Decentralized GPU rental platform on Bittensor. CLI and Python SDK for renting GPUs, managing pods, and deploying ML workloads.
SN 51. Rank 2 of 32 by emission
lium.io
Emission8.0%of the block reward · rank 2 of 32
Participants63 miners · 9 validators
Registration0.5 TAO
Verified2026-09-09links checked and scores merged
The programcompute · github.com@606d4a7 · read 2026-09-22
The work
GPU time by the second. A miner, called a provider, runs a small coordinator plus one or more NVIDIA GPU machines and sets an hourly price per machine. What is paid for is a machine that passes the validator's probe: rented machines are paid from one pool, eligible idle machines of the listed models from another.
own docs · lium.io
a decentralized GPU rental platform on Bittensor. CLI and Python SDK for renting GPUs, managing pods, and deploying ML workloads
How it is scored
Two formulas. A rented machine scores score x gpu_portion x gpu_count / total count of that model, where gpu_portion is a per-model share the platform derives from rental revenue. An idle machine of an eligible model scores gpu_count x hourly_rate x a cap multiplier, min(idle GPUs in its tier, tier cap) / idle GPUs. Both are multiplied by sysbox, driver and uptime gates.
code · raw.githubusercontent.com · neurons/validators/src/incentive/rental_price.py
result.rental_share * result.gpu_count * result.effective_rate / result.total_rental_cost
How the pay splits
Three pools, each proportional inside. Rented machines share the mining pool by score. Idle eligible machines share a second pool sized by their dollar rate per epoch divided by 0.41 and by the epoch's emission in dollars, capped at the burn ceiling. What the idle pool does not claim goes to designated burn UIDs. The packaged ceiling is 0.91, so the rented pool is 0.09.
code · raw.githubusercontent.com · packages/lium-core/src/lium_core/shared_config/defaults.py
total_burn_emission=0.91,
Sources disagree The docs emission page says the rented pool is a fixed 13 percent and the ceiling moved from 91 to 87 percent in mid-2026; the live endpoint lium.io/api/v1/shared-config returned total_burn_emission 0.9 on 2026-09-19 and again on 2026-09-22; the packaged default at the pinned commit is still 0.91.
Who judges
Nine validators were active on the 2026-09-09 chain snapshot, of 64 permitted, though the docs say Lium runs a single validator and point others to a community auditor. Weights are committed then revealed, at most once per 100 blocks. The validator connects to every machine every 15 minutes, checks the sysbox runtime, hardware and a test workload, and reads rental state from the platform's backend.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled True, weights_rate_limit 100, max_validators 64; chain-2026-09-09.json active_validators 9
How often it turns
Emission settles every 360 blocks, about 72 minutes. The validator checks each machine every 15 minutes, accumulates scores across those cycles, then normalizes and sets weights once per tempo. The idle-pool share is recomputed every cycle from live TAO and alpha prices. A new registration is immune from deregistration for 5,000 blocks, about 16 hours.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 5000
Who buys
Renters paying per GPU-second. The pricing page lists a dollar price per GPU-hour for every model and says every pod is billed per second with no minimum. The docs name developers and AI agents as renters. Providers receive rental fees at the rate the live config sets, 0.95 on 2026-09-19. No customer is named and no dated usage or revenue figure appears on any Lium page.
own docs · lium.io
billed per second with no minimum
What has been gamed
Faked or secretly busy hardware. The validator pins hashes of the NVIDIA management library per driver version, treats a GPU over 95 percent busy with under 1 percent memory as a wedged ghost, and flags held VRAM above 2,048 MB on an idle node. Idle-pool gates: price above market p90 times 1.1, disk under 1.5 times VRAM, no power-cap rights. Without sysbox every score is 0.
code · raw.githubusercontent.com · neurons/validators/src/services/const.py
GPU_HELD_VRAM_MB_LIMIT = 2048
5 of 7 fields rest on the chain or the code. A manifest is the site's reading of the subnet's program, not a review; each field links its source. All programs · What a field means · Correct a field
What it says
The subnet in its own words: the line the index took from its front door, and what the team has said on the record.
