GPU estate governance
Efficiency has a price. Now it has a price tag.
Iverson GPU models your AI estate from workload to transformer and prices what breaks.
Every consolidation, curtailment, and cost optimization spends something the dashboard never shows: blast radius, thermal headroom, goodput at risk. Iverson GPU makes that price explicit — then turns the tradeoff dial for you, inside a policy you set.
Read-only. One Helm command. No workload changes. Or run it yourself, right now — the instructions fit on one page.
One governance weight spans the whole tradeoff — and the right setting moves with your workload. A weight tuned on one workload produced 76.5% SLA violations on a single reproducible window of the Philly production trace (replay it live on the frontier page; the paper's 10-window mean is 52%). That is why a controller turns this dial, not a human.
unexpected interruption rate while training Llama 3 on 16,384 GPUs — 419 in 54 days.
Meta, 2024
of paid-for GPU capacity does useful work across measured production fleets.
CAST AI 23k-cluster telemetry · FinOps Foundation 2026
share one NVLink backplane, cooling loop, and power shelf in a rack-scale system. One failure can take them all.
GB200 NVL72 architecture
of global derivatives trading frozen by one shared cooling plant — behind fully redundant chillers.
CME / CyrusOne, Nov 2025
The problems nobody prices
Your dashboards show the symptoms. They rarely price the tradeoff.
Control-loop collision
Autoscalers, rebalancers, and optimizers act on the same pods with no shared objective. The result is churn, flapping, and the 2 a.m. migration storm nobody ordered.
What we do One governance weight arbitrates every loop. In the paper's benchmark, governed placement cut scheduling oscillation 32×.
Correlated-domain blindness
Schedulers see zones and labels. They do not see that eighteen nodes share a cooling loop, a power shelf, and a top-of-rack switch — until all of them fail together.
What we do Physical failure domains are first-class objects. Placement is priced by the blast radius it creates, before anything fails.
The usable-capacity illusion
The dashboard says 119 GPUs are free. Fragmentation, power caps, and quietly failing nodes mean 48 of them can actually take an 8-GPU job.
What we do A safe-usable-capacity ledger, decomposed and priced — the number capacity planning should have been using all along.
Recommendation debt
Optimization suggestions pile up unactioned because every one transfers un-quantified risk to the engineer who clicks apply.
What we do Every recommendation ships with its risk price, confidence, and rollback path — so accepting it becomes rational.
The facility chain
A transformer is a single point of failure your dashboards cannot see.
The cluster API knows nodes and zones. It does not know which racks share a PDU, which PDUs hang off one UPS, or which UPS feeds stand on the same transformer. Iverson walks that chain — supplied once, as node labels or a one-page topology.yaml from your facilities team — and prices the full closure of every domain in it: fail the transformer and everything it feeds fails with it.
On the demo estate: 87% of business-critical GPU capacity stands on Transformer T1 — computed from the snapshot, not estimated.
The same chain runs the other direction: when the grid orders a shed, a governed curtailment pauses idle and batch capacity before it touches a critical training run — the Texas planning page shows both strategies priced.
Dual-corded? Declare feeds: [ups-a, ups-b] and a domain is lost only when every feed is lost. Undeclared redundancy is labeled worst-credible-case — never assumed.
Curtailment
When the grid says shed N kW, which workloads pause — and in what order?
In Texas, curtailment cooperation is becoming a condition of interconnection — and queue pressure in other ISOs points the same way. A blind shed pauses your biggest consumer, usually the critical training run; a governed shed drains idle and batch capacity first and states its time-to-target both ways, expected and worst-case — because the worst case is the compliance clock. Both strategies are priced side by side on the same demo estate.
How it works
Read-only first. Trust is earned in rungs.
Observe
A read-only agent maps your estate: workloads, GPUs, and the physical domains they stand in. First exposure report within a day. No write access, ever.
Govern
Risk-priced recommendations land as pull requests against the policies you already run — consolidation, placement constraints, migration budgets. You approve; we show the receipts.
Self-tune
The controller turns the efficiency–resilience dial continuously, inside hard bounds you set. You choose the envelope; it finds the safe edge — and proves it did.
Built on peer-reviewed research
The governance controller is published, benchmarked, and reproducible.
Self-Tuning Governance of the Energy–Resilience Tradeoff in GPU Cluster Scheduling (IEMCON 2026) evaluates the controller on replayed Microsoft production traces and measured grid-carbon data. The simulator behind our sandbox is held to the paper's reference within 1e-6 relative tolerance (job counts exact) by 39 pinned parity tests — run the experiments yourself.
Honest footnote: results are trace-driven simulation, not customer-cluster measurements. The violation metric is conservative by construction; read paper numbers as paired comparisons, not production SLA figures.
mean SLA violations across the paper's ten production-trace windows when a static weight is transplanted (the single-window live reproduction on /frontier shows 76.5%) — the case for self-tuning.
less scheduling oscillation than greedy energy minimization under a load ramp.
fewer jobs lost when a correlated high-centrality domain fails — exposure avoided before the failure, not recovered after it.
provenance, all three: trace-driven simulation on the Philly production trace — 52% is the 10-window mean, 76.5% the single reproducible window; 32× and 5.5× are paired policy comparisons on the same benchmark
Get your report
Get your free GPU Exposure Report
Tell us where your estate stands and we'll get you the agent. Read-only, one Helm command, first report within a day — no write access, ever.