# Biohazard VFX — render, compute and control nodes

**Decision date / every price and benchmark seen: November 4, 2026.** Colorado, USA. USD unless explicitly GBP. List prices are observations, not multi-unit delivered quotations. All wattages and city-scene throughput below are **unverified engineering estimates**, not hardware we measured. `nodes.json` contains inputs, source URLs, dates, nulls for unknown costs and calculation outputs.

## Buy decision and totals

- **GPU:** buy two identical **single RTX PRO 5000 Blackwell 72GB / 128GB RAM / 4U** nodes; expand to four only after the city-scene test. One process per GPU, not shared with ComfyUI while rendering. The 72GB SKU is now real, rather than treating the older 48GB SKU as the only PRO 5000.[26][66]
- **CPU:** start with **one dual EPYC 7763 / 512GB refurb**, using CTO Servers H2-EPYC-128 as the quote target. Do not buy sixteen or sixty-four until representative Houdini/Nuke jobs and facilities capacity justify it. The seller is in the **UK**, and its current banner requires a valid quotation for every listed product; the advertised price is **not a verified current sale quote**.[4]
- **Control:** three identical **MS-01 i9-13900H / 32GB** machines, mirrored enterprise PLP SSDs, **ZFS replication**, not Ceph.[45][47][163]
- **Fabric:** two MikroTik CRS518 data switches and two CRS326 management/corosync switches; dual NIC links with active-backup failover.[27][262]

| Scope | Observed parts/list-price sum | What's outside it |
|---|---:|---|
| One GPU node | **$17,844.21** | Assembly, tax/shipping, qualified RAM/board revisions |
| Two GPU nodes | **$35,688.42** | Same exclusions |
| Four GPU nodes | **$71,376.84** | Same exclusions |
| Three-node control cluster | **$15,518.94** | Shelves, UPS, patch leads, Proxmox support subscription |
| Startup fabric including workstation/storage NIC retrofits | **$5,677.94** | Long-run fiber/optics and existing storage systems |
| Two GPUs + control + fabric | **$56,885.30** | CPU nodes excluded |
| Four GPUs + control + fabric | **$92,573.72** | CPU nodes excluded |
| Starter: two GPUs + one refurb CPU + control + fabric | **$57,965.26 + £12,895 advisory** | Refurb current quote, FX, delivery/import and all facilities/software costs **unverified** |
| Expanded: four GPUs + two refurb CPUs + control + fabric | **$94,733.64 + £25,790 advisory** | Same; these are not all-in USD totals |

Sums are calculated from the sourced BOMs below; the CPU additions included in the last two rows are $1,079.96 per machine. **Do not use these as delivered budget commitments.** New DDR5 and PLP-drive prices in the retrieved listings are expensive; no old RAM/SSD price has been silently substituted. Binding quantity availability is unverified.

## A. GPU render node — one card per 4U

| Part | Qty | Unit USD | Extended USD | Source |
|---|---:|---:|---:|---|
| PNY VCNRTXPRO5000B72-PB RTX PRO 5000 Blackwell 72GB | 1 | $8,999.99 | $8,999.99 | [66] |
| AMD Ryzen 9 9950X | 1 | $697.99 | $697.99 | [236] |
| Supermicro H13SAE-MF revision 2.0; Ryzen 9000 BIOS | 1 | $505.37 | $505.37 | [231] |
| Kingston KSM56E46BD8KM-32HA 32GB ECC UDIMM | 4 | $1,461.99 | $5,847.96 | [96] |
| Samsung 990 PRO 2TB MZ-V9P2T0B/AM | 2 | $389.99 | $779.98 | [115] |
| Mellanox ConnectX-4 Lx MCX4121A-ACAT dual SFP28 | 1 | $299.98 | $299.98 | [101] |
| Corsair RM1000x CP-9020271-NA 1000W ATX 3.1 | 1 | $209.99 | $209.99 | [226] |
| Noctua NH-D12L | 1 | $99.95 | $99.95 | [225] |
| Sliger CX4170a base | 1 | $249.00 | $249.00 | [190] |
| Sliger ARCTIC P12 Pro PST CO ACFAN00312A | 3 | $15.00 | $45.00 | [190] |
| Sliger GDRAIL-20XX-B | 1 | $109.00 | $109.00 | [190] |

**Per node: $17,844.21; four: $71,376.84.** Sliger's current configurator base is $249, with the black front panel at a **$0 configuration override**, three Arctic fans $45 and selected rails $109. Do not add the $40 spare front-panel price again. These figures were read from the maker's embedded configuration data, not an older comparison-site bundle.[190]

