Single 2U Supermicro chassis hosting the full agent stack, vLLM/SGLang inference, and the local
Bronze/Silver/Gold data layer. Sized to run Gemma 4-31B FP8 with KV-cache headroom and the
18-agent swarm concurrently. Hot-spare GPU in the second slot for zero-downtime through the
six-month pilot gate. Prices shown as ranges: low = current May 2026 estimate,
high = defensive ceiling sized to absorb supply-chain volatility (memory shortage markups, GPU
spot pricing, tariff / lead-time premiums).
48 GB GDDR6 ECC · Ada Lovelace · FP8 Transformer Engine · blower-cooled, dual-slot 300 W. Hosts inference and the swarm runtime; vendor support included.
Qty 1$9,000 – $11,500
RTX 6000 Ada · Spare
GPU · Hot Spare
NVIDIA RTX 6000 Ada
Identical card, second slot. Failover GPU for zero-downtime operation across the 6-month pilot gate. Drilled weekly. RTO target < 15 min.
Qty 1$9,000 – $11,500
NVMe U.2 · 7.68 TB
Storage
NVMe U.2 SSD · 7.68 TB
Enterprise U.2 PCIe Gen4 · 1 DWPD endurance · model weights, telemetry buffer, local log retention, and historian hot-tier. Bronze Parquet ~50 GB/yr/plant compressed.
Qty 2$3,800 – $5,000
The rig at a glance
GPU compute
2× RTX 6000 96 GB VRAM total
System memory
256 GB DDR5 ECC RDIMM
Local storage
15.36 TB NVMe U.2 Gen4 · 1 DWPD
Fabric
25 GbE dual-port OCP NIC · hardware TLS
End-to-end network topology · OT → DMZ → IT → HMI
Compute subtotal — May 2026 estimate · single edge inference rig
$32,300 – $41,500
Hardware BOM · Edge Inference Compute
02 / 04
Airgasan Air Liquide company
Hardware BOM · Part 2 Network · Power · Services
Section I · Hardware (cont.)
Plant network, power, enclosure & services.
Plant-side network, power, and rack enclosure follow ISA / IEC-62443 zoning. All items per-plant.
Includes one-time plant-side electrical install, structured cabling, environmental monitoring,
field-engineer commissioning, year-1 hardware support, and a 15% hardware contingency.
FortiGate 100F
Network · IT/OT firewall
Fortinet FortiGate 100F
Purdue L3/L3.5 segmentation — non-negotiable for the OT bridge. Strict ACLs OT→IT one-way; zero IT→OT writes for Phase 2. Logs forward to Group SOC via syslog.
Qty 1$11,000 – $14,500
10 / 25 GbE TOR Switch
Network · top-of-rack
10 / 25 GbE Managed Switch
Intra-rack networking and telemetry aggregation. SFP28 uplinks to the inference node and firewall. SNMP, sFlow, and 802.1X for operator HMI auth.
Qty 1$4,500 – $6,500
Cabling · Fiber · Patches
Network · cabling
Structured Cabling, Fiber, Patch Panels
Intra-rack copper and rack-to-firewall fiber runs. 1U fiber patch panel with SC/LC, splice trays, and labeled bundles. Documented as-built for plant IT.
Qty 1$2,000 – $3,000
Rack UPS · 6 kVA
Power · UPS
Rack UPS · 6 kVA Online
Online double-conversion · 3U rack form · pure sine ride-through for plant voltage events. Dual feeds to redundant PSUs. SNMP into Prometheus.
Qty 1$5,500 – $7,500
Smart Rack PDU
Power · PDU
Smart Rack PDU · 208 V / 30 A
Metered per-outlet, remote-controlled, SNMP. Feeds the redundant PSUs. Per-outlet alerting wired into Alertmanager.
Qty 1$2,500 – $3,500
208 V · 30 A
IT-room electrical
Power · plant-side install
Plant-Side Electrical Install
Dedicated 208 V / 30 A circuit, breaker, conduit, labor. One-time work in the plant IT room. Coordinated with plant facilities ahead of rig delivery.
Qty 1$8,500 – $13,000
Airflow · Env Monitor
Enclosure · environment
Rack Airflow + Env Monitoring
Blanking panels and airflow containment. Temperature, humidity, smoke sensors. SNMP to Prometheus + plant BMS. No liquid cooling at 300 W/card.
Qty 1$4,500 – $6,500
24U Half-Rack Cabinet
Enclosure · cabinet
24U Half-Rack Cabinet
Lockable front and rear doors · cable management · blanking panels. Standard server rack for the plant IT room. Mounts directly above the UPS.
Qty 1$2,500 – $3,800
Services, support & contingency
Line item
Function
Cost / site
Monitoring stack
Grafana + Prometheus dashboards for agent telemetry, GPU utilisation, inference latency, and rack environment.
$3,000 – $4,500
Install & commissioning labor
Third-party field engineer stand-up — 40 hours. Rack-and-stack, cabling, baseline configuration, smoke test against the digital twin reference.
$8,000 – $12,000
Year-1 hardware support
Supermicro / Dell ProSupport equivalent for chassis and GPU RMA. NBD on-site for the pilot duration.
$5,500 – $8,000
Hardware contingency (15%)
Covers RMA, cable re-work, plant-side remediation. Released to operations if unspent at month 6.
$13,500 – $18,500
Total — May 2026 estimate · single-plant rig (compute + N/P/E + services + contingency)
$103,300 – $142,800
Hardware BOM · Network · Power · Services
03 / 04
Airgasan Air Liquide company
Software Stack System-level only
Section II · Software
System-level software — what runs on the hardware.
Layers that connect directly to the rig hardware — base OS, GPU stack, container runtime,
inference, observability, security, backup. All system-level software below is OSS or
free-from-vendor. Anything with a license fee is listed at the bottom as a PAID
managed subscription. Application-layer Python / agent / data tooling is in the engineering
runbook — out of scope here.
Base OS · GPU stack · container runtime
Layer
Component
Function
License
Base OS
Ubuntu 22.04 LTShardened · CIS benchmark
Group baseline OS image. AppArmor enforced. Unattended security patching via Ubuntu Pro.
Ubuntu Pro
Init / services
systemd
Service supervision, journaling, time sync (timesyncd → plant NTP).
LGPL
GPU stack
NVIDIA driver 550+ · CUDA 12.4+ · cuDNN · NCCL
GPU device driver, compute runtime, comms library. Mirrored on internal artifact registry — no live external pulls.
NVIDIA
GPU monitoring
nvidia-smi · DCGM
GPU health, ECC, thermals, power draw. Exporter feeds Prometheus.
NVIDIA
Container runtime
Docker Engine · Docker Compose · Traefik
Container runtime + reverse proxy / TLS termination. Free Docker Engine on Linux (Docker Desktop is a separate paid item — see managed subs).
Docker Desktop site license for the dev team (Airgas > 250 employees triggers commercial-use clause). 15 seats × $5–$24/mo. Rig runs free Docker Engine on Linux.
Auto-unseal · namespaces · audit forward to Group SIEM. Small-team tier. Replaceable with OpenBao (Apache 2.0 fork) at zero cost.
$2,000 – $5,000
Misc · buffer
Small SaaS / individual-seat subs accumulated through the pilot — observability add-ons, certs, GitHub seats, ad-hoc tooling. Conservative ~$300–$500/mo.
$3,500 – $6,000
TOTAL · YEAR-1
Recurring OPEX — separate from one-time CAPEX. Low = current May 2026 estimate, high = defensive ceiling.
