Your campaign, run on our node, handed back reproducible.

• You send a job spec — targets, compound library, what question the run has to answer.
• We pre-register the verdict criteria before any scoring happens, so the conclusion cannot drift to fit the numbers.
• We run it and monitor it, including the parts that fail and need re-running.
• You get results plus a reproducible report — versioned engine, inputs, outputs, and the code path taken.

This is a managed service, not rented hardware. You are buying an operated campaign and someone accountable for whether the numbers mean anything.

• Screening campaigns across one or several targets.
• Calibration runs, powered properly, on your measured data.
• Blinded holdouts and negative-control panels — the runs that decide whether a result is real.
• Pocket-knockout and structure-dependence probes.
• Adapter fine-tuning where a model needs to learn your chemistry.

A DGX Spark node with 128 GB of unified memory, and access through Catalytic AI to a Blackwell workstation with 96 GB of VRAM at 1,792 GB/s.

Being direct about the limits, because you should hear it from us rather than discover it: this is a single node, not a cluster, and we do not offer an uptime guarantee on it. That is exactly why access is managed rather than self-serve — we absorb the operational reality instead of selling you a promise about it. Campaigns are scheduled against real capacity, and if a deadline needs more iron than we have, we will say so before you commit.

• An SSH login. This is not infrastructure rental.
• An SLA. See above — we will not sell you a number we cannot hold.
• A result dressed up beyond what it supports. If a campaign comes back inconclusive, the report will say inconclusive.