**Integration specification.** Require **H13SAE-MF PCB revision 2.0** and Ryzen-9000-compatible BIOS; Supermicro explicitly distinguishes the required revision. Put the double-slot GPU in the first CPU slot and the ConnectX NIC in the second, so the GPU operates at PCIe Gen5 x8 and NIC at its Gen3 x8 capability. This avoids the trap of trying to install an x8 NIC into a physically x4-only slot on a cheaper board. Four ECC UDIMMs give 128GB; use validated JEDEC settings, not EXPO, and expect the board's lower 2-DIMM-per-channel memory speed. Exact module QVL and supplier PCB revision remain **unverified**.[243][244][173]

The priced Kingston legacy part entered phase-out on September 24, 2026; its named replacement is KSM56E46BD8-32MH. Treat available legacy stock as a pilot-only quote and require an approved replacement/quantity quote before the repeat order; replacement pricing is **unverified**.[93][94]

**Storage:** mirror the two 2TB NVMes for OS and disposable project/model cache; keep authoritative assets and outputs on studio storage. These are consumer SSDs, not PLP Ceph devices.[115] Re-stage a cached scene locally before starting a frame range. **Power:** RM1000x 1000W ATX 3.1 with native 12V-2x6, single PSU. Use the 145mm NH-D12L under the chassis's 153mm limit; confirm the GPU power-plug/bend clearance with the lid on and actual PSU cable length/190mm PSU compartment clearance. Run a closed-chassis soak before cloning the build.[190][225][226]

**VRAM decision.** A city needing 28–33GB is too close to a 32GB device's limit once BVH, render buffers and other allocations count. Reject the 32GB RTX 5090 and RTX PRO 4500 as the standard city node. A pair of 32GB cards is not a 64GB memory pool. Blender can spill to host RAM on supported backends, but explicitly warns of the performance impact. The selected 72GB card leaves real scene headroom and runs at a 300W board limit; its extra VRAM is capacity, not a guaranteed speed increase over the 48GB model.[26][278]

**One short alternative:** RTX PRO 6000 Blackwell Workstation 96GB if a qualified workload really needs >72GB; observed CDW card list price **$16,513.99**, 600W. Stock/quantity is unverified. Do not buy the newer 84GB RTX PRO 5500 on a speculative price: NVIDIA now lists it, but regional pages still disagree between partner availability and “Coming Soon”; a buy-now price/lead time was **unverified**.[63][16][19]

### Cycles throughput: sourced benchmark, conditional city projection

Blender Open Data **5.2.0**, **OptiX**, median score; same version/backend for these comparisons. Score is a benchmark aggregate, not frames/hour.[280]

| GPU | Median score | Submissions |
|---|---:|---:|
| RTX 3090 | 5,087.13 | 102 |
| RTX 5090 | 16,072.21 | 218 |
| RTX PRO 5000 Blackwell 48GB | 11,691.00 | 4 |
| **RTX PRO 5000 Blackwell 72GB** | **10,956.48** | **5** |
| RTX PRO 6000 Blackwell Workstation 96GB | 17,305.66 | 19 |

Calculation from the user-supplied **50s render + 60s load on a 3090**:

`relative speed = 10956.480362 / 5087.132260 = 2.1538`

`render seconds = 50 / 2.1538 = 23.2152`

`frames/hour = 3600 / (23.2152 + 60 / frames_per_process)`

| Dispatch policy | One node | Two nodes | Four nodes |
|---|---:|---:|---:|
| Render-only / ideal persistent resident scene | 155.1 | 310.1 | 620.3 |
| A new process and 60s scene load **every frame** | 43.3 | 86.5 | 173.0 |
| Ten-frame chunks, one 60s load per chunk | 123.2 | 246.4 | 492.9 |

These are **unverified estimates**, not a delivery promise. For 17,327 frames, four nodes imply **27.9 render-only hours versus 100.1 hours loading every frame**, before retries, output I/O, denoise or load contention. Blender's persistent-data setting can help reuse data; ten-frame chunks do not guarantee that animated geometry/shaders avoid rebuilds.[279]

**Important baseline inconsistency:** the 3090 has 24GB, so “30GB per frame” cannot already be a purely resident 30GB GPU allocation on that card. It may mean host RAM, transient/combined memory, or existing out-of-core operation. The benchmark ratio does not model removing that spill. Measure both host peak RSS and peak VRAM; freeze Blender, driver, samples, denoiser, resolution and output settings before timing the new node.[278][284]

