The challenge
Their protein-structure / MD pipeline targets TPUs (JAX/XLA). There is no Apple-Silicon path, the work was impossible on the device, full stop.
Understanding the workload
One recognized a TPU accelerator family, which exists only in the cloud, and bursted accordingly.
Best hardware for the job
Matched 8× Cloud TPU v5e, the family the workload actually needs.
Completion
One provisioned the TPU slice, ran the pipeline, returned results, and tore it down, credential, data, and bill stayed in the lab's own project.
What changed
Impossible → completes in ~92 min
Feasibility
~$14.79
Cost per run
Stayed in the lab's own cloud
Data control
The feeling we build for
It wasn't slow before. It was impossible. Now it's a 90-minute job.
Placement, hardware matching and completion are real system behavior; the figures are editable model inputs.