UpStation automates GPU and IT purchasing by delivering curated content to your LLM. Explore some solutions then book a meeting once you want to engage with the OEMs.
Cables and optics (FS, generic-coded; reaches are rated maxima, OM4 for multimode)
Speed
2 m passive DAC
Passive reach
Active DAC / AEC
AOC 10 m (max length)
Short-reach optic, each (reach)
Long-reach optic, each (reach)
Breakout cable
800G
$239
3 m
AEC 5 m $1,424 (max 5 m)
$3,011 (50 m)
SR8 $879 (50 m)
DR8 $1,319 (500 m) · 2×FR4 $1,869 (2 km)
800G→2×400G 2 m $299
400G
$174
3 m
Active DAC 5 m $611 (max 5 m)
$1,055 (30 m)
SR8 $219 (100 m)
DR4 $549 (500 m) · FR4 $719 (2 km) · LR4 $999 (10 km)
400G→4×100G 2 m $199
200G
$120
3 m
Active DAC 7 m $325 (max 7 m)
$1,034 (50 m)
SR4 $299 (100 m)
FR4 $1,249 (2 km) · LR4 $2,624 (10 km)
200G→2×100G 2 m $119
100G
$54
5 m
Active DAC 9 m $300 (max 9 m)
$221 (30 m)
SR4 $99 (100 m)
CWDM4 $209 (2 km) · LR4 $399 (10 km)
100G→4×25G 3 m $136
25G
$39
5 m
Active DAC 9 m $73 (max 9 m)
$73 (30 m)
SR $49 (100 m)
LR $74 (10 km)
—
10G
$18
7 m
Active DAC 10 m $105 (max 10 m)
$39 (30 m)
SR $25 (300 m)
LR $34 (10 km)
—
1G
RJ45, Cat6 patch
—
—
—
10G SFP+ uplinks: see 10G row
—
Fibre patch (generic FS)
Cable
1 m
3 m
10 m
30 m
50 m
Used with
LC-LC duplex OM4
$8.40
$10
$20
$47
—
SR (10G, 25G)
MTP-12 OM4
$59
$89
$150
$301
$452
SR4, SR8, DR4, DR8
LC-LC duplex OS2
$8.40
—
—
—
—
FR4, CWDM4, LR, LR4
Link by distance
Distance
Use
Cost per link
Inside passive reach
2 m passive DAC
1 cable
To active max
Active DAC or AEC
1 cable
To AOC max, or to short-reach optic reach
AOC, or 2× short-reach optic + multimode fibre
1 cable, or 2 optics + fibre
Beyond
2× long-reach optic + single-mode fibre
2 optics + fibre
Any
Connected cost per port = switch $/port + link
Choosing a link by distance: 2 m passive DAC inside passive reach; active DAC or AEC up to its maximum; beyond that an AOC (10 m priced; longer lengths cost more) or two short-reach optics plus multimode fibre out to the optic's reach; past that, two long-reach optics over single-mode. At 100G and up the optics, not the fibre, set the cost. Connected cost per port = switch $/port + one cable, or + two optics plus fibre. GPU-fabric links are estimated all-optical: GPU node density (1–3 HGX nodes per 40 kW rack) puts most nodes past passive-copper reach of their leaf, since a 3 m DAC spans at most the adjacent rack. Mixing copper in is realistic — expect 25% or more of fabric cables to land as passive DAC once installation is engineered (peer links, co-racked leaf↔spine, nodes in or beside the leaf's rack) — but which links qualify depends on rack elevations, so estimates stay optical and copper is taken as a savings at engineering time.
Example: 1TB → 2TB with 64GB DIMMs = 1,024 GB × $39.53 = $40,478.72 (the H200 line lists this option at +$40,480). Marginal rates from non-exotic SKUs, taken from the eStore configurator (the price its cart charges). Within a family the price rises almost linearly with cores and capacity, so alternative configurations are not priced separately: apply these rates to the base build. Supermicro's drive pricing is high: buyers who source NVMe or HDD elsewhere should price drives from the market and the chassis from here. Exotic parts (Xeon 6768P/6788P/6748P at $250-320/core, 3DS 256GB DIMMs, dual-port or FIPS drives) sit outside these bands.