### Power / circuits

Budget **600W AC per node**, **2.4kW for four**; unverified, not the sum of PSU ratings. Four nodes alone correspond to **10A at 240V, unity power factor**. Plan **two dedicated 240V/20A branches** with A/B UPS/PDUs and distribute the single-PSU nodes. At a conventional 80% continuous planning limit each branch provides 16A/3.84kW; have a licensed electrician confirm the applicable installation, actual PF, receptacles, load, ventilation and service capacity. Do not infer that a 1000W PSU continuously draws 1000W. Four GPU nodes + two 850W CPU nodes + 240W control + 238W switch maximum total **4.578kW**, before storage/HVAC/UPS losses; this is not a one-branch build. Power model and circuit sizing are **unverified planning assumptions**.

## B. CPU sim / comp / out-of-core node

**Selected quote target:** CTO **H2-EPYC-128**, 1U, **2× EPYC 7763**, 128 physical cores/256 threads, **512GB DDR4 (8×64GB)**, 2×480GB SATA boot SSDs, redundant PSU, rails, 12-month RTB warranty. Published **£12,895** on November 4, 2026; **current quote, availability, exact OEM chassis, U.S. delivered cost and PSU wattage unverified**. The CPUs are 280W TDP each, not the whole-system wall load.[4][241]

Require before purchase: exact OEM model; two supported NVMe PCIe x4 paths/riser positions; dual-25GbE NIC coexistence; two OEM-approved **≥1200W PSU modules** each able to carry the node after PSU failure; genuine DRAM capacity and a U.S. support/shipping arrangement. Add **2× 990 PRO 2TB + 1× MCX4121A-ACAT = $1,079.96** from the same sourced storage/NIC listings. If this 1U cannot take that layout, it fails the purchase gate; no invented compatibility claim is being made.[115][101]

The advertised eight DIMMs populate only four of each CPU's eight channels if balanced. Accept this as the initial capacity-first box, not a full-bandwidth sim configuration. Specify **16×64GB = 1TB** on the eventual memory-upgrade quote to populate both sockets fully; upgrade price **unverified**. Prefer one large sim per NUMA-aware host or separately pinned Nuke workers; do not promise distributed speedup for a single simulation. CTO's inexpensive-looking Xeon listing also describes its “512GB RAM” as **Intel DC PDIMM**, a warning not to equate PMem capacity with ordinary DRAM.[4][192]

**Only alternative: new Zen 5 single-socket EPYC 9355P**, 32 cores, 512GB ECC DDR5, H13SSL-NT rev2.x, same 4U design. Better per-core clocks and a simpler NUMA topology can matter for sims/comp; representative-workload speedup is **unverified**. Eight DIMMs use eight of twelve channels; full population would be 768GB. Here is the discrete-parts comparison, not an eStore “starting at” misrepresented as this configuration.[126][264]

| Part | Qty | Unit USD | Extended USD | Source |
|---|---:|---:|---:|---|
| AMD EPYC 9355P | 1 | $4,490.99 | $4,490.99 | [72] |
| Supermicro H13SSL-NT rev 2.x | 1 | $1,994.00 | $1,994.00 | [166] |
| Samsung M321R8GA0EB0-CWM 64GB DDR5-5600 ECC RDIMM | 8 | $3,924.00 | $31,392.00 | [171] |
| Samsung 990 PRO 2TB | 2 | $389.99 | $779.98 | [115] |
| MCX4121A-ACAT | 1 | $299.98 | $299.98 | [101] |
| Corsair RM1000x | 1 | $209.99 | $209.99 | [226] |
| Supermicro SNK-P0084AP4 | 1 | $94.45 | $94.45 | [178] |
| Sliger CX4170a | 1 | $249.00 | $249.00 | [190] |
| ARCTIC P12 Pro PST CO | 3 | $15.00 | $45.00 | [190] |
| GDRAIL-20XX-B | 1 | $109.00 | $109.00 | [190] |

**New-node observed parts sum: $39,664.39.** The eight RAM modules alone account for $31,392 at this vendor's observed $3,924/module. This is a high spot-retail comparison, **not a market-wide fair-price claim or recommendation to pay it**. A builder/volume quote may differ materially; none was available. AMD already describes Zen 6 / EPYC 9006 “Venice” in production; retail barebones/CPU stock, price and lead time for this small-studio build were **unverified**, so this recommendation deliberately uses purchasable Zen 5 parts, not an assumption that Zen 6 does not exist.[171][22][23]

