Every build below is computed from the catalog data and the sizing rules; use them to check arithmetic. Prices in USD as of 2026-11-03; Supermicro lines at eStore B2B net (12% off list, list in brackets); the supplier confirms the final price on order. Continental US; Canada quoted case by case. Ready to buy: book a meeting and UpStation introduces the supplier (how buying works). Every page: llms.txt (map) · homepage (all tables) · catalog.json (data).
Hardware + colo over 36 months (taxes not included; see the power page)
$3,941,896.91
GPUs → nodes: 32 ÷ 8 = 4 nodes. Multi-node training means the GPUs act as one, so HGX class (B300, 288 GB per GPU, 2.3 TB per node); a PCIe server would not serve it. Host memory stays at the listed 2TB (no offload stated).
GPU fabric: 1 port per GPU = 32 × 800G. 32 ≤ 32, so one 32× 800G OSFP switch is non-blocking at 1:1 (single point of failure, accepted; growth past 4 nodes needs the 64-port switch or a fat-tree: at 32 nodes, 8 leaves + 4 spines on 64-port switches). Links: all 32 GPU ports priced optical (2× SR8 + MTP-12 fibre, $2,058 per link) — at 2 nodes per 40 kW rack the cluster spans 2 racks, so most links leave the switch's rack and 800G passive DAC (3 m reach) cannot be assumed; swapping the 16 genuinely in-rack links to DAC is a final-engineering saving, not part of the estimate. No storage on this fabric.
Front-end (converged with storage): each node takes the listed +2× 400G option and connects one port to each switch of an MLAG pair of 32× 400G QSFP-DD switches (400G AOC), 3:1 or better; the storage nodes' 2× 100G ports land on the same pair through 400G→4×100G breakouts (5 nodes × 2 ports = 10 ports, 2 breakout cables per switch). FS's 200G switches are backordered to 2027 (November 2026), so the front-end runs at 400G; peer link 2× 400G DAC. Uplinks to the buyer's core are not priced. Storage stays converged because 16 GB/s across 4 nodes is far below the 40 GB/s-per-node dedicated-fabric threshold.
Storage supply: 2U all-NVMe storage node, dual socket, 24x 7.68TB NVMe (184TB raw) = 184 TB raw, 20 GB/s planning per node. Replica 3 on N nodes plans to 184 × N × (N−1)/N × 0.333 × 0.85: N = 4 gives 156.2 TB (short), N = 5 gives 208.3 TB ≥ 204.5 TB. Throughput 5 × 20 = 100 GB/s ≥ 16. Erasure coding 4+2 would need 6 hosts and plan to 521.6 TB, 2.6× the need, so replica 3 on 5 nodes is the answer. Software (Ceph) is buyer-supplied and not priced.
Racks at 40 kW: B300 = 8 U and 16.96 kW provisioned, so 2 per rack (power binds). Rack A: 2× B300 + GPU switch + OOB switch = 35.24 kW; rack B: 2× B300 = 33.92 kW; rack C: 5× storage + front-end pair = 9.75 kW, 12 U. Racks, PDUs and cabling labour are quoted separately.
Power and colo: 78.91 kW provisioned in total; Montana colo at $260/kW/month × 78.91 kW = $20,517/month (36-month term; other locations on the [power page](/power)). Optional add-ons, not included: CPU nodes for cluster services (provisioning, scheduler, monitoring), such as the listed 1U compute node; settle with the buyer whether they want a services cluster.
Worked design: 2 PB usable archive on 5× HDD storage nodes (RAID 6), 100G front-end pair, Montana colo
Hardware + colo over 36 months (taxes not included; see the power page)
$340,491.55
Per node: 2U HDD storage node, single socket, 24x 24TB SAS (576TB raw) = 576 TB raw. RAID 6 as two 12-drive groups (10+2) = 0.833 efficiency; plannable = 576 × 0.833 × 0.85 (fill ceiling) = 407.8 TB per node.
Nodes: ceil(2,000 ÷ 407.8) = 5 nodes = 2,039.0 TB plannable (1.02× the need, inside the 2× overshoot check). Throughput 5 × 5 GB/s ≈ 25 GB/s streaming (estimate); no stated demand, so capacity binds. Each node is one failure domain: a node outage takes its 407.8 TB offline until repaired.
