Why aren't these in the eight-box lineup?

The lineup's promise is 'same chip in all eight — pick on chassis and price'. These are a different machine class with different trade-offs, so they get their own page and their own honest maths.

Why no prices on this page?

The RTX PRO 6000 Blackwell has a verified US street price of $11,830 against an ~$8.5k MSRP (August 2026) — AI demand has it inflated roughly 40% and moving weekly. We quote at order, pass hardware through at cost, and the invoice includes the receipts.

Are the speed figures measured?

No — and they are labelled. They are our measured Spark numbers scaled by the bandwidth ratio, which is how memory-bound decode behaves. When we build the first unit we publish measurements and replace every estimate on this page.

When are two clustered Sparks the better buy?

When capacity beats speed: €13k buys a 256GB pool that runs bigger quants than the €30k two-GPU tower's 192GB. The tower wins on raw speed and native-FP8 fidelity. We will tell you which side of that line your workload falls on.

The tier above

Big iron that still lives at home

Tower workstations with RTX PRO 6000 Blackwell GPUs — server-class inference on a normal wall socket. Same install, same receipt discipline, same honest maths: we tell you below exactly where the desk boxes stay the better buy.

3.0kW

One EU/UK wall socket. Under ~2.5kW runs on a normal circuit — that is the real ceiling of “at home”.

Tower only

A workstation under load is office-loud; liveable in a study. Rack units are 70dB screamers — disqualified at any price.

6.6×

PRO 6000 memory bandwidth vs a GB10 (1.79TB/s vs 273GB/s). Decode is memory-bound; speed follows bandwidth.

Two configurations we actually stand behind

Built on Supermicro and Gigabyte tower platforms (Xeon W / Threadripper PRO), assembled by an EU integrator with warranty, installed by us. Prices quoted at order — GPU street prices are inflated ~40% over MSRP right now and move weekly; we will not publish a number that goes stale.

Y Tower 96

from €18,900

Hardware at cost, from ~€16,900 + €2,000 install · all-in, ex VAT · indicative — GPU street prices move weekly, so we reprice at order and the invoice carries the receipts.

96GB GDDR7 · 1.79TB/s~750W · any socket

A Threadripper or Xeon tower with one RTX PRO 6000 Blackwell. Six and a half times the memory bandwidth of a GB10 — decode is memory-bound, so single-stream speed scales with it.

DeepSeek V4 Flash via the DS4 87GB build — est. 90–180 tok/s (unmeasured; scaled from our Spark numbers)
gpt-oss-120b (59GB), Laguna NVFP4 (71GB), Qwen3.5 NVFP4 (75.6GB) — all with real context headroom
FLUX.1-dev (95GB) — fits; image generation in seconds, not minutes

32GB less memory than a Spark, no unified pool, three times the wall power, audible fans, roughly three times the price. Our 103GB flagship quant does not fit — the 87GB DS4 build is the flagship path here.

Request this build Email us

Y Tower 192

from €31,800

Hardware at cost, from ~€29,800 + €2,000 install · all-in, ex VAT · indicative — GPU street prices move weekly, so we reprice at order and the invoice carries the receipts.

192GB GDDR7 across two cards~1.6kW · still one socket

The same tower with a second card. This is the cleanest flagship on any home machine we know of: the official 167GB FP8 checkpoint of DeepSeek V4 Flash fits WHOLE, with ~25GB left for context.

DeepSeek V4 Flash at native FP8 — no quantisation compromise at all
Qwen3.5 at UD-Q6 (112.4GB), MiniMax-M2.5 (101GB) — comfortable, fast
Real serving concurrency for a team, tensor-parallel over PCIe 5.0

No NVLink on PRO 6000 cards — fine for MoE inference, not a training rig. Two clustered Sparks cost about €13k and pool 256GB; this costs ~€30k and pools 192GB. You are paying for fidelity and speed, not capacity.

Request this build Email us

Cluster 512 — frontier scale, made of Sparks

Four GB10 boxes over a 200G switch: 512GB pooled, officially supported by NVIDIA's cluster tooling, and measured by three independent parties running GLM-5.2 — a 744B frontier model — at home.

Y Cluster 512

from €27,200

Hardware at cost from ~€19,200 (4× DGX Spark + 200G switch + approved cables) + 4× €2,000 install · all-in, ex VAT · indicative — repriced at order, receipts on the invoice.

512GB pooled · ~460GB usable~960W · one socketNVIDIA-supported 4-node topology

GLM-5.2 (744B) at 22–42 tok/s single-stream, ~104 tok/s aggregate — measured, three independent recipes
Qwen3.5-397B at 37–41.5 tok/s single user — measured
Starts as one box: every Spark you buy today is a quarter of this cluster later

Why it ships configured, not as a parts list: the measured numbers depend on a pinned driver (580.142 — newer leaks memory), community patches (one TP=4 kernel bug produced silent gibberish with no error), and RDMA actually verified on — setups that silently fall back to TCP lose half their throughput. We build it, verify the fabric, run the benchmarks with you, and hand over your numbers. Single-stream speed never beats one box — this is for models that cannot exist on one box. Beyond four nodes nothing has published evidence, so we do not sell it.

Request this build The measured numbers

What we will not sell you for your house

The ceiling is real. Above it, the honest answer is a rack in a facility — or more Sparks.

Four GPUs (384GB, ~€55k+)~2.9kW needs a dedicated circuit, and the honest fit check disappoints: GLM 5.2 (~376GB) does not fit once context exists. You would be buying noise and wiring for FP8-plus-concurrency — most buyers should take the two-GPU tier or cluster Sparks.
Anything rack-mount1U and 2U units are built for server rooms — 70dB+ under load. No price makes that liveable.
8-GPU HGX class10kW+. That is not a home purchase, it is a colocation contract — and if you need one, we will say so and help you spec it instead.

Questions, answered straight

Why aren't these in the eight-box lineup?: The lineup's promise is 'same chip in all eight — pick on chassis and price'. These are a different machine class with different trade-offs, so they get their own page and their own honest maths.
Why no prices on this page?: The RTX PRO 6000 Blackwell has a verified US street price of $11,830 against an ~$8.5k MSRP (August 2026) — AI demand has it inflated roughly 40% and moving weekly. We quote at order, pass hardware through at cost, and the invoice includes the receipts.
Are the speed figures measured?: No — and they are labelled. They are our measured Spark numbers scaled by the bandwidth ratio, which is how memory-bound decode behaves. When we build the first unit we publish measurements and replace every estimate on this page.
When are two clustered Sparks the better buy?: When capacity beats speed: €13k buys a 256GB pool that runs bigger quants than the €30k two-GPU tower's 192GB. The tower wins on raw speed and native-FP8 fidelity. We will tell you which side of that line your workload falls on.