### CPU Cycles parity is not sim / Nuke parity

Same Blender **5.2.0 CPU** medians, compared to selected GPU's **10,956.48 OptiX** score; all seen November 4, 2026.[281][280]

| CPU configuration | CPU score | Submissions | Nodes to equal one selected GPU, ideal render-only aggregate |
|---|---:|---:|---:|
| New comparison: single EPYC 9355P | 1,152.03 | 2 | **9.51 → 10 whole nodes** |
| Reference dual EPYC 7713, not the selected box | 1,934.65 | 1 | **5.66 → 6 nodes** |
| Reference new dual EPYC 9555, not the alternative BOM | 3,849.00 | 3 | **2.85 → 3 nodes** |
| Single EPYC 7763 | 1,182.05 | 2 | Not a dual-7763 measurement |
| **Selected dual EPYC 7763** | **unverified in 5.2** | — | **unverified; do not double the single score** |

Open Data's unfiltered, mixed-version dual-7763 median is **2,441.38 / 46 submissions**; it is context only, excluded from matched-version ratios. A sparse CPU submission can be power/NUMA/VM limited. Do not extrapolate these ratios to Mantra, Karma CPU, Nuke or simulation, or assume all cores scale linearly.[230]

### Power and Colorado electricity

Allow **850W AC/node for the refurb**, **550W for the new alternative**. These are **unverified steady-load planning allowances**, including RAM/fans/SSDs/PSU losses; measure at the PDU before fleet procurement.

Use **Xcel / Public Service Company of Colorado's effective October 19, 2026 rate sheet**, the newest electric summary listed when retrieved. **Schedule C winter total $0.12713/kWh**, summer **$0.16878/kWh**, fixed **$13.28/month**. Do **not** use the promotional rate-page's $0.06058 base charge as an all-in rate. C-TOU winter totals are $0.28313 on-peak / $0.11344 off-peak, not a flat $0.12713 rate.[121][24][64]

| Node design / count | IT kW | A at 240V, 1-phase, PF=1 | 720h kWh | Schedule C winter energy only | SG winter incl. demand + fixed |
|---|---:|---:|---:|---:|---:|
| Refurb ×1 | 0.85 | 3.54 | 612 | $77.80 | $154.35 |
| Refurb ×16 | **13.60** | **56.67** | 9,792 | **$1,244.86** | **$883.15** |
| Refurb ×64 | **54.40** | **226.67** | 39,168 | **Not eligible** ($4,979.43 hypothetical only) | **$3,215.31** |
| New ×1 | 0.55 | 2.29 | 396 | $50.34 | $137.20 |
| New ×16 | **8.80** | **36.67** | 6,336 | **$805.50** | **$608.78** |
| New ×64 | **35.20** | **146.67** | 25,344 | **$3,221.98**, only if whole-site peak remains <50kW | **$2,117.82** |

Calculated at 100% load for a **30-day/720-hour month**, cooling excluded. Add $13.28 once per meter to C energy-only totals, not once per node. The table's SG calculation uses **$0.03689/kWh + $11.15/kW distribution + $19.45/kW winter generation/transmission + $105.76/month fixed**, assuming steady farm demand equals billing demand and the flat-demand SG schedule. Tax/franchise fees, ratchets/minimums and actual existing site loads are excluded; these are not utility bill quotes.[121]

C/C-TOU require **whole-site peak below 50kW**. Sixty-four refurb nodes already exceed it before cooling; use **SG** for that planning scenario, not a misleading C extrapolation. SG can also be better for a continuously loaded smaller farm, but the actual meter/tariff controls. At assumed **PUE 1.4**, 64 refurb nodes become **76.16kW** and approximately **$4,459.13/month on the same SG model**; 64 new nodes approximately **$2,922.64/month**. PF=0.95 makes actual 240V current 5.26% higher than the unity-PF column. These are single-phase equivalent currents, not 208V three-phase panel schedules. This scale needs engineered service, dedicated cooling and likely colo evaluation, not office outlets.[238][121]

## C. HA agent + memory control plane

| Per identical node | Price | Source |
|---|---:|---|
| MS-01 i9-13900H, **32GB + 1TB**, U.S. variant | $1,295.00 | [45] |
| **2× Samsung PM9A3 MZ1L2960HCJR-00A07**, 960GB M.2 22110, new | $3,877.98 | [140] |
| **Per node / three nodes** | **$5,172.98 / $15,518.94** | Calculated |

