nozle
    nozle

    GPUs are scarce.Margin is scarcer

    Token prices fall every quarter. GPU capacity doesn't. And the tenant burning your H100 pool looks identical to the one paying for it. Price every tenant on what they actually cost — tokens, GPU-seconds, and bursts — without rewriting your billing stack.

    Book a teardown
    Start free

    Four things every inference platform needs to ship.

    See cost per tenant. Re-cost pool moves without a price change. Gate quotas on the hot path. Compose tokens and GPU-hours into one plan. One ledger, wired in once.

    Cost, by tenant. Live.

    Every prefill, every decode, every GPU-second pinned to the tenant who triggered it. The noisy neighbor and the underwater account stop hiding inside an aggregate cloud bill.

    • GPU + token + storage costs joined to tenant_id. One ledger, not three spreadsheets.
    • Alerts the second a tenant flips negative. By pool, by region, by SKU.
    • Re-rate any month against any future plan. "What if H100 traffic moved to L40S?"
    tenant cost ledger · livetoday · 14:32
    tenantmrr7d trendgross margin
    • Loam AI
      Cluster · H100
      $48,200
      +364%
    • Spire Robotics
      Pool · A100
      $12,400
      +152%
    • Aperture Labs
      Pool · L40S
      $3,800
      −938%
    • Drift Synth
      Burst · shared
      $640
      −35-22%
    • Forge Vision
      Cluster · H100
      $26,500
      +259%
    • Pico Genomics
      Pool · A100
      $2,100
      −51-38%
    2 underwater · 1 noisygpu cogs joined
    pool swap · plan v223 pool swaps · re-costed
    planpool · was → nowimpact
    • Free
      tier · 01
      vllm-shared→l40s-batched
      +18%
      margin
    • Team
      tier · 02
      a100-shared→h100-shared
      +9%
      throughput
    • Cluster
      tier · 03
      h100-private→h100-private
      —
      no change
    • Enterprise
      tier · 04
      h100-reserved→h100-reserved
      🔒
      contract
    versioned · reversiblere-costed from your cost sheet

    Move traffic between pools rewriting your pricing.without

    An H100 pool fills. A new region opens. Batched throughput drops the unit cost by 40%. Update the pool's rate in your Cost Sheet and every new event is re-costed — the price your tenants see doesn't move.

    • Per-pool cost rates. Different tenants, different pools, same ledger.
    • Plans are versioned and reversible. Roll back v17 of your pool map without a migration.
    • Forecast margin before you commit. Margin trend and forecast per tenant.

    Enforce quotas at the gate. Not the dashboard.

    A free tenant ships a viral demo. An enterprise wants a paid burst before close. A scraper hammers your embedding pool at 3am. Set quotas per tenant, per pool, per SKU — enforced on the hot path, billed when they overflow, lifted with a flag instead of a deploy.

    • Quota checks on the hot path. Atomic check-and-deduct on every request.
    • Block, or bill the burst. A burst is a line item, not an outage.
    • Per-tenant override on the fly. Lift a cap for one customer. No deploy, no PR.
    quota gate · livehot path · /can
    tenantquota · usageverdict
    • Loam AI
      Cluster · H100 · reserved
      240 GPU-min/hr182 GPU-min/hr
      within
    • Spire Robotics
      Pool · A100 · shared
      8,000 tok/s6,420 tok/s
      within
    • Aperture Labs
      Pool · L40S · shared
      1,200 tok/s1,200 tok/s
      blocked
      └─ free tier · hard limit hit
    • Drift Synth
      Burst · shared
      0 GPU-min/hr+12 GPU-min/hr
      burst
      └─ burst until 14:32 · auto-bill at $0.04/min
    • Forge Vision
      Cluster · H100 · reserved
      180 GPU-min/hr92 GPU-min/hr
      within
    3 within · 1 blocked · 1 burst0 PRs · 0 deploys
    plan.compose · v22 draftno deploy
    prompt · plain english
    4 lines
    >Charge $0.0008 / 1k input tokens, $0.0024 / 1k output.
    >Plus $2.40 per H100 GPU-hour, billed by the second.
    >First 1M tokens free. Reserved capacity gets 25% off.
    >Bursts past quota auto-bill at $0.04 / GPU-min.
    plan · compiled
    json · valid
    • meter[in]$0.0008/ 1k tok · free_tier: 1,000,000
    • meter[out]$0.0024/ 1k tok
    • meter[h100]$2.40/ GPU-hour · per-second
    • burst[h100]$0.04/ GPU-min · over-quota
    • discount25%reserved_capacity
    v22 · liveversioned · reversible

    Mix tokens and GPU-hours in one plan.

    Inference bills three units at once: tokens, GPU-seconds, and reservations. Describe the plan in English; ship it versioned, without a deploy. Hybrid pricing stops being the quarter you don't ship features.

    • Compose any inference shape. Token, GPU-min, request, reservation — combine per plan.
    • Versioned at the tenant level. Grandfathering is a flag, not a migration.
    • A/B prices, then keep what wins. Promote the version that lands the margin you want.
    Not an inference platform?

    See the four shapes of pricing.

    Usage, hybrid, agentic, margin. Pick the one your customers actually buy on — Nozle ships all four on the same fabric.

    Open the overview

    Stop guessing.
    Start knowing.

    Wire Nozle into your stack and see your first margin report in 60 minutes.

    Book a teardown
    Start free
    Setup
    60 min
    Card required
    no
    Sales call
    no
    First invoice
    in 60 min