Alternative, distributed across nodes with Ceph erasure coding 4+2 (0.667 efficiency, 6-host floor): plannable = 576 × N × (N−1)/N × 0.667 × 0.85; N = 7 gives 1,959.4 TB (short), N = 8 gives 2,285.9 TB, so 8 nodes ($448,109.68 for the nodes) plus NVMe DB/WAL drives at 1–4% of HDD capacity via the rate card. Survives a whole node.
Network: each node has 2× 100G; one port to each switch of an MLAG pair of 32× 100G QSFP28 switches (10 ports of 32 per switch), peer link 2× 100G DAC, uplinks to the buyer's core not priced; the listed −$874 option swaps to 2× 25G if the core is 25G. OOB: 5 nodes + 2 switches on one 48× 1G RJ45 + 4× 10G switch.
Rack and power: 5 × 2 U + 3 U of switches = 13 U, 4.59 kW provisioned: one rack, one 208V 1-ph 30A circuit pair. Montana colo $1,193/month.
DIY 8× RTX 5090 node on the OneChassis 7U barebone (PCIe 5.0, DDR5), parts at market, Montana colo
Inputs: 8 retail GeForce RTX 5090 cards in one node; no supplier build; buyer assembles; parts bought at the dated market prices on the [rate-card page](/rate-card); 36-month Montana colo.
Hardware + colo over 36 months (taxes not included; see the power page)
$138,980
Host: the OneChassis PCIe 5.0 / DDR5 barebone ($5,285, chassis + TURIN2D24G-2L+ board + risers + 5× 2000 W) is the only listed host that takes retail 3.5–4-slot RTX 5090s. 2× EPYC 9354 (32C each, $1,695 market), 12× 32GB DDR5-4800 = 384 GB (one DIMM per channel, $800 each), one PM9A3 3.84TB U.2 ($533).
Cards: 8× RTX 5090 at the street median $5,999 = $47,992, 72% of the build. Street range $5,200–$7,450 per card moves the build between $60,408 and $78,408. MSRP $1,999 (NVIDIA, Jan 2025); first-party stock has largely evaporated and the purchasable price is 2.5-4x MSRP. The low end of the range is in-store pickup when a shop has stock; the high end is online marketplace stock that ships today
Total $66,800 against $76,679.60 for the Octoserver 8× RTX 5090 build (8× RTX 5090 · 2× 96C AMD · 512GB · 1× 3.84TB + 1× 1TB · 2× 25G, custom passive cards, 3-year warranty, assembled and tested): the DIY route saves $9,879.60 (13%) at the median, with 192 fewer cores and 128 GB less memory, no system warranty, and the buyer's own assembly, burn-in and GeForce-driver risk. Same host with the PCIe 4.0 / DDR4 board, 2× EPYC 7543 and 16× 32GB DDR4 (512 GB): $58,140.
Power: 8 × 575 W + 2 × 280 W + 12 DIMMs × 8 W + 1 drive × 15 W + 400 W fans/board = 5,671 W DC ÷ 0.92 = 6,164 W nameplate, 7.71 kW provisioned; the 5× 2000 W supplies give 8 kW at N+1. One 208V 3-ph 30A circuit pair. Montana colo $2,005/month, or self-hosted $386–$771/mo in an operator-built facility (the self-hosted band on the [power page](/power): $50–$100/kW/month, free cooling, basic UPS, one or two network providers at the low end); 7U per node, so 5 nodes per rack by space, 5 by power at 40 kW.
GeForce drivers exclude data-center deployment: this is a lab, workstation-class or on-premises node, not a hosted-service build. Every price here is a dated market read, not a vendor quote; re-check before ordering.
DIY 8× RTX 5090 node on the OneChassis 7U barebone (PCIe 4.0, DDR4), parts at market, Montana colo
Inputs: Same 8× retail RTX 5090 node on the PCIe 4.0 / DDR4 board (ROME2D32GM-2T): EPYC 7002/7003 CPUs and DDR4 at second-hand prices; 36-month Montana colo.