Store's public variant JSON reports the selected MS-01 configuration available; this is not the cheaper barebone price with RAM accidentally omitted.[45] Keep the included 1TB NVMe as boot. Put the two PM9A3s in the two 22110-capable secondary slots as a **local ZFS mirror**, one slot running PCIe3 x2. The drives provide hardware PLP.[163] CDW's observed unit price is $1,938.99; actual quantity/ship date is **unverified**. Some third-party pages mislabel this exact MZ1L SKU as U.2 or 2280; obtain the correct **M.2 22110** part and inspect it before deployment.[25][140][268]

**Networking:** both onboard SFP+ ports at **10GbE**, one to each CRS518. Both 2.5GbE copper ports negotiate **1GbE**, one to each independent CRS326 for redundant corosync/management paths. Use separate VLANs and physical management uplinks; do not carry replication bursts on corosync. No add-in NIC required.[25][27][262]

**VM/storage design:** one active 8-vCPU/16GiB Linux VM contains the agent gateway, its SQLite-backed memory API and CPU embedding worker. A separate 4GiB VM holds the initial manager. Reserve 6–8GiB/host for Proxmox/ZFS, and test N+1 placement. This is not an HA GPU/large-LLM inference cluster: heavy AI jobs stay on the GPU fleet. Estimate **80W/node / 240W cluster**, unverified.

- Replicate the complete guest data volume to **both peers every minute**. It is asynchronous: RPO is **the last successful replication**, potentially older than one minute when jobs lag/fail. ZFS replication supports HA but is not synchronous Ceph durability.[47]
- Keep SQLite, WAL and SHM together on **local ext4 inside the VM's virtual block disk**, one active VM/writer, fsync-safe settings; never share the SQLite file over NFS/SMB or rsync a live database file alone. WAL specifically requires same-host shared-memory coordination. Use SQLite online backup/guest-aware backups to a different machine and offsite; snapshots/replication are not independent backups.[203]
- On host failure, Proxmox fences the failed host, establishes quorum and **restarts the VM elsewhere**. Official documentation says typical error detection/failover is **about two minutes**; plan **1–3 minutes plus guest/application recovery**, not an uptime SLA, live transfer of a dead process or zero interruption. Two remaining hosts can keep a three-voter cluster quorate; do not disable fencing to make a broken test “work.”[46]
- **Planned live migration** needs a healthy source and target, compatible CPU baseline and QEMU/machine versions, matching bridges/VLANs, spare target RAM, and guest disks available through replication or online storage-copy. Local ZFS does not forbid live migration: disk deltas and RAM can move. Avoid PCI/USB/GPU passthrough in these VMs. Test migrations and crash recovery with active SQLite writes; target restart time and actual RPO must be observed before production.[259][47]

**Ceph decision:** do not deploy it for these three small control machines. The chosen ZFS design accepts a defined asynchronous RPO and avoids a three-node shared-storage write/recovery dependency. The PLP drives are retained for durability even though Ceph was not selected. Proxmox subscription/support price: **unverified**, excluded from hardware totals.

## D. Switch fabric — buy redundant paths now

| Part | Qty | Unit USD | Extended USD | Source |
|---|---:|---:|---:|---|
| MikroTik CRS518-16XS-2XQ-RM | 2 | $1,595.00 | $3,190.00 | [27] |
| MikroTik CRS326-24G-2S+RM | 2 | $209.00 | $418.00 | [262] |
| MikroTik XQ+DA0001 | 2 | $39.00 | $78.00 | [206] |
| MikroTik XS+DA0003 | 28 | $39.00 | $1,092.00 | [204] |
| MCX4121A-ACAT | 3 | $299.98 | $899.94 | [101] |

**Total $5,677.94.** Two CRS518s supply 16×25GbE SFP28 + 2×100GbE QSFP28 each; the SFP28 ports also take 10GbE/1GbE devices. Switch/DAC figures are maker MSRP (switch also listed in stock at Streakwave at $1,595), not a cart quotation.[87] The two CRS518s' current specifications give **95W maximum each**; two CRS326s add **24W each**, total **238W max**. The CRS518 marketing paragraph still says 77W; budget the higher spec-table number.[27][262][245]

Per data switch allocate **4 GPU + 2 CPU + 1 workstation + 2 storage endpoints at 25G + 3 control nodes at 10G + 2 management-switch uplinks at 10G = 14 SFP28 ports**, leaving **two spare**. Twenty-four endpoint cables plus four management uplinks explain the **28 DACs**; two 100G cables join A/B as an inter-switch LAG. GPU NICs are already in the GPU BOM, CPU NICs in its additions, and only the workstation/two storage retrofits are charged here: no double counting. Storage systems themselves are excluded.

