company’s bill.

See it

Cash-pay advisory
from $250/hr

Check fit