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
SN51 Lium rents out GPU time by the second, like a holiday-let site where the properties belong to strangers and the site pays hosts a retainer even when nobody is staying.
The commodity, explained from zero
A GPU is the graphics chip that does most AI computing, and its time is sold by the hour the way a hire car is. Lium's unit is one GPU for one second. The front page is a list of machines, ninety-two available today, with a price per hour on each: a consumer RTX 4090 from about 0.35 dollars, an H100 at about 1 dollar, an eight-GPU B300 server at 64 dollars. The pricing page states "Every pod is billed per second with no minimum."
It has a price because AI work needs these chips, they are scarce, and most people who need one for an afternoon will not buy one. Each listing shows where the machine sits, how reliable it has been, and how fast its network is, the way a holiday let shows location and reviews.
The materials say the renters are developers and AI agents, and give them a one-line install and an API. No customer is named and no dated usage or revenue figure appears on any Lium page; the only numbers are supply.
Why it is on Bittensor at all
Two things change. First, anyone with a supported NVIDIA card, from a datacenter H200 down to an RTX 3060, can list it; the site calls this "permissionless". Second, the subnet's newly minted tokens pay hosts while their machines sit idle, provided the machine passes seven eligibility checks. That retainer for empty rooms is where the holiday-let comparison breaks: no letting site pays a host for a vacant week, and this one does, from token emission rather than rent. Whether Lium is cheaper than a centralized cloud is claimed on the front page ("Unmatched Prices") but not shown against any competitor's price.
How the work gets done
Miners, called providers here, run a small coordinator program plus one or more GPU machines, and set their own hourly price. Validators, the judges who decide who gets paid, connect to each machine, check for a required container runtime called sysbox, measure its hardware and speed on a test workload, and score it; without sysbox a machine "earns no emission and cannot be rented." Emission, the subnet's share of new tokens, is split into three pools: a fixed 13 percent for machines currently rented, up to 87 percent for eligible idle machines, and the rest burned. The pages fetched do not say what share of a renter's dollars reaches the provider. The docs say Lium runs a single validator, operated by the Lium team.
How you would know it works
The front page at lium.io is a live marketplace: each pod shows its price, its uptime, and its reliability percentage, and a reader can browse it without an account. That is the artifact. There is no dated usage figure and no public benchmark.
What is missing
The on-chain description is a slogan, "revolutionizing the democratization of compute", and the contact field is empty. There is no staker or tokenomics page. The one validator is run by the team. There is no status page and no service-level agreement; support is a Discord ticket or an email. The docs do not state an immunity period or when a provider is dropped.
Go deeper
Sources for this explainer
- https://lium.io
- https://lium.io/pricing
- https://lium.io/llms.txt
- https://docs.lium.io/bittensor-subnet/overview
- https://docs.lium.io/providers/rewards/emission.md
- https://docs.lium.io/providers/rewards/penalties.md
- https://docs.lium.io/providers/architecture.md
- https://docs.lium.io/validators/index.md
- https://docs.lium.io/security/official-support-and-scams.md
Metaphor: a holiday-let site whose properties belong to strangers. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.
How legible it is
Four audiences, four questions each, rated on what a first-time reader can find in five minutes. The scale measures findability, not quality.
