Durable Python jobs.
Lower cloud costs.
Portal gives Python workloads durable execution across clouds and CPU/GPU architectures. Run long-running jobs on affordable spot compute, recover from interruptions, and switch hardware without starting over.
ClickTap to play Snake
Give this to your coding agent
set up https://portal.clusy.io
Fully application-side. Cross-architecture.
Portal moves the program itself, not the machine it runs on. Everything is captured from inside the Python process: the variables, the tensors, the open files, the exact point in the loop. That is why the next machine can have a different chip, a different kind of virtualization or a different cloud behind it, and the program simply carries on from the same step. There are no VM snapshots, no kernel modules and nothing to install in the hypervisor.
- Any architecture
- From x86_64 to arm64, from GPU to CPU, and back again.
- Any kind of machine
- VMs, containers and microVMs, on any provider.
- Verified on arrival
- Every tensor is checked against its digest when it arrives.
What we ran on Portal
In one night we ran five real workloads across AWS, Modal and Lambda: a 728 GB climate analysis, a simulation of a billion particles, every ship in US waters, the human proteome, and a job that changed clouds while it was running. Each one landed on the cheapest machine that fit it, and every number here comes from the run itself.
Read the case studies
Twenty years of hourly global temperature, all in memory267 GB of ERA5 loaded into one 728 GB array on a single spot machine, then analysed cell by cell.728 GBstate in memory47 minload and analysis71%saved
A billion-particle universeAn N-body simulation of cosmic structure: 1.07 billion particles on a 2048-cubed mesh, from z = 50 to today, on 64 vCPUs.1.07 bnparticles2 h 16 minwall time52%saved
Every ship in US waters, July to December 2024NOAA's AIS feed for half a year: 178 daily files and 66 GB, parsed into compact arrays and drawn at a hundredth of a degree.1.57 bnpositions80,593vessels50%saved
The human proteome, folded overnightESMFold, a 3-billion-parameter structure model, over 16,650 human proteins in eight GPU shards across two clouds.16,650proteins folded8GPU shards, two clouds27%under AWS list price
A job that changed clouds while it was runningA copy of the ships job, moved 100 seconds in: from an AWS spot machine to a Lambda GPU box and back, with its arrays, open loops and progress intact.411 Mpositions it held2moves, AWS to Lambda to AWS17 minend to end- Karpathy's autoresearchThe same run priced on eight providers, with AWS on demand as 100. On Portal it came to 22, 78% less than AWS.22%of the AWS price, on Portal62%the next cheapest, Lambda Labs179%the dearest, AzureAzure179%Modal163%Google Cloud160%AWS100%CoreWeave89%Nebius65%Lambda Labs62%Portal22%
Scalable enterprise solution.
Portal can run as a dedicated deployment inside your own cloud account, working with the clusters you already have as well as the clouds, and sized to the way your team works. A customized deployment can move up to 5 TB of state within a preemption notice, so even your largest runs come through a spot interruption intact.
Talk to us about a deployment- Private, and fast because of it
- A dedicated controller and checkpoint store inside your own cloud account, so credentials, checkpoints and records never leave it. Moves run at the speed of your own network.
- Your clusters, in the same view
- Enroll the machines you already own with a small agent that only dials out. Run on them or on your clouds from the same console, with leases, usage and clean-up in one place.
- Scheduling that saves compute
- Every run lands on the cheapest machine that fits. Policies move it when a better price appears, use spot where it is safe, and suspend what sits idle or runs past its budget.
- Teams, quotas and access
- Give people exactly the access they need, from viewer to owner. Set spending caps per connection and budgets per run, and check cloud quotas before a launch.
- Every action on the record
- Sign-ins, launches, moves, resumes and connection changes go into an append-only audit log that stays in your account. A coding tool only ever holds the scopes a person approved.