netoperationalslot297,841,022nodes412 / 487utilization68%p5014.2 ms/tok
llama-3.1-70b24.1ms/tokqwen-2.5-72b18.7ms/tokmixtral-8x22b31.4ms/tokdeepseek-r152.3ms/tokphi-48.1ms/tokcommand-r-plus22.9ms/tokgemma-2-27b14.6ms/tokllama-3.1-70b24.1ms/tokqwen-2.5-72b18.7ms/tokmixtral-8x22b31.4ms/tokdeepseek-r152.3ms/tokphi-48.1ms/tokcommand-r-plus22.9ms/tokgemma-2-27b14.6ms/tokllama-3.1-70b24.1ms/tokqwen-2.5-72b18.7ms/tokmixtral-8x22b31.4ms/tokdeepseek-r152.3ms/tokphi-48.1ms/tokcommand-r-plus22.9ms/tokgemma-2-27b14.6ms/tok
buildv0.7.4-rc.2
◇ operator program · openonboarding · 47 nodes this weekaverage revenue · $24k / gpu / month

Run a node.
Earn per token.

You bring the GPUs. We bring the queue. Stake $NIMBUT, register your fleet, get matched to inference jobs from autonomous agents around the world. Settled in USDC, hourly.

◇ avg gpu/mo
$0.0k
at 70% utilization
◇ payout cadence
1h
settled in USDC
◇ minimum stake
0 $NIMBUT
tier 1 · single node
◇ slash cap
2%
per attested fault

◇ flow · from cold gpu to first payout

Five hops. Two hours. Hot model first.

Onboarding flow is one CLI installer and one staking transaction. Most operators are taking real traffic within the same business day.

i.

Install agent-host

curl nimbut.compute/install | sh — one binary, no docker required.

ii.

Stake $NIMBUT

Lock 5k+ $NIMBUT as collateral. Tiers unlock more models & higher routing priority.

iii.

Warm models

Pick from the catalog or upload custom. The host pre-loads weights; warm models pay better.

iv.

Take traffic

The router starts sending you matched jobs within minutes. Output is signed; cost is metered.

v.

Get paid

USDC settles to your operator wallet every hour. No invoices. No 30-day net.

◇ tiers · stake more · earn more

Four rungs.

More stake means: more models you can serve, heavier routing priority, eligibility for premium models, and a seat on the operator committee.

tier 01

Scout

0k$NIMBUT
  • Serve up to 8 models
  • Standard routing priority
  • Single-region only
  • 0.4% slash cap / fault
tier 02

Carrier

0k$NIMBUT
  • Serve up to 24 models
  • Priority routing
  • Multi-region (up to 3)
  • Premium model eligibility
tier 03

Anchor

0k$NIMBUT
  • Full catalog access
  • Top routing priority
  • Multi-region (unlimited)
  • Custom model uploads
  • Committee seat eligibility
tier 04

Constellation

0M$NIMBUT
  • Everything in Anchor
  • Sovereign region (your hardware, your jurisdiction)
  • Direct enterprise contracts
  • Reduced protocol take-rate

◇ specs · what your hardware needs

The bar.

Bare-metal preferred but cloud is fine. We auto-detect your GPUs, run a benchmark, and assign your node to compatible models.

componentminimumrecommendedpremiumnotechecks
gpurtx 4090 (24g)4× a100-80g sxm8× h100-80g sxm5fp8 supported auto
vram total24 GB320 GB640 GBper node auto
cpu16 vcpu64 vcpu128 vcpuamd epyc / xeon scalable auto
ram64 GB512 GB1 TBecc preferred auto
nvme2 TB8 TB16 TBfor model cache auto
network1 gbps25 gbps100 gbpslow jitter required auto
power redundancysingle psudual psua+b feeduptime sla auto
ip diversity/29/27/24for failover auto

◇ math · what one h100 actually earns

Real numbers. Real operators.

From a sample of 18 reference fleets running in fra-4 and nyc-3 over the last 30 days. Numbers below are post-protocol take-rate, pre-electricity.

◇ 4× h100 sxm5 · fra-4 · 30d trailing revenue
peak / day
$3,418
avg / day
$2,807
30d total
$84,210
◇ model mix · same fleet
qwen-2.5-72b62%
llama-3.1-70b21%
mixtral-8x22b11%
phi-46%
◇ warm models pay 1.6×–2.4× premium

If you have idle GPUs,
we have a queue.

The agent economy is starving for compute. Reference fleet on Tier 2 returns ROI on stake in roughly 14 days.