[price]
Hardware that is already paid for
A datacenter has to buy the card, house it, cool it and earn a margin on top. A gaming PC at 3 a.m. only has to beat its own electricity bill.
class B from $0.22 per gpu hour
// decentralized, verifiable compute
NIGHTSHIFT
Your computer works for you, even when you don't use it.
Lend the hours your graphics card sits idle. People pay by the second to run AI models on it, several machines check each other's answers, and you are paid in $SHIFT.
// the idea
Every computer has a night shift.
You bought a graphics card that can run an AI model. You use it a few hours a day. The rest of the time it is the most expensive paperweight in the room.
Somewhere else, someone is renting that exact kind of card by the hour from a datacenter, and waiting in a queue for it.
Nightshift connects the two. Your machine takes jobs when you are away, hands them off the second you sit down, and is paid for every second it computed.
No machine is trusted on its word. Every result is computed more than once and compared before anyone is paid.
24 hours: u you, # booked (5.4h), - on offer, unbooked
per day
$1.05
per 30 days
$31.49
An estimate from the launch rate card, paid in $SHIFT, with 90% of each payment going to the machines. Not a promise: what you earn depends on how much work the network has, and electricity is on you.
// try it
Run a task. See who ran it.
Pick a job and send it across the mesh. Watch it get cut up, computed three times over, checked, and paid. The receipt at the end names every machine.
prompt
Summarize the Bitcoin whitepaper in three sentences.
- model
- llama-3.1-8b
- split
- 2 halves, mapped then reduced
- replicas
- 3 machines per shard, majority hash wins
- needs
- 8 GB of memory or more
[------------------------] 0.00s
[.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.][.]
hover a machine to see who it is
- [.] idle
- [>] assigned
- [#] computing
- [=] returned
- [$] paid
- [z] handed off
- [x] mismatch
> tail -f job.log
waiting for a job_
> cat output
the result lands here once a majority of replicas agree on its hash.
// verification
Don't trust a stranger's GPU. Check it.
A network of anonymous machines is only useful if a wrong answer cannot get through quietly. Four rules make sure of that.
- 01
Every shard runs more than once
Each piece of a job goes to 3 machines that do not know each other. The result only counts when a majority return the same hash.
- 02
Machines have money on the line
A machine posts $SHIFT before it gets work. Vote against the majority and part of that stake is gone, along with the pay for the job.
- 03
Owners come first, jobs survive
The moment an owner touches the keyboard, the job checkpoints and moves to another machine. You pay for seconds computed, never for the interruption.
- 04
The receipt names the machines
Every job ends in a public receipt: which machines ran which shard, what each one returned, and what each one was paid.
To pass a wrong result, liars must land on 2 of the 3 randomly drawn machines.
Chance a wrong result is accepted
2.8%
about 1 shard in 36, at 3x the compute.
Binomial tail, assuming machines are drawn uniformly at random. Every replica that votes against the majority loses stake, so each failed attempt costs the attacker money.
// how it works
Two sides. Three steps each.
- [1]
Scan
See what your machine is worth to the network: card, memory, class.
- [2]
Set your hours
Choose when it may work. Touch the keyboard and it stops within a second.
- [3]
Get paid per second
Each second your card computes is paid in $SHIFT, straight to your wallet.
- [1]
Pick a model
Open models that fit on consumer cards: language, speech, images, embeddings.
- [2]
Send the job
It is cut into shards and sent to several machines at once.
- [3]
Read the receipt
Get the result, the hash every replica agreed on, and the list of machines that did it.
// against the cloud
The bar is a datacenter. Here is how we clear it.
Renting from a big provider is cheap, dependable and simple. A network of home machines has to answer on all three, or it is a toy.
[reliability]
No machine is trusted, so none is critical
Home machines disappear without warning. Nightshift assumes they will: every shard is replicated and checkpointed, and a lost replica is replaced mid-run.
3 replicas per shard by default
[simplicity]
One request in, one receipt out
No instances to size, no images to build, no idle bill. You send a job and you are charged for the seconds it computed.
billed by the second
// the token
$SHIFT
The unit the network runs on. Jobs are paid in it, machines earn it, and machines stake it to be allowed to work.
Machines that did the work90%
Verifier pool5%
Burned5%
Applied to every job, on settlement.
// questions
Asked before you ask.
[?]Will this slow my computer down?
Not while you use it. Work only starts when the machine has been idle for the delay you set, and it stops the moment you come back. The job is handed to another machine.
[?]Can the machine owner see my data?
Assume yes. Jobs run in a sandbox, but the owner controls the hardware. Do not send secrets or personal data. Nightshift is built for open models and public or low-sensitivity inputs.
[?]How can results be checked if AI models are random?
Jobs run with a fixed seed and deterministic sampling, so honest machines produce identical bytes and identical hashes. A different hash means a different computation.
[?]What stops a machine from returning garbage?
It would need to land in the majority of the replicas for a shard, by chance, with other machines returning the same garbage. When it fails, it is not paid and its $SHIFT stake is cut.
[?]Why is there a token?
It is the unit the network settles in. Buyers of compute pay in $SHIFT, machines are paid in $SHIFT, and machines stake $SHIFT so that cheating has a price.
[?]What can I run?
Open-weight models that fit on consumer cards: small and mid-size language models, speech to text, image generation and embeddings. Frontier-size models that need a rack are out of scope.
Put your machine on the night shift.
It takes one scan to see what your card is worth to the network. It takes one click to see the network do a job.