The hive is the cook tier: owned GPUs that slow-cook the diabetic models, and owned storage that holds the data they're cooked from — and the records they'll never touch. Here's the hardware, and why each piece.
The cook fleet runs on NVIDIA RTX PRO 6000 Blackwell cards (sm_120, CUDA 13). Each card carries 97 GB of VRAM; a two-card rig gives 194 GB.
| Spec | Value | Why it matters for diabetic cooks |
|---|---|---|
| VRAM / card | 97 GB | 27B anchor + activations fit on one card |
| VRAM / rig (2×) | 194 GB | flagship cooks in one place, no sharding |
| Compute | sm_120, CUDA 13 | modern attention via SDPA; Triton JITs for the arch |
| Ownership | ours | marginal cost ≈ kWh, not cloud markup |
Honest note: Blackwell is new
silicon. Prebuilt flash_attn/xformers kernels don't ship sm_120 yet, so we cook
with torch-native SDPA attention. We caught that at canary, not mid-burn — and wrote it down. →
Crockpot Cook
Records live on Synology NAS boxes — the diabetic's own vault, on their own network.
| Role | What lives there | What never leaves |
|---|---|---|
| Home vault | records, meds, labs, foot-care logs, appointments (15-folder structure) | all raw PHI |
| Cook staging | open/deeded corpora, model weights, receipts | (no PHI ever staged for cooks) |
| Receipts ledger | hash-chained, non-PHI proofs of every action | — |
The hive cooks a model, proves it beats base, and the model flows down to the edge brain on the diabetic's desk. Receipts flow back up; PHI never does. That hand-off is the whole architecture. → The Edge