Tested on a real run · 4 Nov 2026 · RTX 3060 Laptop

Your training runshouldn't die withthe GPU.

Simulation · not livestep 112 / 400

Ready on Provider A

Advance manually or stop the simulated provider.

Spare GPUs are cheap, but they vanish halfway and you can't see what they did. GPUChain moves your run when a GPU dies, refuses work that doesn't hold up, and only pays for the steps that made it into your model.

Watch the recorded run Break it yourself

Measured on the 4 Nov run, on a real RTX 3060. One run, one seed, local chain. See how

to notice the provider stopped: 0.6 s
until it was training again elsewhere: 65 s
junk checkpoint, refused and unpaid: 24.8%
USDC paid for accepted steps. 4.10 refunded: 0.15 + 0.75

Run timeline

00:13Training starts on Provider A00:28We pull the plug on Provider A00:29Failure noticed at step 15801:09Checkpoint at step 150 accepted by both verifiers01:33Training resumes on Provider P02:03Provider P's checkpoint fails the quality test, refused02:30Training resumes on Provider B03:01Checkpoint at step 200 accepted by both verifiers03:56Checkpoint at step 400 accepted by both verifiers03:56Final model accepted04:08Escrow pays out on-chain

Model: DistilBERT + LoRA
Dataset: AG News
GPU: RTX 3060 Laptop
Chain: Anvil (local)
Money: Test USDC

01What goes wrong today

A provider disappears
The run is interrupted.
A replacement provider resumes from a trusted checkpoint
Someone sends back junk
You find out much later, if ever
A test you both agreed on refuses it before it counts
Paying for it
By the hour, whatever happened
By the step, and only for steps that made it into the model
The job times out
A refund can take manual follow-up
The escrow permits a refund after its deadline and grace period

When	Renting a GPU today	With GPUChain
A provider disappears	The run is interrupted.	A replacement provider resumes from a trusted checkpoint
Someone sends back junk	You find out much later, if ever	A test you both agreed on refuses it before it counts
Paying for it	By the hour, whatever happened	By the step, and only for steps that made it into the model
The job times out	A refund can take manual follow-up	The escrow permits a refund after its deadline and grace period

02Try it

Go on, try to break it.

A simulation that runs in your browser, with made-up numbers. Start a run, then stop a provider or make it cheat. Watch where the work goes and who gets paid. The real recording is in section 04.

Ready when you are

step 000 / 200

Solid: accepted and paidStriped: not checked yetLines: checkpoints every 25 steps

Nothing yet. Press start, then try to break it.

Where the 1.00 USDC goes

Paid per accepted step. Refused or lost work earns nothing.

Provider A0.002 / step0.000
Provider B0.003 / step0.000
Provider C0.0025 / step0.000
Back to you1.000

Steps lost to kills: 0
Steps refused: 0

A simplified simulation running in your browser, with made-up speeds and prices. The real run is further down.

03How it works

From failure
to payout.

Nobody has to step in. You can watch every step happen on the dashboard.

01
Detect
0.6 s
A provider stops answering. The coordinator marks it lost and fences off its worker.
02
Recover
65 s
A replacement provider resumes from a trusted checkpoint, with micro-batches resized to fit its memory. In this run the providers shared one GPU.
03
Verify
2 of 2
In this run two verifiers evaluated selected checkpoints against the last accepted one. A failed test meant that work did not count.
04
Pay
per step
Escrow pays each provider for the steps it put into the accepted model and refunds the rest.

04The real run

On 4 Nov we stopped a training process halfway through, on purpose. Then we made a replacement provider cheat.

to notice the provider had stopped
until it was training again (41 s of that checking the checkpoint)
steps lost, out of 400
accuracy of the junk checkpoint. Refused. The one before was 89.2%
final accuracy, on 2,000 test examples

Training loss, every step, straight from the recording

Here's what that looked like.

Lower is better. Hover or drag across it to see any step.

Provider AProvider P (fault injected)Provider B

The budget was 5 USDC. Provider A got 0.15 and Provider B 0.75 for the steps that made it into the model. Provider P got nothing, and 4.10 went back to the client. It's one run with one seed on a local chain, so think of it as a demonstration, not a benchmark.

05Honestly

What works today

Kill a provider; a replacement can resume
Bad checkpoints get refused
Escrow pays per accepted step
Timeout refund after the deadline and grace period
Live dashboard and replay

What we haven't done yet

A second physical GPU (one GPU with memory caps for now)
Base Sepolia (a local chain for now)
Independent verifiers (our team runs them for now)
An audit, or real money

06Who made this

We're a team of four. We kept asking one question: if you rent a GPU from someone you've never met and it dies halfway, what happens to your run, and to your money? This is our answer, built for a hackathon and tested on a real GPU.

PGPiyush GargLead · coordinator and integrationLead · coordinator and integration
VPVaibhav ParmarML runtimeML runtime
AKArvind KushwahaContracts and verifiersContracts and verifiers
PSPrateek SharmaDashboard, design and demoDashboard, design and demo

Stop a provider. Watch the run carry on.

