Not true
Computing happens with no server at all, or a tiny microcontroller can replace a production web and AI server.
WORKSHOP ISSUE No. 01 · BUILD IT LOCALLY
A PRACTICAL CIVIC COMPUTING PROJECT
Build a Local Internet in a Box
Two honest server plans—one compact and printable, one rack-mounted and expandable—show how a public website, local AI, searchable records and video processing can operate outside a hyperscale server farm.
NO CLOUD REQUIRED FOR THE CORE WORK
01
THE BIG IDEA
What “AI without a data center” actually means
True
A server is simply a computer that provides services. Websites, databases, search, video routing and inference on already-trained models can run on hardware in a library, newsroom, office or municipal building.
Where data centers still matter
Training frontier-scale models, serving millions of simultaneous users, global redundancy and enormous storage pools. Those needs are real—but they are different from running a local civic platform.
02
FOLLOW THE DATA
How a local AI-powered site works
Public website path
A browser reaches your domain. A reverse proxy terminates HTTPS, then sends the request to PerfectWorld, the database, search index or local AI API. A small VPS can be an optional edge relay when the local ISP uses CGNAT; it does not need to store the records or perform the heavy inference.
Video path
RTSP/HLS camera feeds enter MediaMTX. FFmpeg or a detector samples frames. The GPU performs object detection, transcription or summarization. PostgreSQL stores searchable metadata while ZFS stores selected clips. The public site receives results—not necessarily every raw frame.
Local LLM path
PerfectWorld calls an OpenAI-compatible endpoint exposed by llama.cpp or another local inference engine. Quantized model weights stay on local disks and are loaded into unified memory or GPU memory. Retrieval can search your own public-record corpus before generation.
03
CHOOSE YOUR MACHINE
Two builds, one open stack
KIT BUILD / LOCAL CIVIC NODE
CivicNode Micro
A compact 128 GB unified-memory server with mirrored NVMe and stackable two-disk cassettes.
PLANNING RANGE$5,800–$8,800before tax, labor and site work
Hardware headline
- AMD Ryzen AI Max+ 395 class: 16 cores / 32 threads
- 128 GB LPDDR5x unified memory; up to 112 GB can be assigned to graphics on supported configurations
- Two mirrored 4 TB M.2 NVMe drives
- One stackable cassette with two 24 TB SATA drives; add cassettes or a disk shelf later
- 5 GbE baseline networking
- Open-source Olimex ESP32-PoE-ISO environmental monitor
Good fit
- PerfectWorld / BorderLandMail web, database, search and document services
- A city-size community portal with sensible caching and modest concurrent demand
- Private local 7B–70B-class quantized model experiments with low-to-moderate concurrency
- A handful of 1080p video-analysis feeds when models and sampling rates are tuned
- Department, newsroom, library, school, workshop or disaster-recovery node
Limits to print in red
- Not high availability by itself—pair it with a second node and off-site backups for public service
- The soldered memory cannot be upgraded later; choose 128 GB if large local models matter
- Video retention, not web pages, usually consumes most storage
- No true hot-swap backplane in the printed cassette; shut down before servicing drives
- Exact AI and video throughput must be benchmarked
Electrical planning
- Idle
- 70–130 W
- Busy
- 190–330 W
Ordinary grounded 120 V outlet; dedicated circuit preferred when adding disks and UPS.