Use **active-backup bonds**, not an untested cross-switch LACP/MLAG configuration. Put storage/replication and render traffic on separate VLANs; monitor drops and bandwidth; keep MTU consistent. **10GbE fallback is the same CRS518 cages and XS+DA0003 DACs at 10G**, not a second switch purchase. Coded-cable acceptance on NVIDIA NICs must be tested. Long workstation runs need correct optics/fiber rather than pretending a 3m DAC reaches another room; cost **unverified**.[27][204]

This fabric is for the **2–4 GPU / one–two CPU** stage, **not** the 16/64 CPU hypothetical. Add leaf pairs with 100G spine uplinks for that expansion; high-port-count Mellanox/Arista or larger leaves need fresh quotes. Expansion network cost **unverified**. Existing storage throughput and a single NAS head can still bottleneck a nominal 25G fabric.

## E. Licenses and manager cost that scale with workers

All prices seen **November 4, 2026**, USD, software excluded from hardware totals. Budget simultaneous licensed workers, not just installed executables; exact checkout sharing across multiple processes must be tested against the purchased entitlement.

| Software | Decision / worker-cost consequence | Source |
|---|---|---|
| Blender / Cycles | No per-render-node license fee; use pinned Blender 5.2 builds. | [289][278] |
| **Houdini Indie** | **$299/year or $449/two years**; revenue/contracting entities ≤$100K and funding <$1M; max 3 Indie + 3 Engine Indie per entity. Not a general 16/64-node commercial-farm path; cannot mix its limited-commercial pipeline with commercial Houdini. | [196][51] |
| **Engine Indie** | **Free; max three dedicated farm machines** via the three Engine Indie entitlements. Each includes Engine/Mantra/Karma components. A paid Indie subscription's secondary installation is for that same artist, not an extra artist seat or unlimited farm permission. Check the exact issued entitlement; do not infer an arbitrary number of simultaneous render nodes from installation count. | [49][184] |
| **Houdini FX commercial** | Workstation **$3,505/year rental**, or **$4,495 perpetual** with **$2,740/year AUP** shown separately. Max 5 workstation licenses/facility; floating FX is **$5,480/year**. Need FX for artists, not every exported render worker. | [196] |
| **Houdini Engine commercial** | `hython`/`.hip` cooking or sims need Engine/appropriate Houdini entitlement: **$525/year workstation**, facility limit 5; global floating **$795/year single**, **25-pack $16,500/year**. Beyond 25 volume terms require quote. Free Unreal/Unity Engine is **not** a free command-line farm license. | [196][132] |
| **Karma / Mantra** | FX/Core workstation bundles **5 Karma**, floating **10**; extra Karma **$195/year** each. Bundled perpetual Karma requires active AUP. Rendering plain pre-exported USD through `husk` avoids full `.hip` cooking/Engine, but still uses Karma entitlement; see callback caveat below. Mantra needs Renderer token per rendering machine; comparison page lists five bundled and additional commercial Mantra tokens free on request. Do not buy FX per pure-render worker. | [133][196][184] |
| **Nuke Render (`nuke_r`)** | **$462/year per additional render entitlement**. One concurrent headless worker per licensed host budget: **16 → $7,392/year; 64 → $29,568/year**, before existing bundle credits. `-x` defaults to `nuke_r`; `-xi` uses an interactive entitlement instead and is not a free farm workaround. Plugin nodes may require separate farm licenses. | [258][185][272] |
| **Deadline 10** | **$0 license fee**, but in maintenance mode **since November 7, 2025**. Do not start a new studio's long-term dependency on it; free licensing does not mean active support for Blender 5.2/Houdini 22 plugins. Deadline Cloud is separate, not this on-prem-free product. | [150][55] |
| **OpenCue** | **$0 license, Apache-2.0**. Suitable future unified-farm migration when scale justifies operating its services; pipeline/admin labor not priced. | [182] |
| **Flamenco — chosen initially** | **$0 license, GPL-3.0**. Start with Blender; custom JavaScript job types can wrap Nuke/husk/hython CLI work, but integration is **not built or verified by this spec**. Keep its SQLite-backed manager single-active like the memory VM. | [179][180] |
| **HQueue** | Included, **no HQueue-server license**; Houdini/Engine/Renderer/Karma jobs still need their own licenses. Not an exemption from the three Engine Indie farm-machine limit. | [184] |
| **Tractor** | **$100 per license + $20/year maintenance**; confirm concurrent-task licensing, not one license blindly equated with a many-task physical node. Not selected for the startup farm. | [148] |