Nameplate max for PCIe GPU servers = (GPU TDP x count + CPU TDP x 2 + 8 W per DIMM + 15 W per drive + 400 W fans/board/NICs) / PSU efficiency (Titanium 0.94, Platinum 0.92); HGX servers use the vendor configurator total, which matches NVIDIA system guidance. Vendor configurator totals for PCIe boxes are kept in each product's basis as a cross-check only (Supermicro pads chassis overhead by 1.5-2.4 kW; Octoserver omits fans and PSU loss). Minimum circuit kW = nameplate / 0.8 so the circuit never runs above 80% (NEC continuous). Pick the smallest standard circuit whose rated kW (V x A x sqrt3 for 3-phase) >= that minimum; provision it twice (A+B) so either feed carries the box alone. Colo kW to buy = nameplate / 0.8; monthly colo = that kW x the term rate ($175 / $225 / $250 per kW-month for 5 / 3 / 1-year terms). Quoted to UpStation 2026-09-08, available in Canada and the US: $175/kW/mo on a 5-year term, $225 on 3-year, $250 on 1-year. Billed on the provisioned kW (nameplate / 0.8).
Size a GPU cluster on accelerators (GPU count and aggregate GPU memory), never on host cores. Every listed GPU server has 8 GPUs: nodes = ceil(GPUs / 8). Nodes ship fully populated; partial population is not offered. Model memory: required GPU memory (GB) = parameters (billions) × bytes per parameter × 1.2 (weights plus overhead) + KV cache (grows with context length × batch).
HGX or PCIe: Does the model at its precision fit one GPU's memory, and can each GPU work alone? Yes: PCIe. No: HGX. Distributed training is never served by PCIe servers. An HGX cluster of 2 or more nodes carries a dedicated non-blocking GPU fabric (1 port per GPU), a storage tier sized to feed the GPUs, and up to 2× the power per node of a PCIe server; a PCIe inference deployment carries neither the fabric nor the parallel-file tier.
Networks: GPU fabric 1 port per GPU (H200 400G, B300 800G) at 1:1, never carrying storage: one switch while the ports fit it, else on R-port switches leaves = ceil(ports ÷ (R/2)), spines = max(2, ceil(leaves × R/2 ÷ R)) (32 nodes on 64-port switches: 8 leaves + 4 spines). Front-end 2 ports per node at 3:1 on an MLAG pair; OOB 1× 1G per node and per switch. 1 cable per link; fibre links need 2 transceivers, DAC/AOC none. Estimate every GPU-fabric link as optical (2× short-reach optics + fibre, or AOC): node density puts most nodes in a different rack from their leaf, past passive-DAC reach.
Racks: 40 U and 40 kW per rack unless the site says otherwise; nodes per rack = min(by U, by kW); A+B feeds. Management: 3× the 1U compute node from 4 GPU nodes up, one OOB switch per rack.
GPU server
RU
Provisioned kW
Per 40 kW rack
Rack kW
Binds on
8x NVIDIA B300 (HGX B300 NVL8), Intel
8
16.96
2
33.92
power
8x NVIDIA H200 (HGX H200), Intel
8
11.34
3
34.02
power
8x NVIDIA RTX PRO 6000 Blackwell Server Edition, Intel
4
8.47
4
33.88
power
8x NVIDIA RTX PRO 6000 Blackwell Server Edition, AMD
Full rules, feed-rate tables, protection math and the questions to settle first: How solutions are sized. Worked designs computed with them (a 4× B300 cluster with storage; 2 PB usable on HDD nodes): Reference builds and worked designs.
