GPUs clock out. Your jobs clock in.
Nightshift is building a way to run open AI models on graphics cards that would otherwise sit idle. Each job gets a price before it starts, a spending cap it can’t pass, and a time card when it finishes.
No. EX-0417
| Quoted | in 0.30, out 0.60 per 1K tokens |
|---|---|
| Cap held | 0.840 cr |
| Clock in | 02:14:07 |
| Metered | 1,200 in, 312 out |
| Clock out | 02:14:11 |
| Charged | 0.547 cr |
| Released | 0.293 cr back to balance |
Example card. No job ran and nothing was charged.
- Website
- Live
- Inference pilot
- In development
- Paid access
- Not open yet
- Prices on this page
- Examples only
Five punches, from request to receipt.
Every job follows the same five steps, and each one leaves a line on its time card.
Find a free GPU
Your job goes to a machine whose owner has cleared it for outside work and that has a slot open. A chip that only looks quiet doesn't count.
Get a quote
You see the per-token rate and the most the job could cost. Rates move with how much spare room there is. Once you accept, that quote is fixed for this job.
Set a cap
You approve a maximum. That amount is held while the job runs, and the job can't spend past it.
Run and meter
The open model does the work. Input and output tokens are counted as the answer is written.
Settle and file
You're charged for what was measured, and the rest of the cap is released. The time card records the job, the usage and the cost.
A time card explains the bill. It doesn’t prove the answer is right. Lower prices depend on how much capacity is free, so they aren’t guaranteed.
Paid-for GPUs often wait.
In a two-month trace of a Microsoft training cluster, GPUs already assigned to jobs averaged 52% utilization [1]. That’s one cluster and one kind of work. It shows slack exists, not that the slack is for rent.
A day on one machine
- Owner’s own work
- 57%
- Nightshift jobs
- 21%
- Still idle
- 22%
Hatched gaps belong to owners who don’t share their GPUs, so they stay empty. Jobs leave a short margin before the owner’s next task. Illustration of a simulated day, not measured data.
Finding: Headroom is common.
Even GPUs that are already handed to a job often run well below capacity. [1]
What Nightshift does: Use only capacity that's offered.
A machine joins the pool when its owner allows outside work and a slot is free. Low activity on its own isn't permission.
Finding: Cheap capacity can vanish.
Serving language models on preemptible machines lowers cost, but those machines can be reclaimed at short notice, which slows responses. [2]
What Nightshift does: Price in the failures.
Rates follow how many slots are taken, above a floor that covers compute, operations and failed jobs. Quiet hardware doesn't automatically mean a discount.
Finding: Cold starts cost time.
Loading model weights is a big part of starting a model on demand. Faster loading helps, but it's one piece of the problem. [3]
What Nightshift does: Start with one warm model.
The pilot runs a single pinned model in one region, so it can stay loaded between jobs.
The pilot has to show acceptable answers, reliable requests and a lower total cost before we claim any of them.
- Jeon et al. Analysis of Large-Scale Multi-Tenant GPU Clusters for DNN Training Workloads, USENIX ATC 2019.
- Miao et al. SpotServe: Serving Generative Large Language Models on Preemptible Instances, ASPLOS 2024.
- Fu et al. ServerlessLLM: Low-Latency Serverless Inference for Large Language Models, OSDI 2024.
Work that can wait for a quiet hour.
Batch text jobs suit spare capacity best: lots of small requests, a clear way to check the output, and nobody waiting on a reply in real time. These are planned workflows, not customer results.
| Job | What it does | Usually run by |
|---|---|---|
| Catalog sorting | Puts product descriptions into the right category, checked against examples you've already reviewed. | Store builders |
| Field extraction | Turns loose text into fields an app can read, such as names, dates and amounts. Unknowns stay marked as unknown. | Developers |
| Ticket triage | Tags support messages by what the customer wants before they reach a queue. | Support teams |
| Review summaries | Writes short digests of long documents for a person to check before anyone relies on them. | Researchers |
| Archive tagging | Labels a content library so people can actually search it. | Product teams |
| Bill matching | Lines up time cards with your own request log, so every charge has a matching job. | Finance teams |
Four calls, one time card.
The planned API keeps every step explicit. Ask for a quote, authorize a cap from your wallet, run the request against that quote, then fetch the time card.
- POST
/v1/quotesRates, output limit and maximum charge - POST
/v1/authorizationsYour wallet signs the cap - POST
/v1/chat/completionsRuns against the accepted quote - GET
/v1/timecards/{id}Usage, charge and released amount
Planned API. It isn’t live and the names may change. In the first pilot your wallet authorizes each request, so there’s no separate Nightshift API key.
{
"timecard": "ex_0417",
"model": "open-weight-example@rev",
"quote": { "input_per_1k": 0.30, "output_per_1k": 0.60 },
"cap": 0.840,
"usage": { "input_tokens": 1200, "output_tokens": 312 },
"charged": 0.547,
"released": 0.293,
"clock_in": "02:14:07",
"clock_out": "02:14:11",
"example": true
}Before you clock in.
Straight answers about what exists today and what is still a plan.
What is Nightshift?
A project building AI inference on spare GPU capacity. It's aimed at text jobs like sorting, pulling out fields and summarizing, run on open models. This site explains the plan. The service itself isn't live yet.
Why would spare capacity cost less?
Rates are planned to follow how many serving slots are free, above a floor that covers real costs. When more room is free, a quote can come in lower. It can also come in higher, and no discount is promised. You always see the rate before you approve anything.
Can a job cost more than my cap?
No. Each quote fixes the model, the rates and an output limit, and you approve a maximum charge. The job stops at that limit, so the charge can't go past it.
What happens to the part of my cap I don't use?
It's released when the job finishes. You pay only for the tokens that were measured. In the planned pilot, a request that fails isn't billed.
Do I need an API key?
Not in the first pilot. Your wallet signs each authorization, so there's no separate Nightshift key to manage. Your wallet's own key still matters, so keep it safe. Delegated budgets come later.
Which models will it run?
The pilot starts small: one approved open-weight model at a pinned revision, in one region, with plain text requests and no streaming. That model's license, quality and serving cost get checked before anyone pays. Streaming, tools and more models are later work.
What does a time card prove?
That Nightshift signed a specific billing record and included it in a published batch. It doesn't independently prove which GPU did the work, that the metering was accurate, or that the answer is correct.
Can I withdraw funds I haven't used?
The plan returns available balance to the same wallet, token and network it came from. Money held for a running job stays held until that job settles.
What is $SHIFT?
$SHIFT is the project's token. This site makes no promise about its price or returns. Before you buy anything, check the contract address against the one shown here and on our official X account.
Off-peak hours start soon.
Read the plan, then follow along while the pilot comes together.