The Sovereign Lab

10 towers. 130 cores. 188 GB VRAM. 125 TB storage. $15K. No cloud.

10 towers. 130 cores. 188 GB VRAM. 125 TB storage. $15K. No cloud.


Hardware

The Cluster

NodeCPUGPURAMNVMeRole
northGatei9-14900KRTX 5090 (32 GB)192 GB DDR5~8 TBFlagship compute
ironGatei9-14900KRTX 4070 (12 GB)96 GB DDR53.6 TBNUCLEUS deploy, validation
southGateRyzen 5800X3DRTX 4060 (8 GB) + swappable128 GB DDR4~5 TBHeavy compute
strandGateDual EPYC 7452 (64c)RTX 3090 + RX 6950 XT256 GB ECC~20 TBBioinformatics, multi-GPU
biomeGateTR 3970X (32c)RTX 5060 + Titan V + K80256 GB DDR4~5 TBHBM2 test bench
eastGatei9-12900RTX 4070 + Akida NPU32 GB DDR52 TBUtility, neuromorphic
westGatei7-4771RTX 2070 Super32 GB DDR32 TBCold storage / NAS
swiftGateRyzen 5800XRTX 3070 FE64 GB DDR4~2 TBMobile compute
flockGatei9-13900KRTX 3070 Ti64 GB DDR52 TBMesh compute
kinGatei7-6700KRTX 307032 GB DDR4~1 TBStaging

Plus 4 SFF nodes (1 GMKtec NucBox M6, 3 Intel NUCs).

Aggregates

MetricValue
CPU cores130+
GPU VRAM (installed)~188 GB
GPU VRAM (float pool)+68 GB (2x 3090, 2x Titan V, 2x MI50)
HBM256 GB (Titan V + MI50)
System RAM~1.2 TB
NVMe~49 TB
HDD (ZFS)~76 TB
NPUs3x BrainChip Akida AKD1000
HSMs4x SoloKey FIDO2
Total investment~$15,000

Acquisition Strategy

Most bought used. Prices paid:

ComponentSourcePrice
Tesla K80eBay$50
Titan V (x2)eBay$400 each
RTX 3090 (x2)FB Marketplace$700 each
Dual EPYC workstationSurplus~$1,000
10G NIC (Mellanox CX-3, x4)eBay$15-25 each
10G switch (MikroTik CRS305)Amazon$130

What Sovereignty Means (Testable)

Four claims. Each verifiable.

You own the hardware. The CPUs, GPUs, and drives are physical objects in a physical room. No API endpoint can be deprecated. No terms of service can change. No pricing tier can be adjusted. If the internet goes down, the science still runs.

You own the data. Every dataset fetched from NCBI, UniProt, or KEGG is BLAKE3-hashed at download time and stored locally. The fetch is a one-time event. The data doesn’t expire. If the upstream source changes an assembly, your local copy preserves the version you validated against, and the provenance record shows both states.

You own the compute. No metered API calls. No GPU-hour billing. The RTX 4070 runs Vulkan f64 GPU compute through DF64 double-float emulation — the same precision NVIDIA reserves for datacenter cards, unlocked through a sovereign shader pipeline (coralReef) that bypasses CUDA.

You own the provenance. Every computation produces a DAG entry (rhizoCrypt), committed to a permanent ledger (loamSpine), attributed via Ed25519-signed braids (sweetGrass). The chain is cryptographic. You can verify any result without trusting anyone.


Dependency Map

40+ external dependencies across 7 clusters. Honest assessment:

Cloudflare (highest priority replacement)

DNS, TLS termination, tunnel for external access. Replacement path specified: BearDog TLS (ChaCha20-Poly1305 with ACME), Songbird NAT traversal, self-hosted authoritative DNS. Baselines capturing hourly.

GitHub (longest pole)

Source hosting, CI, binary releases, Pages. Forgejo installed as calibration instrument. 74 workflow files to port.

Package Registries

crates.io, PyPI, Conda. Low urgency. Vendor escape hatch exists (cargo vendor, pip download, conda pack).

AI APIs

Anthropic, OpenAI. Optional. Ollama works locally. Long-term path is sovereign inference through barraCuda WGSL compute.

Science Data APIs

NCBI, UniProt, KEGG. Irreplaceable external data sources, but once fetched, data is local forever. Not a service dependency — a data dependency with caching.

Internal Primal Gaps

5 of 6 resolved by upstream Phase 60. MethodGate enforced on 10/13 services. Ionic token auth live. Resource envelopes enforced.

Irreducible Externals

Domain registrar. Linux kernel. NVIDIA GPU drivers. Let’s Encrypt certificate chain. $5/month VPS for NAT relay. Accepted constraints.

Summary: ~20% sovereign by service count, ~80% by criticality. Everything touching the science (compute, data, provenance, attribution) is fully sovereign. What remains external is infrastructure plumbing.


The ABG Model

An external bioinformatics research group connects through JupyterHub at lab.primals.eco via Cloudflare tunnel. Four tiers:

TierAccessResourcesEnforcement
AdminFull control48 GB RAM, 16 coresRoot-equivalent
ComputeKernels, dispatch, workspace32 GB RAM, 8 coresvenv, wheelhouse, per-user scratch
ReviewerDashboards only, no executionRead-only showcaseNoKernelManager, chmod 550, symlinks
ObserverRendered output + provenanceRead-onlyNo terminal, no kernel, no file access

Every restriction mechanism-enforced: filesystem permissions (root-owned), kernel blocking, iptables owner-match (drops internet for restricted users, preserves LAN), hidepid=2, ACLs on system binaries.


Monthly Operating Cost

CostAmount
Electricity~$150
Internet~$80
Domain / DNS~$5
Hardware depreciation~$250
Total~$485

No VC funding. No grants. No institutional backing. One person pays the electricity. The attribution pipeline (sweetGrass) exists so that when community contributions flow, credit flows back proportionally.

The parts list and the monthly bill are transparent. The cost makes sense when you see what it replaces.

The repos are open. Build your own.


Verify

  • Deployment: github.com/sporeGarden/projectNUCLEUS
  • All services: github.com/ecoPrimals
  • Hardware inventory: documented in the ecosystem
  • Live system: lab.primals.eco
  • License: AGPL-3.0-or-later