NVIDIA B200 (HGX) servers: out of stock industry-wide; B300 is the listed alternative
GB200 / GB300 NVL72 rack-scale systems and liquid-cooled racks
AMD Instinct (MI300/MI325/MI355) servers
2-GPU and 4-GPU servers (every listed GPU server has 8 GPUs; hosts without GPUs have 8 slots)
Dell, HPE and Lenovo priced builds (Tier 1 pricing shown is Supermicro; other OEMs on request)
InfiniBand switches, and Arista or Dell/SONiC Ethernet switches (quote)
Vendor-coded optics (listed optics are generic-coded for the FS switches)
Racks, PDUs, structured cabling and installation labour
Software licences other than the PicOS switch OS; managed services; GPU cloud or rental
Questions and answers
What is UpStation? UpStation sells pre-priced NVIDIA GPU servers, network switches, optics and cables, and colocation, at published prices, so buyers and their AI assistants can design and price a deployment without a sales call. Delivery: Continental US; Canada quoted case by case. Orders and quotes: book a meeting.
What does an 8× NVIDIA B300 server cost? $572,364 for the Tier 1 OEM build (Supermicro SYS-822GS-NB3RT-01-G2: 8× B300 · 2× 64C Intel · 2TB · 1× 960GB M.2 · 8× 800G), ships 1-3 wk. It draws 13,566 W nameplate, so provision 16.96 kW; colocation is $2,968/mo 5-yr · $3,816/mo 3-yr · $4,240/mo 1-yr.
What does an 8× NVIDIA H200 server cost? $319,421.64 for the Tier 1 OEM build (Supermicro SYS-821GE-TNHR: 8× H200 · 2× 48C Intel · 1TB · 1× 240GB SATA · 2× 25G), 6-10 wk. Provision 11.34 kW; colocation $1,984/mo 5-yr · $2,552/mo 3-yr · $2,835/mo 1-yr.
What is the cheapest 8-GPU server listed? $76,679.60 for 8× RTX 5090 (passive server edition) from Octoserver; the cheapest 8× RTX PRO 6000 build is $163,909.60 (Tier 2). A host with 8 GPU slots and no GPUs starts at $15,146.20.
What circuit does an 8× RTX PRO 6000 server need? Any 8x RTX PRO 6000 server at 600 W caps: plan 6.8 kW nameplate (5.5 kW with cards capped at 450 W), provision 8.5 kW, one 208V 3-phase 30A circuit pair (A+B) per server; where three-phase is unavailable, 208V single-phase 50A. A 30A single-phase (L6-30) cannot carry it even at 450 W caps. Four servers per 40 kW rack. 200-240 V input only; the box needs four 3000-3200 W supplies in 3+1 for redundancy (Dell XE7745 matrix, ASUS/Supermicro/Exxact/GIGABYTE all ship that). Computed per build on the power page; the Tier 1 build is 6,780 W nameplate, 8.47 kW provisioned.
What does colocation cost? Colocation is billed on provisioned kW: $175/kW/month on a 5-year term, $225/kW/month on a 3-year term, $250/kW/month on a 1-year term, US or Canada. Provisioned kW = nameplate W ÷ 0.8 so no circuit runs above 80%, with A and B feeds each sized for the whole box. Example: one B300 node at 16.96 kW is $2,968/mo 5-yr · $3,816/mo 3-yr · $4,240/mo 1-yr.
How are prices set, and how do I get a firm quote? Every price is the vendor's published price (Supermicro, Octoserver, FS.com) in USD as of 2026-09-09. UpStation confirms the final price on order. Book a meeting with the bill of materials ready.
Can I get a different configuration? Each class has one priced base build. Intel and AMD equivalents are listed with a host delta. Memory, CPU and drive changes are priced with the rate card (marginal $/GB, $/core, $/TB) applied to the base build; options such as NICs, drives and warranty have signed price deltas. Anything marked quote, or outside the rate-card bands, is quoted by UpStation.