Parts list and cost worksheet
| Done | Category | Qty | Part | Why / specification | Planning cost |
|---|---|---|---|---|---|
| Compute | 1 | Framework Desktop Mainboard, Ryzen AI Max+ 395, 128 GB The kit centers on Framework's documented standard Mini-ITX board and open mechanical CAD. A complete Framework Desktop can also be used. |
Standard Mini-ITX; 16C/32T; Radeon 8060S; two M.2 sockets; 5 GbE | $2,800–$3,400 | |
| Cooling | 1 | Framework heatsink plus 120 mm PWM fan Do not substitute a cooler without checking socket, mounting pressure and case clearance. |
Use the matched six-heatpipe cooler and high-static-pressure 120 mm fan | $90–$160 | |
| Power | 1 | 400–500 W FlexATX ATX 3.x power supply Never mix modular PSU cables from different models. Verify the exact board and drive connectors. |
24-pin ATX, CPU power, at least four SATA power connectors, 80 Plus Gold preferred | $140–$240 | |
| Fast storage | 2 | WD Red SN700 4 TB M.2 2280 NVMe One drive may be used initially, but a mirror is strongly recommended for production. |
Mirror as ZFS; NAS-oriented endurance up to 5100 TBW for the 4 TB model | $600–$1,040 | |
| Bulk storage | 2 | Seagate Exos X24 24 TB SATA HDD Retail prices vary dramatically. Choose identical interface/sector format and retain a cold spare when possible. |
7200 RPM enterprise CMR; mirror for approximately 24 TB raw fault-tolerant capacity | $1,400–$2,200 | |
| Storage I/O | 1 | PCIe x4 six-port SATA HBA (ASM1166 class) For a larger shelf, move to a server HBA and a chassis with the matching physical slot and cables. |
AHCI/JBOD mode; low-profile bracket or riser as required | $70–$150 | |
| Monitoring | 1 | Olimex ESP32-PoE-ISO Use the non-WROVER version with the included I²C example, or remap I²C pins for WROVER variants. |
Open Source Hardware; isolated PoE; Ethernet; Wi-Fi/BLE; UEXT | $30–$55 | |
| Monitoring optional |
1 | BME280, leak probe, reed switch and small OLED Sensor pins are configurable in the included firmware and ESPHome example. |
Temperature/humidity/pressure, floor leak, enclosure door and local status | $35–$90 | |
| Resilience | 1 | 1000–1500 VA line-interactive pure-sine UPS Size from measured load and required runtime, not nameplate wattage alone. |
USB monitoring; replaceable battery; network shutdown support | $280–$550 | |
| Network optional |
1 | Managed 5/10 GbE switch and CAT6A patch cables Optional when connecting directly to an existing managed network. |
VLAN support; PoE port for the ESP32 monitor | $220–$650 | |
| Enclosure | 1 | CivicNode Stack printed enclosure hardware All printable files and parametric OpenSCAD source are included. |
PETG/ABS/ASA, M3 heat-set inserts, M3 screws, rubber drive grommets, fan guards | $90–$180 |
OFF-THE-SHELF TWIN
Minisforum MS-S1 MAX 128 GB plus a four- or eight-bay ZFS storage appliance
Comparable Ryzen AI Max+ 395 / 128 GB compute in a finished chassis, with dual 10 GbE and a PCIe expansion slot; the separate storage appliance improves serviceability.
Tradeoff: Less open mechanical design than the kit and a two-box footprint, but faster to deploy.
4U BUILD / PRODUCTION NODE
PerfectWorld Pro Rack
A serviceable EPYC server with ECC RAM, IPMI, two 32 GB GPUs and ten hot-swap bays.
PLANNING RANGE$16,000–$27,500before tax, labor and site work
Hardware headline
- AMD EPYC 7443P: 24 cores / 48 threads and 128 PCIe 4.0 lanes
- 512 GB ECC RDIMM with eight memory channels populated
- Two Radeon AI PRO R9700 32 GB blower GPUs (64 GB aggregate; software must explicitly shard or parallelize)
- Two mirrored enterprise NVMe drives for VMs, databases and indexes
- Eight 24 TB enterprise disks in RAIDZ2: 192 TB raw, approximately 144 TB before formatting/reserve after two parity disks
- Dual 10 GbE onboard, IPMI remote management and optional 25 GbE uplink
Good fit
- PerfectWorld Pro with multiple agencies, departments, modules and background jobs
- Concurrent local AI services, retrieval, transcription, embeddings and document processing
- Larger multi-feed video-processing workloads with separate ingest, inference and retention tiers
- Virtual machines, containers, snapshots, staging and production on one expandable platform
- A primary on-premises node paired with a second replication/backup node
Limits to print in red
- Two 32 GB GPUs do not automatically behave like one 64 GB GPU; tensor/model parallel software is required
- A single rack server still has a shared failure domain—use replication and tested restores
- Heat, noise, dust, power quality and physical access must be managed like a server room
- The selected EPYC generation is value-oriented, not the newest platform
- Camera throughput varies widely with codec, resolution, detector and frames sampled per second
Electrical planning
- Idle
- 280–500 W
- Busy
- 900–1,350 W
Dedicated 120 V / 20 A or 208–240 V circuit after an electrician verifies the load and UPS.
