Skip to content

Hardware & Elastic Burst

"Elastic burst," not "all on one box." The reference target is a single NVIDIA DGX Spark (GB10: ARM Grace + Blackwell). Most of the factory runs there. Heavy or ARM-incompatible models run on remote GPUs on demand, over a private mesh — and the site says so wherever that's the case, because a "one box" claim that quietly needs a cluster is exactly the gap a skeptic turns into a headline.

One box, elastic burst — local DGX Spark bursting heavy models to remote GPU over a private mesh

Illustrative diagram. Patient data stays on the local box; only derived, non-identifying work bursts — with one stated exception (GPU genomics alignment) noted below.

What runs where

Runs locally on the DGX Spark Bursts to remote GPU (on demand)
Orchestration, the capability registry, the MCP tool-surface, the workflow composer Frontier co-folding (Chai-1), RFdiffusion, Evo 2 (and gated Chai-2 binder design)
Governance gates + 21 CFR Part 11-style lineage x86-only CUDA containers (e.g. Parabricks, several BioNeMo NIMs)
The RAG intelligence agents (vector DB + LLM), multi-omics join, DuckDB Anything ARM-incompatible or too heavy for one box
All patient data Only derived, non-identifying work — never raw PHI

Two reasons a model bursts: raw compute (frontier models exceed one box), and architecture (the Spark is aarch64, so x86-only CUDA containers must run on a remote x86 GPU regardless of raw power).

The one exception, stated plainly: GPU genomics alignment (Parabricks, x86-only) needs the raw sequencing reads, so that step is the single case where identifying data goes off-box. It must run in a dedicated, secure remote environment; the reference deployment demonstrates it on the public HG002 sample. No other step sends identifying patient data to a burst GPU.

Privacy is architectural

All patient data stays on the local box. Only derived, non-identifying assets are ever sent to a burst GPU — never raw PHI. This is a design constraint, not a policy note; see Governance & Lineage.

Note

"Elastic burst" is a hard honesty invariant: no page implies everything runs locally, and no burst sends raw patient data off the box.