unalloc joins OpenCost Kubernetes allocations with LiteLLM, OpenAI and Anthropic bills into one exact ledger and reports how much of your AI spend has no owner. Six case studies, one of them on a real H100, show where that attribution breaks.
$ unalloc report --fixtures --dimension team
Most failures happen where systems meet, not inside any one of them. Each number is read from the study's results when this page is built, and links to them.
of a distributed deployment's GPU bill had no owner when team was set only on LeaderWorkerSet leader pods.
of that bill resolved to a Helm chart name after the obvious fallback label, while the headline unallocated share fell to 4%.
difference in a retrieval-heavy tenant's share between a token meter and an equal time-share meter, on an H100 running vLLM, at every load tested.
GPU utilization at every load while throughput rose 7×. Power draw (469 W → 660 W) tracked the work; utilization did not.
of gateway spend counted twice when the gateway and the provider bills were both turned on, in a $41,420 month.
of true provider spend reported by billing adapters that read only the first page of results. Fixed in 0.2.0.
Each study emits payloads in the providers' own formats and runs them through unalloc's real adapters. Raw data and scripts are in the repository.
real servingvLLM 0.29.0 serving Qwen2.5-7B-Instruct on a DigitalOcean H100, driven with four tenants at 2–16 requests/s. 6,241 requests, 0 errors, about $2.15.
Token and time-share meters disagree by 12–14 points on the RAG tenant; utilization reads 97–99% at every load.
Code and data →
real inferenceTensor- and pipeline-parallel inference on torch.distributed, verified token-identical to a single process, then a month of LeaderWorkerSet allocations.
Leader-only labels leave 66% unowned; falling back to name sends 61% to a chart name.
Code and data →
real inference · CPUA from-scratch PyTorch decoder serving a 96-request, four-tenant trace, with cached decoding checked against recomputation.
Token counting and measured compute disagree by 33 points of the pool; less on a batching GPU.
Code and data →
simulationDiscrete-event engine with paged KV blocks, prefix caching, chunked prefill and continuous batching, metered five ways.
Step time sees almost no idle capacity; KV memory leaves 83% of the bill with no request.
Code and data →
real CLI · synthetic spendThe unmodified CLI against mock OpenCost, LiteLLM, OpenAI and Anthropic APIs with real auth schemes and cursor pagination.
Every source on double counts all gateway spend; the gateway alone covers 72% of invoices.
Code and data →
mixedLabeling Pareto, per-feature unit economics, self-host break-even and a CI budget gate on the same ledger.
Three label fixes take a 67%-unallocated org to 1.9%.
Code and data →Sort unowned spend by dollars. In the hybrid-org study three label changes took unallocated spend from 66.9% to 1.9%.
unalloc labels --dimension teamJoin cluster and API rows on a feature label. A RAG answer cost $22.32 per 1k requests with its vector database, $13.38 counting the LLM bill alone.
unalloc report --dimension featureOne self-hosted GPU beat a mid-tier API above about 0.23 requests/s and a small-tier API above about 1.7 requests/s in the simulation. Utilization decides it, not list price.
python -m case_studies.use_casesFail CI when too much spend has no owner. Exit code 2 above the budget.
unalloc report -D team --budget 10Record each input by digest and keep the payloads. Later, verify rebuilds the report, or says which input changed, went missing, or which rule moved the result.
unalloc report -D team --json --keep-inputs run/ > run.json
unalloc verify run.json --inputs run/The full paper: the tool, six studies, related work and limitations. A preprint on arXiv (cs.DC, cross-listed to cs.PF), not yet peer reviewed. PDF built from this repository.
The readable version: what was measured, what surprised me, and advice for teams setting up LLM showback.
Machine specs, results and the captured bring-up and teardown of the GPU run.
Every command to repeat the H100 run on DigitalOcean, including the two things that went wrong.
Walk through each study with live code.
To cite the findings, cite the paper, arXiv:2609.24991.
@misc{urista_kvcache_2026,
author = {Urista, Timothy},
title = {Who Pays for the {KV} Cache? Attributing Shared {AI} Inference Spend Across {Kubernetes} and {LLM} Provider Bills},
year = {2026},
eprint = {2609.24991},
archivePrefix = {arXiv},
primaryClass = {cs.DC},
doi = {10.48550/arXiv.2609.24991},
url = {https://arxiv.org/abs/2609.24991}
}To cite the software, cite the release. Every release is archived on Zenodo at doi:10.5281/zenodo.22761012, which resolves to the latest — currently 0.3.2. The paper describes 0.2.3; 0.3.0 adds result provenance and unalloc verify without changing any attribution result. 0.3.1 clarifies source overlap and research limitations in the documentation, and 0.3.2 makes the bundled demo non-overlapping with sample invoices.
@software{urista_unalloc_2026,
author = {Urista, Timothy},
title = {unalloc: find the AI spend nobody owns},
version = {0.3.2},
year = {2026},
doi = {10.5281/zenodo.22761012},
url = {https://github.com/timurista/unalloc}
}