Parts list and cost worksheet
| Done | Category | Qty | Part | Why / specification | Planning cost |
|---|---|---|---|---|---|
| Chassis | 1 | Sliger CX4712 4U rackmount chassis Confirm rail kit, backplane/cable option and GPU clearance before ordering. |
Ten trayless 3.5-inch SATA/SAS bays, seven 120 mm fans, eight expansion slots | $440–$550 | |
| Motherboard | 1 | ASRock Rack ROMED8-2T Update BIOS/BMC before production and preserve a tested configuration backup. |
Single-socket SP3, eight DIMMs, seven PCIe 4.0 x16 slots, dual 10 GbE, IPMI | $550–$900 | |
| CPU | 1 | AMD EPYC 7443P Used/refurbished pricing is attractive; buy from a seller with return testing and verify carrier/torque procedure. |
24C/48T, 2.85–4.0 GHz, 200 W, 128 PCIe 4.0 lanes, eight-channel DDR4-3200 | $500–$1,100 | |
| Memory | 8 | 64 GB DDR4-3200 ECC RDIMM Use server-qualified modules and match rank/organization where possible. |
512 GB total; one matched DIMM per channel | $1,600–$3,200 | |
| AI accelerators | 2 | AMD Radeon AI PRO R9700 32 GB Use supported ROCm versions. Model sharding depends on the inference engine and model. |
32 GB GDDR6 ECC on Linux, 300 W, dual-slot blower, ROCm support | $2,600–$3,000 | |
| Fast storage | 2 | 3.84 TB enterprise NVMe SSD with power-loss protection M.2 2280/22110 fit and heatsink clearance must match the board. |
Micron 7450 PRO or equivalent; mirror for VMs, database and search indexes | $1,000–$2,000 | |
| Bulk storage | 8 | Seagate Exos X24 24 TB SATA HDD Eight drives provide 192 TB raw; usable capacity is lower after parity, formatting, snapshots and free-space reserve. |
RAIDZ2 / two parity drives; keep at least 15–20% free space | $5,600–$8,800 | |
| Storage I/O | 1 | Broadcom 9500-16i Tri-Mode HBA Use the exact SFF cable type required by the chassis/backplane. Avoid hardware RAID beneath ZFS. |
PCIe 4.0, SAS/SATA/NVMe support; operate as HBA/JBOD for ZFS | $800–$1,300 | |
| Power | 1 | 1600 W 80 Plus Platinum ATX power supply Do not use adapters of unknown quality. Verify branch circuits and UPS with an electrician. |
Two native 12V-2x6 GPU leads, strong 12 V rail, sufficient SATA power | $380–$650 | |
| Cooling | 1 | SP3 cooler, seven 120 mm PWM fans and ducting materials Blower GPUs are selected because they exhaust heat predictably in dense racks. |
Front-to-back airflow; server-room dust control | $300–$650 | |
| Network optional |
1 | Dual-port 25 GbE NIC and optics/DAC Onboard dual 10 GbE is sufficient for many installations. |
Optional high-speed uplink for replication or storage | $250–$700 | |
| Resilience | 1 | 2200–3000 VA rack UPS with network card Measure load at idle and full inference; test automatic shutdown and restart quarterly. |
Pure sine, monitored shutdown, replaceable batteries | $1,200–$2,800 | |
| Rack | 1 | Locking 18U+ rack, rails, PDU, cable management and temperature sensors Do not put the system in an unventilated closet. |
Grounded, ventilated, front/rear clearance | $700–$1,800 |
OFF-THE-SHELF TWIN
Dell PowerEdge R760xa configured with 512 GB RAM, two 32 GB-class GPUs and an external storage shelf
A support-contract alternative with enterprise rails, redundant power, remote management and validated GPU cooling.
Tradeoff: Usually several times the DIY hardware cost after GPU, memory, storage and support are configured; request a current quote rather than relying on list price.
04
DO THE MATH
Capacity calculator
05
PRINT THE CHASSIS
CivicNode Stack v1 enclosure
The original enclosure is a panelized, stackable prototype sized for a standard 170 × 170 mm Mini-ITX board, a FlexATX supply and removable two-drive cassettes. Every part fits a nominal 220 × 220 mm print bed. Add another storage deck when the archive grows.
- Complete STL packAll printable parts
- Parametric OpenSCAD sourceChange dimensions and regenerate
- Printing and assembly guideOrientation, inserts and tolerances
- Reference assembly STLVisual fit study—not a one-piece print
Prototype rule: print one mounting coupon and test-fit the real board, PSU and drives before committing to the full enclosure. Component revisions, connectors and cable bends vary. This printed chassis is not a safety-listed commercial enclosure; use a tested metal chassis or obtain qualified thermal/fire review before unattended public-service deployment.
06
FROM PARTS TO PUBLIC SERVICE
Eight-stage build plan
- Scope the service.List public pages, records, AI tasks, camera count, retention, privacy rules and recovery objectives before buying hardware.
- Print a fit coupon.Verify Mini-ITX holes, panel thickness, heat-set inserts, rear I/O opening and drive sled tolerance.
- Assemble and burn in.Update firmware, run memory and storage tests, monitor temperature, and test full-load power before storing public data.
- Create mirrored storage.Mirror boot/database NVMe. Use a separate ZFS pool for records and video. Schedule scrubs and SMART tests.
- Install the service layer.Use a supported Linux release, containers or VMs, a reverse proxy, database, backups, monitoring and a firewall.