Watch the recorded run Open the live dashboard

GPUChain

Open dashboard

Tested on a real run · 4 Nov 2026 · RTX 3060 Laptop

Your training runshouldn't die withthe GPU.

Simulation · not livestep 112 / 400

Ready on Provider A

Advance manually or stop the simulated provider.

Watch the recorded run Break it yourself

Measured on the 4 Nov run, on a real RTX 3060. One run, one seed, local chain. See how

to notice the provider stopped: 0.6 s
until it was training again elsewhere: 65 s
junk checkpoint, refused and unpaid: 24.8%
USDC paid for accepted steps. 4.10 refunded: 0.15 + 0.75

Run timeline

Model: DistilBERT + LoRA
Dataset: AG News
GPU: RTX 3060 Laptop
Chain: Anvil (local)
Money: Test USDC

01What goes wrong today

A provider disappears
The run is interrupted.
A replacement provider resumes from a trusted checkpoint
Someone sends back junk
You find out much later, if ever
A test you both agreed on refuses it before it counts
Paying for it
By the hour, whatever happened
By the step, and only for steps that made it into the model
The job times out
A refund can take manual follow-up
The escrow permits a refund after its deadline and grace period

When	Renting a GPU today	With GPUChain
A provider disappears	The run is interrupted.	A replacement provider resumes from a trusted checkpoint
Someone sends back junk	You find out much later, if ever	A test you both agreed on refuses it before it counts
Paying for it	By the hour, whatever happened	By the step, and only for steps that made it into the model
The job times out	A refund can take manual follow-up	The escrow permits a refund after its deadline and grace period

02Try it

Go on, try to break it.

A simulation that runs in your browser, with made-up numbers. Start a run, then stop a provider or make it cheat. Watch where the work goes and who gets paid. The real recording is in section 04.

Ready when you are

step 000 / 200

Solid: accepted and paidStriped: not checked yetLines: checkpoints every 25 steps

Nothing yet. Press start, then try to break it.

Where the 1.00 USDC goes

Paid per accepted step. Refused or lost work earns nothing.

Provider A0.002 / step0.000
Provider B0.003 / step0.000
Provider C0.0025 / step0.000
Back to you1.000

Steps lost to kills: 0
Steps refused: 0

A simplified simulation running in your browser, with made-up speeds and prices. The real run is further down.

03How it works

From failure
to payout.

Nobody has to step in. You can watch every step happen on the dashboard.

01
Detect
0.6 s
A provider stops answering. The coordinator marks it lost and fences off its worker.
02
Recover
65 s
A replacement provider resumes from a trusted checkpoint, with micro-batches resized to fit its memory. In this run the providers shared one GPU.
03
Verify
2 of 2
In this run two verifiers evaluated selected checkpoints against the last accepted one. A failed test meant that work did not count.
04
Pay
per step
Escrow pays each provider for the steps it put into the accepted model and refunds the rest.

04The real run

On 4 Nov we stopped a training process halfway through, on purpose. Then we made a replacement provider cheat.

to notice the provider had stopped
until it was training again (41 s of that checking the checkpoint)
steps lost, out of 400
accuracy of the junk checkpoint. Refused. The one before was 89.2%
final accuracy, on 2,000 test examples

Training loss, every step, straight from the recording

Here's what that looked like.

Lower is better. Hover or drag across it to see any step.

Provider AProvider P (fault injected)Provider B

05Honestly

What works today

Kill a provider; a replacement can resume
Bad checkpoints get refused
Escrow pays per accepted step
Timeout refund after the deadline and grace period
Live dashboard and replay

What we haven't done yet

A second physical GPU (one GPU with memory caps for now)
Base Sepolia (a local chain for now)
Independent verifiers (our team runs them for now)
An audit, or real money

06Who made this

PGPiyush GargLead · coordinator and integrationLead · coordinator and integration
VPVaibhav ParmarML runtimeML runtime
AKArvind KushwahaContracts and verifiersContracts and verifiers
PSPrateek SharmaDashboard, design and demoDashboard, design and demo

Stop a provider. Watch the run carry on.

Watch the recorded run Open the live dashboard

Your training runshouldn't die withthe GPU.

Go on, try to break it.

From failureto payout.

Detect

Recover

Verify

Pay

On 4 Nov we stopped a training process halfway through, on purpose. Then we made a replacement provider cheat.

Here's what that looked like.

What works today

What we haven't done yet

Stop a provider. Watch the run carry on.

Your training runshouldn't die withthe GPU.

Go on, try to break it.

From failureto payout.

Detect

Recover

Verify

Pay

On 4 Nov we stopped a training process halfway through, on purpose. Then we made a replacement provider cheat.

Here's what that looked like.

What works today

What we haven't done yet

Stop a provider. Watch the run carry on.

From failure
to payout.

From failure
to payout.