**Example license gate:** four GPU plus two CPU machines rendering Karma simultaneously need six worker entitlements. A single FX workstation bundle's five leave **one extra Karma token: $195/year**, provided the jobs are plain exported USD with no Engine-triggering callbacks; cooking `.hip` on each worker needs Engine licenses too. A four-node dedicated Indie batch farm already exceeds the three Engine Indie workers and needs a different valid entitlement plan.[133][184]

**Husk caveat:** pre/post-render Python callbacks that invoke Houdini libraries, or non-basic procedural settings, can also check out Engine/Core/FX even when the input is USD. SideFX explicitly documents these exceptions. Qualify the exact job before budgeting a Karma-only worker.[286][287]

AI-video / SAM models, paid ComfyUI nodes and Nuke plugins may impose separate commercial/worker terms; exact installed models/plugins and prices **unverified**, not assumed free. The user-supplied Fox rate of ~$2.03/job-hour is a billing observation, not a local GPU-hour benchmark; no cloud-equivalence or payback claim is made from it.

## Acceptance gates before a repeat order

1. **One pilot GPU:** confirm revision/BIOS/ECC reporting, per-device peak VRAM, CUDA/PyTorch model support, OptiX and Karma device logs; render the actual heaviest frame and a ten-frame range. Repeat cold and warm; compare pixels and record PDU watts. No fallback-to-CPU success masquerading as GPU qualification.
2. **One pilot CPU:** verify real DRAM capacity/channel layout, NUMA placement, cache I/O and closed-rack power/noise. Benchmark the studio's Nuke script and Houdini sim; do not purchase a CPU farm on Cycles scores alone.
3. **Control plane:** kill one host during an embedding/memory write, verify fencing, SQLite `integrity_check`, actual last committed record/RPO and restart latency. Test planned live migration and restore from independent backups.
4. **Network/facilities:** pull each link and switch power independently; preserve corosync quorum and service access. Load-test storage with simultaneous scene staging, obtain written CPU/quantity quotes, and get electrical/HVAC sign-off.

**Unresolved rather than invented:** binding refurb/U.S. delivered price; exact 1U OEM/NVMe/PSU approval; dual-7763 5.2 score; quantity availability; board/RAM QVL; true city throughput/residency; measured wall watts; support/subscription, facility, UPS and expanded-fabric costs. Hardware has not been purchased, built or benchmarked by this research task.

