ergonomic-vastai
Run serious AI models, privately,
for cents an hour.
You need AI that doesn't send your data to anyone — for research, for sensitive documents, for work that can't leave your hands. Your laptop can't run the good models, and cloud AI platforms charge monthly whether you use them or not.
ergonomic-vastai fixes this. It rents the right GPU from the cheapest market, installs the fastest serving stack for that card, connects the model to your tools, and shuts everything down the moment you're done. You pay for hours, not months.
What that looks like
One command. Cost shown before anything is rented. Auto-shutoff so a forgotten session can never become a bill.
$ ergonomic-vastai up qwen3.8-27b
host RTX 6000Ada · 49GB — cheapest card that fits, of 64 offers
cost $0.52/hr · hard cap $2.07 · auto-off after 30 min idle
ready ~9 min
● ready your private endpoint: http://127.0.0.1:57778/v1
$ ergonomic-vastai down
destroyed, verified gone — total: $0.08
What we do for you
Pick the hardwareWe shop the GPU spot market live and rank
offers by real cost — including download time you'd otherwise pay for.
Tune the enginevLLM, SGLang, llama.cpp, MLX: we choose and
configure the fastest one for your model and card, from measurements, not guesswork.
Guard your walletCost shown before renting, a hard time
limit on every run, idle auto-shutoff on the machine itself, and teardown that waits for
the provider to confirm the machine is gone before it stops watching.
Keep it privateThe endpoint answers only on the machine's
own loopback; the single way in is your ssh tunnel. We never see your prompts, because
nothing transits our servers. Engine telemetry is switched off and request logging stays
off, by configuration we ship rather than promise.
Who can see what, plainly: a rented market GPU belongs to its host. That is fine for open models and work you could run on any cloud. For data that must not touch a stranger's machine, the same command targets hardware you name (--trust own), and everything above still holds. We publish this because you should pick the machine to match the data, not discover the difference later.
The economics
A 27-billion-parameter model: $0.52/hour.
An evening of experiments: about $2.
A month of daily research use: $15–40, only the hours you run.
Managed AI clouds charge 2.5–4× for the same GPU. Subscriptions charge
you every month either way. We add 10% to the market price of the card.
Who it's for
Researchers & academicsSensitive data, IRB constraints,
grant budgets. Run open models on hardware you control, spend only what the work
needs.
Anyone whose laptop hit the wallYou tried running models
locally and met the 8GB reality. Rent a 49GB card for the evening instead of buying a
$5,000 one.
Pricing
Do it yourself
$0
open-source tool, your own GPU-market account
The full tool, free forever. You pay the GPU market directly; ergonomic-vastai adds nothing.
Get notified at launchManaged
market +10%
founding members: $20/mo, locked
No accounts to create, no market to learn. We carry the hardware relationships and the tuning; you run one command and pay one bill.
Become a founding memberRented market GPUs are right for open models and shareable-risk data. For data that can't touch anyone's hardware but yours, we ship the same tuned stack as inspectable packages for machines you own — talk to us.