How many storage nodes does 2 PB usable need? With RAID 6 in 10+2 groups (0.833 efficiency) and an 85% fill ceiling, one 2U HDD storage node, single socket, 24x 24TB SAS (576TB raw) plans to 407.8 TB, so 2,000 TB needs 5 nodes (2,039.0 TB plannable, $258,717.25 for the nodes). Distributed with Ceph erasure coding 4+2 it takes 8 nodes (2,285.9 TB plannable) because Ceph also keeps (N−1)/N in reserve for self-healing and needs at least 6 hosts. Storage software is buyer-supplied; see the worked design.
How many switches does a B300 cluster need? One 800G port per GPU at 1:1. 4 nodes (32 ports) fit one 32× 800G OSFP switch ($37,499); 8 nodes fit one 64-port switch; 32 nodes (256 ports) need 8 leaves and 4 spines on 64-port switches, 256 host links and 256 uplinks. Add a front-end switch pair and one 48× 1G OOB switch per rack; the worked design prices the 4-node case with storage.
How many GPU servers fit in a rack? At 40 kW and 40 U per rack: 2× B300 (33.92 kW), 3× H200 (34.02 kW), 4× RTX PRO 6000 Tier 1 (33.88 kW), 4× RTX PRO 6000 Tier 2 (33.48 kW), 5× RTX 5090 (39.95 kW). Power binds before space in every case; a site limit replaces the 40 kW.
Notes
Prices as of 2026-09-09; final price confirmed on order. Continental US delivery; Canada case by case.
Intel and AMD builds for every node; memory and drives set the price. Exotic parts: quote.
3 years parts and labour included; on-site and extended warranty are options.
Draft, estimate and price-pending lines are marked; everything else is confirmed.
2026-09-08: First public catalog: B300, H200, RTX PRO 6000 (Tier 1 and Tier 2), RTX 5090, three GPU hosts without GPUs, FS switches with cables and optics, compute and storage nodes, rate card, power and colo model.
2026-09-09: Site rebuilt for people and AI assistants: visible tables, Markdown twin (catalog.md), short brief (brief.txt), computed power and colo fields and reference builds in catalog.json, JSON-LD, robots.txt and sitemap. Prices unchanged.
2026-09-09: Solution sizing added: how many nodes, switches, cables, storage nodes and racks a design needs (GPU count → nodes, fat-tree switch counts, storage protection math, racks and power), with two computed worked examples. Rack units and GPU memory per server added. Prices unchanged.
2026-09-09: Intel Xeon 6 host (SYS-422GA-NRT, 2x 6944P, TDX-capable) replaces the AMD AS-5126GS-TNRT as the Tier 1 host without GPUs; the AMD host stays as an equivalent. 48x 10G switch priced (S5860-48SC). Draft markers removed from every vendor-priced line. The AMD host is also kept as its own line.
2026-09-09: Supermicro lines re-priced from the eStore configurator (the price its cart charges) after an audit found that the AMD 5U host, the 1U compute node, both storage nodes, the H200 option deltas and the memory and drive rate-card rows had been priced from a stale secondary price list: hosts and the compute node moved a few percent, the storage nodes rose substantially because Supermicro's drive pricing is two to three times market (rate card updated). New Intel 5th Gen Xeon host (SYS-521GE-TNRT, 2x Gold 5520+, TDX) $37,937.10; Intel Xeon 6 host $47,884.46.
2026-09-09: Fabric cabling rule made explicit: GPU node density (1–3 HGX nodes per 40 kW rack) puts most nodes in a different rack from their leaf switches, so estimates price every GPU-fabric link as optical (2× short-reach optics + fibre, or AOC); passive DAC is a final-engineering substitution for proven in-rack or adjacent-rack links (25%+ of fabric cables landing as copper is a reasonable expectation, taken as a savings then, not in the estimate). Worked design updated to match. Prices unchanged.
Generated 2026-09-22; prices as of 2026-09-09; re-checked monthly.