Hardware + colo over 36 months (taxes not included; see the power page)
$129,276
Host: the OneChassis PCIe 4.0 / DDR4 barebone ($4,585; the board is $400 cheaper than the TURIN board and restocks mid-October). 2× EPYC 7543 (32C Milan, $595 market; a 7742 64C is $680 and a 7532 32C $318), 16× 32GB DDR4-3200 = 512 GB (two DIMMs per channel, $240 each: $3,840 against $9,600 for 384 GB of DDR5 in the PCIe 5.0 design), one PM9A3 3.84TB U.2.
Total $58,140: $8,660 under the PCIe 5.0 / DDR5 design and $18,539.60 (24%) under the Octoserver build, with the same 512 GB of memory as Octoserver but 64 cores against 192. Cards are again 83% of the build.
Minimum-cost variant: 2× EPYC 7532 ($318 each) and 8× 32GB DDR4 (256 GB, one DIMM per channel) take the node to $55,666; 64 cores and 32 GB of host RAM per GPU are enough for inference replicas that load a model once. Below that the only lever is the cards: every $500 per card is $4,000 per node.
Getting the cards: NVIDIA has no server-quantity channel for GeForce; retailers cap quantities per customer, so the $5,200 low end is one or two cards at a time at in-store pickup and quantity comes from marketplace sellers at $7,450 and above ([market notes](/buying)). Price a fleet at the top of the range and treat acquisition as a procurement project; Octoserver's custom passive card ($5,200, 30 days) is the only RTX 5090 in the catalog that can be ordered in quantity.
Network: the ROME2D32GM-2T has 2× 10GbE RJ45 on board, so no NIC, transceivers or DACs: Cat6a to the listed 48× 10GBASE-T leaf ($5,659, [S5860-48XMG](/p/UPS-NET-10GT-48)), or a used Arista 7050TX-64 at about $175 (unsupported). 10G carries what 25G carries for an independent-GPU farm; it is a simplicity choice ([front-end for PCIe farms](/sizing)). The PCIe 5.0 node's TURIN board has 2× 1GbE and takes a NIC.
What PCIe 4.0 costs: the RTX 5090 is a PCIe 5.0 x16 card and runs at 4.0 x16 here, half the host bandwidth; inference and single-GPU training are unaffected in practice, multi-GPU all-reduce over PCIe is slower. Rome and Milan parts are three to five years old, with no warranty on used CPUs; the DDR4 is the same generation and plentiful.
Power: 8 × 575 W + 2 × 225 W + 16 DIMMs × 8 W + 1 drive × 15 W + 400 W fans/board = 5,593 W DC ÷ 0.92 = 6,079 W nameplate, 7.6 kW provisioned; one 208V 3-ph 30A circuit pair. Montana colo $1,976/month, or self-hosted $380–$760/mo operator-built.
DIY 8× RTX PRO 6000 Server Edition node on the OneChassis 4U barebone, parts at market, Montana colo
Inputs: 8 passive RTX PRO 6000 Server Edition cards in a 4U; buyer assembles; parts at the dated market prices on the [rate-card page](/rate-card); 36-month Montana colo.
Hardware + colo over 36 months (taxes not included; see the power page)
$214,323
Host: the OneChassis OCG7466M6 4U ($6,100, board included) with 2× EPYC 9354, 12× 32GB DDR5-4800 (384 GB) and one PM9A3 3.84TB, as in the 5090 design. Cards: 8× RTX PRO 6000 SE at $15,000 street = $120,000 (Octoserver sells the passive card at $14,800 in stock, inside this range; Supermicro's accessory card is out of stock. Data-center driver licence included, unlike GeForce. Unsolicited broker allocation offers above list appeared in October 2026; the p75 stays $18,000).
Total $139,623 against $163,909.60 for the Octoserver Tier 2 build (8× RTX PRO 6000 · 2× 96C AMD · 768GB · 1× 3.84TB + 2× 1TB · 2× 25G) and $177,430.31 for the Supermicro Tier 1 build: the DIY route saves $24,286.60 against Tier 2 at the median, again with fewer cores, less memory and no system warranty. RTX PRO 6000 carries a data-center-eligible driver licence, so unlike the 5090 build this one can be hosted.
Power: 6,382 W nameplate, 7.98 kW provisioned; Montana colo $2,075/month, or self-hosted $399–$798/mo operator-built; 4U per node.