On-chain identity 5/7- subnet_name
- lium.io
- github_repo
- github.com
- subnet_contact
- not set
- subnet_url
- lium.io
- discord
- p383_54249
- description
- revolutionizing the democratization of compute
- additional
- not set
Stakers and validators●●●●● 3.3
Should I allocate here? · rank 12 of 32 for this audience
Q1What it is●●●●● 4
Rent High-Performance GPUs at Unmatched Prices
with per GPU-hour prices on the front page and a pricing page per model, the chain description is the tagline revolutionizing the democratization of compute rather than a commodity statement, took the lower score
Q2Who it is for●●●●● 3
Showing 83 / Available Pods
live marketplace with per pod uptime and reliability percentages, supply not demand, no dated usage or revenue figure on any own page, the rewards page gives only an illustrative earnings example
Q3How it resists gaming or fails●●●●● 3
emission page names three pools (rented 13 percent fixed, unrented 0 to 87 percent, burn), a mining_score formula, seven eligibility conditions and a penalties page, three clicks from the front page and written for providers, validators page says one validator run by the Lium team, no subnet failure modes
Q4Identity and documentation●●●●● 3
identity 5 of 7, contact and additional not set, discord field is a username string, github and url resolve, no staker or tokenomics page, the validators page points to a community auditor repo
Miners●●●●● 4.0
Can I compete, and what wins? · rank 3 of 32 for this audience
Q1What it is●●●●● 4
Probe each node for the required sysbox-runc runtime, hardware specs, and synthetic workload performance.
Provider runs a CPU coordinator plus GPU nodes, validators probe and score, renters connect for GPU-hour workloads, no request or response schema
Q2Who it is for●●●●● 4
meaningful rental revenue depends on actual customer rentals matching your GPU type and price
supported GPU families listed with a full weighted list in Architecture, registration cost is dynamic with btcli subnet list --netuid 51 named as where to read it (chain burn 0.5 TAO today), competitiveness statement is soft so took the lower score
Q3How it resists gaming or fails●●●●● 4
scoring in prose on the emission page with a cap dilution example (64 of 128 GPUs pays 0.5 of rate), penalties page names three automatic penalty codes and a Secure to Spot tier drop that halts incentives, no immunity period or deregistration policy
Q4Identity and documentation●●●●● 4
provider quickstart plus node quickstart and troubleshooting pages, latest release executor-v1.123 dated 2026-08-19, repo pushed 2026-09-09, docs pages carry no last updated date
Buyers and enterprises●●●●● 4.3
Can I use this today? · rank 6 of 32 for this audience
Q1What it is●●●●● 5
Authenticated endpoints accept your API key in the X-API-Key request header.
base URL https://lium.io/api with a curl example and an OpenAPI 3.1 spec at lium.io/api/openapi.json, front page lists live pods with prices and a pip install lium.io quick start
Q2Who it is for●●●●● 5
Every pod is billed per second with no minimum.
price per GPU-hour for every GPU model from 0.03 to 5.88 dollars, renters, developers and AI agents named as the audience on the pricing and docs pages
Q3How it resists gaming or fails●●●●● 3
support is a Discord ticket form in the #contact-lium-support channel, support@lium.io on the docs welcome page, a broken pods page for renters, no status page (status.lium.io does not resolve), no SLA, no response time
Q4Identity and documentation●●●●● 4
API reference with OpenAPI spec and SDK and CLI references, support@lium.io on the docs welcome page, subnet_contact is not set on chain so no contact match is possible
Newcomers●●●●● 3.5
What is this and why does it matter? · rank 15 of 32 for this audience
Q1What it is●●●●● 4
a decentralized GPU rental marketplace built on Bittensor Subnet 51
on the docs welcome page one click from the front page, providers, renters and validators explained in plain words, the chain description is a slogan
Q2Who it is for●●●●● 2
Rent High-Performance GPUs at Unmatched Prices
is the only difference claim on own pages, a cost claim with no comparison against a centralized cloud, third-party pages make the up to 90 percent cheaper comparison
Q3How it resists gaming or fails●●●●● 4
live marketplace of 83 available pods with per pod reliability and uptime percentages and live prices, no dated benchmark, no usage figure with a date
Q4Identity and documentation●●●●● 4
subnet_name lium.io matches the domain, description is a slogan without a subject, url resolves without a scheme, the docs welcome page serves as the about and learn page
Provenance
How this rating came to be. Verified means the scorer fetched the page and the words are on it; secondhand means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
scored by agent b 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 40 live, 0 unreachable, 1 manual
Something wrong? Corrections of fact are applied as they arrive during the window; rating disputes are batched at its close. How to file one · GitHub issue · email.