## Sources

[4] https://www.ctoservers.com/dual-socket-amd-epyc--1u-server--2x--epyc--7763--512gb-ram--nvme-4955-p.asp
[16] https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-5500
[19] https://www.nvidia.com/en-sg/products/workstations/professional-desktop-gpus/rtx-pro-5500
[22] https://newsroom.amd.com/news/amd-epyc-cpus-deliver-every-layer-agentic-ai-stack
[23] https://newsroom.amd.com/news/amd-announces-production-ramp-of-next-generation-a
[24] https://www.xcelenergy.com/company/rates_and_regulations/rates/rate_books
[25] https://store.minisforum.com/products/minisforum-ms-01
[26] https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-5000
[27] https://mikrotik.com/product/crs518_16xs_2xq
[45] https://store.minisforum.com/products/minisforum-ms-01-workstation.js
[46] https://pve.proxmox.com/pve-docs/chapter-ha-manager.html
[47] https://pve.proxmox.com/pve-docs/chapter-pvesr.html
[49] https://www.sidefx.com/faq/indie-new
[51] https://www.sidefx.com/faq/question/indie-restrictions
[55] https://aws.amazon.com/media-services/thinkbox
[63] https://www.cdw.com/product/pny-nvidia-rtx-pro-6000-graphic-card-96-gb-gddr7/8326705
[64] https://co.my.xcelenergy.com/s/business/rate-plans/time-of-use
[66] https://www.newegg.com/d/Best-Sellers/Workstation-Graphics-Cards/s/ID-449/Page-1
[72] https://www.newegg.com/p/pl?d=amd+epyc+9355p
[87] https://streakwave.com/mikrotik-crs518-16xs-2xq-rm-cloud-router-switch-650mhz-64mb-qsfp28
[93] https://www.kingston.com/en/memory/search?partid=KSM56E46BD8KM-32HA
[94] https://www.kingston.com/en/memory/search/pcn?partid=KSM56E46BD8KM-32HA
[96] https://www.cdw.com/product/kingston-ddr5-module-32-gb-dimm-288-pin-pc5-44800-unbuffered/8543445
[101] https://www.neutronusa.com/prod.cfm/2153509
[115] https://www.newegg.com/samsung-2tb-990-pro-nvme-2-0/p/N82E16820147861
[121] https://xcelnew.file.force.com/sfc/dist/version/download/?oid=00D1U0000011ttV&ids=068R300000nIWel&d=/a/R300000YLSLv/pSAmvUy.Aj2efQEs2fh8MVzV4TzrQ6j1JkmLbs3w0js
[126] https://www.supermicro.com/en/products/motherboard/H13SSL-NT?mlg=0
[132] https://www.sidefx.com/buy
[133] https://www.sidefx.com/products/karma
[140] https://www.cdw.com/product/samsung-pm9a3-960gb-m.2-solid-state-drive/6815347
[148] https://renderman.pixar.com/store
[150] https://awsthinkbox.zendesk.com/hc/en-us/articles/7865096598551-AWS-Thinkbox-products-are-now-available-free-of-charge
[163] https://download.semiconductor.samsung.com/resources/brochure/Samsung%20PM9A3%20NVMe%20PCIe%20SSD.pdf
[166] https://ctspoint.com/h13ssl-nt-supermicro
[171] https://networkdevicesinc.com/products/m321r8ga0eb0-cwm
[173] https://download.asrock.com/Manual/B650D4U.pdf
[178] https://store.supermicro.com/server-accessories/heatsink/4u-heatsink.html
[179] https://flamenco.blender.org
[180] https://flamenco.blender.org/download
[182] https://github.com/AcademySoftwareFoundation/OpenCue
[184] https://www.sidefx.com/faq/question/indie-renderfarm-setup
[185] https://support.foundry.com/hc/en-us/articles/207823889-Q100104-How-to-set-up-Nuke-and-Nuke-licenses-for-a-render-farm
[190] https://sliger.com/products/rackmount/4u/cx4170a
[192] https://ctoservers.com/hpe-proliant-dl380-gen10-server-dual-20-core-gold--6230-cpu--512gb-ram--nvme-100gb-qsfp-28--proxmox-4757-p.asp
[196] https://www.sidefx.com/products/compare
[203] https://sqlite.org/wal.html
[204] https://mikrotik.com/product/xs%5Fda0003
[206] https://mikrotik.com/product/xq_da0001
[225] https://www.newegg.com/p/pl?d=noctua%20nh-d12l
[226] https://www.corsair.com/us/en/p/psu/cp-9020271-na/rmx-series-rm1000x-fully-modular-power-supply-cp-9020271-na
[230] https://opendata.blender.org/benchmarks/query/?group_by=device_name&response_type=datatables
[231] https://amazon.com/dp/B0CWGHD2XG?channelId=500&lv=shuf&plpRedirect=mhFallback
[236] https://www.cdw.com/product/amd-ryzen-9-9950x-16-core-processor/8067000
[238] https://co.my.xcelenergy.com/s/business/rate-plans/comparison-tool
[241] https://amd.com/en/products/processors/server/epyc/7003-series/amd-epyc-7763.html
[243] https://www.supermicro.com/en/support/faqs/faq.php?faq=41469
[244] https://www.supermicro.com/en/products/motherboard/h13sae-mf
[245] https://cdn.mikrotik.com/web-assets/product_files/CRS326-24G-2SRM_211044.pdf
[258] https://www.foundry.com/trial-and-buy
[259] https://pve.proxmox.com/pve-docs/qm.1.html
[262] https://mikrotik.com/product/CRS326-24G-2SplusRM
[264] https://amd.com/en/products/processors/server/Epyc/9005-series/amd-Epyc-9355p.html
[268] https://www.memorysolution.de/en/products-online-shop/ssd-hdd/datacenter-ssd/samsung-pm9a3-ssd
[272] https://learn.foundry.com/nuke/content/comp_environment/configuring_nuke/command_line_operations.html
[278] https://docs.blender.org/manual/en/latest/render/cycles/gpu_rendering.html
[279] https://docs.blender.org/manual/en/latest/render/cycles/render_settings/performance.html
[280] https://opendata.blender.org/benchmarks/query/?group_by=device_name&blender_version=5.2.0&compute_type=OPTIX&response_type=datatables
[281] https://opendata.blender.org/benchmarks/query/?group_by=device_name&blender_version=5.2.0&compute_type=CPU&response_type=datatables
[284] https://store.nvidia.com/de-de/consumer/graphics-cards/geforce-rtx-3090-suprim-x-24g
[286] https://www.sidefx.com/faq/question/how-are-licenses-used-when-issuing-the-husk-command
[287] https://www.sidefx.com/docs/houdini/ref/utils/husk.html
[289] https://www.blender.org/about/license