Skip to content
Iverson GPU

Pricing

Free to see. Priced to govern.

Observing your own estate should not cost money. Governance — priced recommendations, policy floors, changes that ship as reviewed PRs — is what we charge for, per GPU actually under governance.

Observe

Free

read-only, forever

See what your estate is actually doing before anyone asks you for money.

  • Read-only collector (get/list/watch — no write verbs)
  • The Exposure Report, computed from your snapshot
  • In-browser rendering at /report/live — your estate file never leaves your machine

Govern

from $60

per governed GPU / month · design-partner terms available

The report stops describing and starts recommending — every action priced, risk-scored, reversible.

  • Priced recommendations, ranked by value per unit of risk
  • Policy engine: blast-radius floors, migration budgets, loop referee
  • Governed PRs — demonstrated on our own production estate (a draft PR with its apply/rollback pair); pilots scope one governed PR end-to-end, not an autopilot

Design-partner pricing is negotiable; the number anchors, the partnership decides — and it targets ≥4× measured ROI, with the pilot's own findings setting that number before any annual figure is discussed.

Facility / Estate

per MW

or per site · same deal: hypothesis pricing, design-partner validated

For colo, neocloud, and power-constrained operators — governance of the building, not just the cluster.

  • Facility closure blast radius — transformer/UPS/PDU chains, feeds:[] redundancy
  • Curtailment planning: governed shed vs. blind pause, time-to-target stated expected and worst-case
  • Cross-cluster merge — shared facility domains joined across estate files
  • Topology worksheet + validator (cycles and orphan feeds are hard errors)

The unit of value here is a megawatt, not a GPU — the first facility partner sets the rate.

Autopilot

With Govern

later — when it clears the bar

Self-tuning inside your envelope: the controller moves the dial, your policy floors bound it.

  • Continuous λ-tuning within the limits you set in Govern
  • Every autonomous move logged, priced, and reversible
  • Ships to Govern customers first; no separate SKU until it earns one

Does the math work

The arithmetic you should run on us, run on our own demo: the sample fleet's inbox carries $126K/mo of priced actions; Govern on its 1,728 GPUs would run ~$104K/mo. If the inbox does not pay for the product, the product is priced wrong — that is the test we invite, not the one we dodge.

both numbers derive live from the fleet page's data: $126K/mo is the sum of the inbox's $-denominated actions · ~$104K/mo is 1,728 GPUs × $60

how it ladders to the fleet page's figures: $126K/mo is the priced-inbox total — every open $-denominated action at face value; the fleet's $151K/mo "recoverable" applies a 55% realization haircut to idle-allocated spend; its $146K "awaiting approval" is the dollars on the subset of actions already simulated and queued for a human yes

Talk to us before you pay anyone anything

This is hypothesis pricing, being validated with design partners. There is no billing system yet — the first conversations set the terms. What a design-partner engagement actually looks like is written out on /pilot — four weeks, scoped in writing, with go/no-go criteria. If power is your constraint, start with /curtailment.

Talk to usor start with the free Exposure Report

Hypothesis pricing, validated with design partners — the per-GPU number is an anchor, not a rate card. No billing yet; nothing on this page charges a card.