Loading stats from /api/stats …

PRIVATE P2P INFERENCE · VERIFIED HARDWARE · 32GB VRAM MIN

Inference at the cost of electricity.

Inferencja w cenie prądu. Optymalizujemy zdecentralizowaną inferencję P2P do kosztu energii. Builderzy dostają prywatne LLM-y za $0.02/$0.05 za 1M tokenów. Inżynierowie i entuzjaści edge monetyzują bezczynne GPU — z podziałem zysków.

We optimize decentralized P2P AI inference to electricity cost for builders, sharing profits with engineers and edge enthusiasts monetizing idle GPUs.

Become a Provider →View calculator

$0.02 in · $0.05 out / 1M prefix cache 60s free host 1:1 · marlin · flashinfer

Phase 0 · Spec NVFP4

Context1M ctx · 2M KV

QuantNVFP4 · FP8 KV

Pricing$0.02 / $0.05 · 1M

GPU minRTX 5090 32GB

Blackwell GB202 · 1M ctx = 2M KV cache ~6GB KV + ~16–22GB model → ~28–30GB · 32GB min headroom @ 0.90 util

CUDA 13.3+driver 580+Ubuntu 24.04

NVFP4 · 1M ctx$0.02 / $0.05 per 1MCUDA 13.3 · 580+ · 47900/47901RTX 5090 32GB min · Ubuntu 24.04 · 50GB HF · 60GB+ free

Run a node — Become a Provider

curl -fsSL https://seedinfer.com/install.sh | bash -s -- --authkey hskey-xxx· host 1:1 marlin/flashinfer/fp8/0.93/128/4096 · fleet/providers·/docs

Become a Provider → Fleet Docs

ECONOMICS · UNIT COST

Why 32GB VRAM is the threshold

Long-context inference is memory-bound. The economics are set by KV-cache size — not FLOPs. At 1M tokens, only 32GB+ cards sustain NVFP4 without offload.

Memory

1M ctx = 2M KV cache

NVFP4 W4A16 + FP8 KV · ~28–30GB with overhead · 1M KV alone ~6GB · 32GB min for headroom at 0.90 util.

NVFP4FP8 KV32GB min

Compute

Marlin · FlashInfer · FP8

Blackwell-optimized kernels · chunked prefill · prefix caching · host 1:1 routing · no offload.

marlinflashinferfp8host 1:1

Power

Electricity = dominant cost

At scale, power dominates unit cost. Builder price covers electricity + profit split.

BuilderProvider

Element requires refinement

Builder → Provider split · live metering pending

Revenue split

Builder pays$0.02 in / $0.05 out · 1M

↓

Provider~60%

share of gross · electricity-first margin

↓

Networkremainder · infra & fallback

Element requires refinement

Live payout curve requires verified metering

Network traffic

Requests / min

Tokens / min — stacked Input / Output

Input Output

Token Distribution

0.0% output · 0.0% input

Live Network Flow

Live · MapLibre

Global provider map — 0 cities

MapLibre GL · dots sized by providers

Top regions

Provider locations (city tiles)

CALCULATOR · ≥32GB VRAM

Net-Profit Calculator

Static inputs only — live profit requires verified metering. Adjust your rig specs; result shows refinement placeholder.

VRAM32 – 96 GB

32 GB min64 GB96 GB max

RTX 5090 32GB minimum · H100 80GB · Blackwell headroom static demo only

Electricity$/kWh

Dominant cost driver · local tariff required for live calc

Uptime%

Availability factor · metering pending

Inputs are UI-only and disabled · no mock calculation is performed.

Estimated net profit

Element requires refinementLive profit = (tokens × price × 60% share) − (power × tariff × uptime) · metering not yet connected

Check eligibility → /provider

≥32GB VRAM required · see economics threshold above

Provider fleet · 0 total · showing 0

No providers match filters.

CONCIERGE · WHITE-GLOVE

VIP Managed Node — we configure it for you

Remote Headscale join (tag:provider)
Attestation & verification
50GB HF sync & NVFP4 setup
Bench marlin 0.93/128/4096
Payouts EVM (6 chains) + Solana
SLA & monitoring

≥32GB VRAM only

Element requires refinement

Submit handling requires verification backend

Quick install

curl -fsSL https://seedinfer.com/install.sh | bash -s -- --authkey hskey-xxx

Headscale join after eligibility check · tag:provider

ROADMAP · 4 PHASES

Roadmap · from electricity cost to parity

FAZA 0 · NOW

active

Nemotron Lightning 1M

$0.02/$0.05 · 2M KV · Headscale RTX 5090 32GB · whitelist 20 × $0.40/day + $5 credits + 99% share

FAZA 1 · Q4 2025

is coming

Qwen 3.6 35B A3B + Gemma 4

Vision + MoE · 100 nodes · 60% share stable

FAZA 2 · Q1 2026

planned

1000+ nodes · Edge expansion

WhiteGlove scale · 96GB HBM · referral

FAZA 3 · 2026

vision

Electricity-cost parity

Builder price = electricity + 10% → 0 margin

Element requires refinement

Live progress metrics pending — verified metering & timeline integration required

Models catalog + pricing · Faza 0

Faza 0: tylko Nemotron cache 60s free / 5min max

Faza 0: tylko Nemotron Lightning 1M (2M KV) — $0.02 / 1M input · $0.05 / 1M output

Tailnet w budowie — Headscale control plane (tailnet.seedinfer.com) + provider RTX 5090 (tag:provider). Qwen 3.6 35B A3B i Gemma 4 26B A4B jako „is coming” / „coming soon”. Cache: prefix caching 60s free, 5 min max.

seedinfer/nemotron-lightning-1m

Brak danych — oczekiwanie na /api/stats … (Faza 0: seedinfer/nemotron-lightning-1m)

coming soon

is coming · provider tailnet under construction

qwen3.6-35b-a3b

is coming

Qwen 3.6 35B A3B

Input

$0.06

Output

$0.50

is coming · vision + MoE · 35B A3B

Providers: —

gemma-4-26b-a4b

is coming

Gemma 4 26B A4B

Input

on Modal

Output

—

is coming · 26B A4B · also 8-bit on Modal

Providers: —

Faza 0 · Nemotron $0.02 / $0.05 · 1M ctx · 2M KV · cache 60s free / 5min maxQwen 3.6 35B A3B · is comingGemma 4 26B A4B · is comingCache 60s free · 5min max · prefix caching