Consumer GPUs,
an invite away.
Keep your existing PyTorch · Jupyter · OpenAI-compatible API — just swap the endpoint and run. From $0.30/hr (by GPU and VRAM), billed per second. It's a pool of personal PCs, not a data center, so there are limits — and we've listed those limits below as they are.
Review, then an invite → 60 minutes free with a trial credit · No spam
"Is it cheap? Is it instant? Is it safe?"
We answer all three without exaggeration — the good and what we can't do, as they are.
Cheap — cheaper than big clouds
Comparable GPU instances at big cloud providers run $0.5~$1/hr, while we start at $0.30 (consumer GPUs). Once invited, you start right away with no foreign card, and since it's per-GPU hourly and billed per second, short runs cost only that much.
Image batch bills only QC-passed outputs (failed ones aren't charged)
Instant — fast with a prewarm pool
We keep workers running ahead of time to cut cold starts. GPU dispatch itself is fast, but a real API round-trip — including the network — is hundreds of ms.
dispatch p99 5.94ms (process-internal, n=300, excludes network)
Data — don't upload anything sensitive
Isolation is process-level — it runs on a personal PC, so it's not fully separated from the host. That's why you should not upload sensitive data, personal information, or proprietary models. It's fine for public and experimental workloads.
Confidential computing (CC) needs H100-class hardware · currently unsupported
Up and running in 3 minutes
Keep your existing PyTorch / Jupyter workflow. One line of API and you're done.
Get an API key
Sign up and issue an API key from the dashboard. Once your invite is approved, a trial credit is granted.
Run your workload
Keep your existing PyTorch / OpenAI-compatible SDK. Just point the endpoint at TechNode.
Pay only for what you use
Per-second billing. The prewarm pool is always on standby, so there's no cold-start cost.
Your existing tools, as is
No new SDK to learn. Use PyTorch · Jupyter · SSH · OpenAI-compatible API as is.
Frequently asked questions
If a node suddenly dies, does my job get cut off too?
The session automatically reconnects to another node — but the in-progress state held in memory (kernel variables) isn't preserved, so you'll need to re-run the notebook (live migration of GPU state isn't possible on consumer GPUs). For long training runs, save checkpoints yourself. The reassignment mechanism itself measured 0 failures over 1.25 million cycles in a 19.6-hour soak test (this is the stability of the yield mechanism, not "availability").
Can the host (GPU provider) see my models and data?
In principle, yes. Because isolation is process-level, the host can access the workload with admin rights on their own PC (fully blocking this would require H100-class confidential computing, which isn't supported on consumer GPUs). So don't upload sensitive data, personal information, or proprietary models. That said, the current beta pool runs on PCs the operator runs directly (RTX 4060 Ti · 4070 SUPER), so there's no exposure to any third-party host — it's fine for public and experimental workloads.
How does pricing work?
RTX 4070 SUPER (12GB) is $0.30/hr, RTX 4060 Ti (16GB) is $0.37 — tiered by VRAM (billed per second). Once invited, you first try it free for 60 minutes with a trial credit, then top up when you want more. We'll walk you through the payment process when you request a top-up (automatic card payments are pending Toss merchant review).
Do I have to rewrite my existing PyTorch code?
No. PyTorch · Jupyter · SSH · OpenAI-compatible API all work as is. Just change the endpoint URL to TechNode.
Apply for the beta now
Rent from the GPU pool right away · for more, top up with a paid plan. If you just want beta updates, leave your email.