GPU access is expensive and fragile
AWS and GCP GPU instances cost $1–30/hr with 10-minute cold starts. You're paying for idle time while CUDA installs.
GPU compute,instantly.
Your ML workflow deserves better than SSH tunnels and dead notebooks. Provision A100s in 30 seconds. Sync files in milliseconds. Run jobs that survive your laptop closing.
30s cold startT4 · A100 · H100CLI-first
The problem
ML infrastructure wasn't built for developers
You've felt this. The GPU quota request. The notebook that worked yesterday. The training job that died when your WiFi dropped.
Your laptop can't handle ML
16GB RAM doesn't train transformers. You need 64GB+ and GPUs — provisioned on demand, not bought upfront.
Environment setup is painful
Every new machine means reinstalling CUDA, PyTorch, drivers. Days lost to dependency hell before you write a line of code.
Collaboration is broken
Sharing a Jupyter server means SSH tunnels or JupyterHub configs. Each teammate needs isolated environments, not shared passwords.
Long-running jobs die
SSH disconnects kill your training run at hour 47. Detached jobs keep running — results are waiting when you come back.
"Works on my machine"
Notebook execution isn't reproducible across machines. Sandboxed environments with pinned runtimes eliminate the guesswork.
Features
Everything you need to ship ML
One CLI. GPU environments. File sync. Detached jobs. No YAML required.
Provisioned dev environments
Jupyter and VS Code in the cloud with instant GPU access. T4, A100, or H100 — ready in ~30 seconds.
Hot file sync
<50ms latency syncing local files to cloud workspaces. Edit locally, execute on GPU — no git push dance.
Detached notebook jobs
Run .ipynb files as background jobs. Stream logs, collect artifacts — training survives your commute.
Persistent storage
PVC-backed workspaces with checkpoint and restore. Your datasets and models persist across sessions.
Scale to zero
Environments auto-suspend after 30min idle. Resume in <10 seconds. Pay nothing while you sleep.
Credit-based billing
Per-second compute billing. No monthly GPU reservations. See exactly what each job cost before you scale.
How it works
Three commands to production
No Terraform. No Kubernetes manifests. Just a CLI that speaks your language.
01
Provision
Spin up a GPU environment from your terminal. Jupyter or VS Code, pre-configured with CUDA and PyTorch.
$ flexx up --project image-classifier --gpu a100 Provisioning environment... ✓ Environment ready in 28s Jupyter: https://env.flexx.dev/abc123
02
Sync & work
Stream local files to your cloud workspace with sub-50ms latency. Edit in your IDE, run on GPU.
$ flexx sync . ✓ Synced 847 files (42ms avg latency)
03
Run & collect
Execute notebooks as detached jobs. Stream logs, download artifacts. Suspend when done — resume in seconds.
$ flexx run train.ipynb --gpu a100 Job started: job_7x9k2m $ flexx artifacts job_7x9k2m --download Job completed · 2.4 GPU-hours · $1.87
Pricing
Pay for compute, not idle time
Per-second billing. Scale to zero. No surprise bills from forgotten instances.
Planned for launch — join early access to lock in these rates
Free
$0forever
$10 credits included
Explore flexx with no credit card required.
- T4 GPU access
- 1 active environment
- 5GB persistent storage
- Community support
- Per-second billing
Most popular
Pro
$50/month
+ pay-as-you-go credits
For individual ML engineers shipping real work.
- T4, A100, H100 access
- 5 active environments
- 100GB persistent storage
- Detached notebook jobs
- Priority cold starts
- Email support
Team
Custom
Volume discounts available
For teams with shared billing and isolated environments.
- Unlimited environments
- Shared project storage
- Team billing & invoicing
- SSO & RBAC
- Dedicated support
- Custom GPU quotas
GPU compute rates
T4
$0.35/hr
$0.000097/s
A100 (40GB)
$2.10/hr
$0.000583/s
H100
$4.50/hr
$0.00125/s
Billed per second. Environments auto-suspend after 30min idle — storage billed separately at $0.10/GB/mo.
Early Access
Be the first to train.
We're building flexx for engineers who ship. Sign up for early access — if there's demand, we ship.