"The next AI economy will not be won by people with the most prompts. It will be won by people with the most defendable assets."

Marketing · Sovereign top to bottom. Built on Swarm's own iron.

The Swarm & Bee marketing engine generates brand-disciplined posts and images on owned baremetal compute and stores them on owned sovereign storage. No OpenAI. No third-party R2-equivalents. No external API dependency. Every marketing asset carries the same provenance discipline as the Royal Jelly cells: sha256 receipts, defendable.eth-anchorable provenance, canonical wordmarks, no fake claims.

The four sovereignty rules

• No OpenAI. Image generation runs on a dedicated NVIDIA RTX 4500 Ada 32 GB owned by Swarm & Bee LLC. Apache-2.0 FLUX.1 schnell weights. No tokens billed to a third party. No prompts sent to an external API.
• No external storage. Generated PNGs land on the Synology DS1525+ NAS at /volume1/marketing-assets/ — the same sovereign storage that holds the order book and the medical vault.
• No SaaS dependency. Allowed external dependencies: Cloudflare edge/CDN/DNS, Tailscale mesh, GitHub source, Resend transactional email. Everything else runs on owned compute.
• Brand discipline at the model. The marketer-skill (Claude Code) enforces canonical wordmarks (defendable.eth, MicroScalers™, Build-A-Buility™), the thesis line, receipts-only claims, no AI buzzwords.

The pipeline · brief to library in <10 seconds

1. Brief — Donovan or Harvey invokes /marketer-skill from Claude Code with a 1–3 sentence brief.
2. Skill — The marketer-skill returns: expanded brief, 3 LinkedIn captions, 3 X captions, 1 Discord post, 1 FLUX.1 schnell image prompt, 5–8 hashtags. All brand-locked.
3. Shim — sb-marketing-shim on the dedicated marketing rig generates the image in ~3–6 seconds (RTX 4500 Ada · FLUX.1 schnell · 4 steps).
4. Library — PNG saved to /volume1/marketing-assets/library/<sha256>.png on the NAS with brief metadata JSON. Asset URL ready to publish.

The marketing rig · dedicated single-purpose 24/7 service

Hardware: NVIDIA RTX 4500 Ada 32 GB GDDR6 · 106 TFLOPS FP16 · 200 W TGP. AMD Ryzen 9 7950X (Zen 4 · 16C/32T) · 64 GB DDR5 · 2 TB NVMe. Ubuntu 24.04 LTS. On the Swarm tailnet as marketing-rig.

Single-GPU dedicated, not a cook rig. Marketing image gen is per-image (one user, one prompt at a time) — diminishing returns from multi-GPU parallelism on this workload. The 4-GPU cook rigs (4× RTX PRO 6000 Blackwell on Threadripper 9970X) are right-sized for training; this is sized for a service. Single-rig opex: ~$25/mo at $0.10/kWh industrial power.

Why baremetal beats cloud image generation

Comparable cloud image APIs (OpenAI gpt-image-1, Replicate, Together) charge $0.04–$0.10 per image. The marketing rig breaks even at <100 images/month and produces >10×'ed iterations/hour (no API rate limits). The brand discipline (skill) runs locally; only the image-rendering step needs compute, and that runs on owned silicon.

The marketer-skill (ships before the rig)

The brand-discipline layer is independent of the image-generation backend. The skill is live in Claude Code today; it produces FLUX-syntax prompts that can be pasted into any compatible runner (ComfyUI, Diffusers, hosted FLUX) until the sb-marketing-shim is online on the dedicated rig. The skill auto-checks every output against:

• No fake claims (revenue · users · clients-by-name unless confirmed)
• Canonical wordmark casing on every mention
• Thesis line included naturally in long-form posts
• Receipts cited from real numbers (234 GPUs, 956,175 medical cells, 240,423 CRE / $42,300 BTC first deeded sale, $0.10/kWh)
• No AI buzzwords (leverage / synergize / transformational / cutting-edge)
• Hashtags follow the 5–8 broad/mid/narrow-branded mix rule
• Image prompts carry the "Powered by Swarm & Bee 🐝" overlay

Architecture spec

Full deploy spec — Dockerfile sketch, FastAPI endpoints, NAS path conventions, canary-watchdog integration, BOM, monthly opex math — committed to the swarmandbee-app source tree at docs/marketing-shim-architecture.md. The spec is the document the build team executes from when the marketing-rig comes online.

Identity and verification

Marketing-rig: marketing-rig.tail80f341.ts.net on the sudosuops.github Tailscale tailnet. Asset library: /volume1/marketing-assets/ on the Synology DS1525+ NAS (192.168.0.102). Image generation: FLUX.1 schnell (Apache 2.0, Black Forest Labs). All marketing artifacts produced under defendable.eth trademark discipline. Inquiries: [email protected].