One workflow.
Every cluster and cloud.
Build the workflow. It provisions, routes and runs itself across HPC and any cloud. First results today.
target.fasta
1 target sequence
stagedligands.smi
12,842 ligands
stagedBuild Boltz-2 inputs
queuedBoltz-2 structure and affinity
94% gpuqueuedRank hits
queued
Sample run. Every step placed automatically.
The bottleneck is not the science. It is the setup.
Compute is not hard to buy any more. Getting from a set of credentials to a result still is. Teams receive a hostname and an allocation, and then spend weeks on environments, schedulers and data movement before the first job lands.
Once everything is finally running, research teams still spend 20 to 40% of their time managing infrastructure, debugging pipelines and handling deployments.
That is the tax this exists to remove. Not a faster GPU: the weeks before the GPU does anything, and the fraction of every week after that goes on keeping it working.
Four steps, none of them infrastructure
There is no cluster to configure, no scheduler to learn, and no container to build before you start.
- 01
Build the workflow
Drag stages onto the canvas and connect them. Dependencies resolve themselves. No DSL to learn and no scripting required, with a code editor there for the people who want one.
- 02
It routes itself
Each stage is profiled as it runs and matched to compute by cost, speed and availability, across HPC schedulers, AWS, Azure, GCP, Scaleway and your own hardware. AI uses that profile to size the next run, so jobs stop shipping oversized to be safe, or getting killed for being too tight.
- 03
Watch it run
Live task status, resource use and logs in one view, with results rendered as you go. You find out a long job is failing while it is still running, not the next morning.
- 04
Keep the result
Every run records its inputs, its environment and its outputs. Reproducibility is a property of the platform rather than something each person has to remember to maintain.
Where the money actually goes
Raw compute is close to a commodity and we do not try to undercut it on price. The spend that hurts a research team is somewhere else. Point at a piece of the bar to see where.
The same list, compared
| Where it shows up | On a normal cloud bill | With Temple Compute |
|---|---|---|
| Idle & reserved capacity | A node reserved between projects, billed while nothing runs on it. | Capacity scales to the workload and shuts down right after. |
| Oversized requests | Jobs padded to be safe, or killed for requests that were too tight. | Every run is profiled; AI sizes the next one from that history. |
| Serial wall-clock time | One pipeline run at a time, however many items are waiting. | Fan-out runs the whole pipeline per item in parallel. |
| A platform hire | Nobody on the team is a cloud or HPC engineer, so one gets hired. | Routing, sizing and monitoring are the platform's job, not a hire's. |
| Line items on an invoice | vCPU-hours, storage, licensing and egress, tallied at month-end. | One token unit, covering all of it, priced before the run starts. |
Wall-clock time, not just spend
Each item waits for the last one to finish. A hundred thousand molecules run one pipeline after another.
Each stage routes itself
Build the chain once. Every stage is matched to whatever is cheapest and fastest right now, across HPC schedulers and the major clouds.
See the pricing page for how tiers and tokens work, or how we compare to our competitors.
Real workflows, not a feature list
27 documented pipelines you can read end to end, from virtual screening to molecular dynamics setup. Each one shows which stage runs where and why.
horus-runtime is free and ungated
The engine under Temple Compute OS is open source. No waitlist, no early access. Read the code, run it yourself, or see exactly how it works.
Or go straight to the documentation.
Because the next discovery shouldn't wait for infrastructure
Ask for a beta account for launch pricing or if you have a workflow to run now.