- Add local AI deliberately.Start with a small quantized model in llama.cpp. Measure latency, memory and concurrency; then choose whether a larger model earns its cost.
- Move video compute local.Point camera streams at MediaMTX, sample frames for inference, write metadata to the database and retain only what policy allows.
- Prove recovery.Restore onto a second machine, simulate internet and power loss, document contacts, and repeat restore tests on a schedule.
07
OPEN-SOURCE TOOLBOX
Suggested software layers
Host
Ubuntu Server LTS for the simplest GPU path, or Proxmox VE for VM/container separation. Use ZFS where its memory and operational requirements are understood.
Web
Caddy or Nginx, PHP/Node, PostgreSQL, Redis and PerfectWorld. Put admin interfaces behind VPN or strong access controls.
AI
llama.cpp for portable quantized inference; vLLM or another supported ROCm engine when the Pro build needs batching or multi-GPU parallelism.
Video
MediaMTX routes RTSP/HLS/WebRTC. FFmpeg decodes or transcodes. OpenCV, Frigate or a purpose-built detector performs selected inference.
Operations
Prometheus/Grafana, SMART monitoring, ESP32 telemetry, UPS shutdown, restic backups and WireGuard for private maintenance access.
Deployment examples
The package includes Compose templates, a Caddy example and an ESP32 monitor project. They are starting points, not one-click production security.
08
THE PART THAT MAKES IT REAL
Reliability, safety and public trust
Three copies
Production data, a local backup and an encrypted off-site copy. A mirror protects availability from one disk failure; it is not a backup.
Two nodes
A second machine can hold replicated services and a recent database copy. The public DNS or edge proxy can switch traffic when the primary is unavailable.
One written policy
Define video retention, access logs, AI review, public-record preservation, redaction and incident reporting before the system collects sensitive material.
Electrical and fire safety: use listed power supplies, grounded outlets, appropriate UPS equipment and proper airflow. Do not fabricate mains wiring. A qualified electrician should verify any new circuit, rack PDU or 208–240 V installation.
09
CHECK OUR WORK
Research sources
20 primary and official references
- Framework Desktop specifications — Framework Computer. Mini-ITX dimensions; Ryzen AI Max+ 395; 128 GB; two M.2; 5 GbE; 400 W FlexATX; cooling dimensions
- Framework Desktop open mechanical documentation — Framework Computer / GitHub. Standard Mini-ITX and ATX connectors; CC BY 4.0 mechanical documentation
- Ryzen AI Max+ unified-memory local AI platform — AMD. Ryzen AI Max+ 395, Radeon 8060S and unified-memory platform
- AMD Ryzen AI Max+ developer platform and ROCm — AMD. ROCm and local model development on 128 GB Halo systems
- Minisforum MS-S1 MAX — Minisforum. Off-the-shelf 128 GB Ryzen AI Max+ 395 alternative, dual 10 GbE and expansion
- ESP32-POE-ISO Open Source Hardware Board — Olimex. OSHW certification, isolated PoE, board dimensions and supported toolchains
- WD Red SN700 — Western Digital. 4 TB M.2 NAS SSD and up to 5100 TBW endurance
- Exos X24 — Seagate. 24 TB enterprise SATA models and five-year limited warranty on listed models
- Sliger CX4712 — Sliger. 4U chassis bays, fan positions and expansion capacity
- ROMED8-2T — ASRock Rack. SP3 socket, DIMM channels, PCIe slots, SATA, dual 10 GbE and IPMI
- AMD EPYC 7443P specifications — AMD. Core/thread count, clocks, TDP, PCIe lanes and memory channels
- Radeon AI PRO R9700 — AMD. 32 GB GDDR6, power, cooling form and professional AI positioning
- Broadcom 9500-16i — Broadcom. PCIe 4.0 Tri-Mode HBA and 16 internal ports
- Dell PowerEdge R760xa — Dell Technologies. Off-the-shelf GPU rack-server comparison
- llama.cpp — ggml-org. Local and cloud inference, quantization, AMD HIP, Vulkan and CPU/GPU hybrid inference
- llama.cpp Docker images — ggml-org. Official CPU, ROCm and Vulkan server images
- MediaMTX documentation — MediaMTX. Self-hosted publishing, proxying, recording and playback of RTSP, HLS, WebRTC and other streams
- Proxmox VE administration guide — Proxmox. Virtual machines and containers
- ZFS on Linux — Proxmox. ZFS storage, snapshots and integrity features
- Parallelism and Scaling — vLLM. Tensor, pipeline and data-parallel deployment concepts
Research and planning date: August 8, 2026. The module deliberately uses planning ranges instead of pretending that volatile hardware prices are permanent.