The Engines
Every listing herein: true specifications, true price, true power drawn from the wall — and beside each figure, what it means and when it matters.
NVIDIA RTX 5090
The entry point. Great for gaming and small local models, but 32 GB is nowhere near enough for the big open models this store specializes in.
What it means: this is the ceiling upon how large a model's weights may be while resident on this engine. A model requiring more memory than this simply will not fit — no amount of additional compute remedies it.
What it means: 0.48 average homes' worth of continuous draw (575 W ÷ 1,200 W per home). At that rate, 156.5 hours of running consumes as much energy as one full electric-carriage battery (90 kWh).
Street price as of mid-2026. Launch MSRP (Jan 2025) was $1,999, but AI-workload demand has pushed real-world prices to roughly $3,499-$4,499 — we price at street level, not MSRP.
A single card only — no NVLink, so a second one does not add memory.
Inquire & Purchase →NVIDIA RTX PRO 6000 Blackwell
A serious workstation card. Enough for many mid-size models, but still short of the 100B+ frontier models this store advises on.
What it means: this is the ceiling upon how large a model's weights may be while resident on this engine. A model requiring more memory than this simply will not fit — no amount of additional compute remedies it.
What it means: 0.50 average homes' worth of continuous draw (600 W ÷ 1,200 W per home). At that rate, 150.0 hours of running consumes as much energy as one full electric-carriage battery (90 kWh).
NVIDIA's own listed price as of mid-2026 (up from an $8,565 launch price roughly 16 months earlier).
The largest VRAM of any single card, but still just one card — no NVLink bridge, so multiple units don't pool.
Inquire & Purchase →NVIDIA H200 (single GPU)
NVIDIA's memory-focused datacenter GPU. The building block for the NVLink server clusters below.
What it means: this is the ceiling upon how large a model's weights may be while resident on this engine. A model requiring more memory than this simply will not fit — no amount of additional compute remedies it.
What it means: 0.58 average homes' worth of continuous draw (700 W ÷ 1,200 W per home). At that rate, 128.6 hours of running consumes as much energy as one full electric-carriage battery (90 kWh).
Datacenter H200 cards run roughly $30,000-$40,000 depending on vendor/volume; we use $35,000 as a representative mid-point.
Sold standalone, this card's 141 GB is only usable by itself. It only pools with other H200s inside a matched NVLink server — see the HGX configurations below.
Inquire & Purchase →H200 HGX 4-GPU NVLink Server
The smallest real NVLink cluster we sell — a genuine shared-memory pool, not just 4 cards in one box.
What it means: this is the ceiling upon how large a model's weights may be while resident on this engine. A model requiring more memory than this simply will not fit — no amount of additional compute remedies it.
What it means: 2.33 average homes' worth of continuous draw (2800 W ÷ 1,200 W per home). At that rate, 32.1 hours of running consumes as much energy as one full electric-carriage battery (90 kWh).
4 x H200 ($35,000 each) in a factory-built NVLink baseboard (HGX-class node).
Real cluster: these 4 GPUs share one NVLink memory domain, so their memory truly adds up to 564 GB.
Inquire & Purchase →H200 HGX 8-GPU NVLink Server
The largest single-node NVLink cluster in our line-up — the standard building block behind most H200-based AI datacenters.
What it means: this is the ceiling upon how large a model's weights may be while resident on this engine. A model requiring more memory than this simply will not fit — no amount of additional compute remedies it.
What it means: 4.67 average homes' worth of continuous draw (5600 W ÷ 1,200 W per home). At that rate, 16.1 hours of running consumes as much energy as one full electric-carriage battery (90 kWh).
8 x H200 ($35,000 each) in a factory-built NVLink baseboard (the standard HGX/DGX H200 8-GPU configuration).
A full 8-GPU NVSwitch domain — all 8 GPUs' memory is addressable as one 1,128 GB pool.
Inquire & Purchase →NVIDIA GB300 NVL72 (full rack)
A full rack-scale AI supercomputer. Overkill for a single model, exactly right for a company that needs headroom for the biggest open models plus room to grow.
What it means: this is the ceiling upon how large a model's weights may be while resident on this engine. A model requiring more memory than this simply will not fit — no amount of additional compute remedies it.
What it means: 114.17 average homes' worth of continuous draw (137000 W ÷ 1,200 W per home). At that rate, 0.7 hours of running consumes as much energy as one full electric-carriage battery (90 kWh).
Full configured racks run roughly $3.7M-$6.5M depending on vendor (HPE/Dell/Lenovo/Supermicro) and configuration; we use $5,000,000 as a planning mid-point — get a vendor quote before buying.
The largest real cluster that exists today: 72 Blackwell Ultra GPUs across the rack share one 20 TB NVLink memory pool.
Inquire & Purchase →→ What clustering truly means, and whence the power/homes figures come