SHAH'S HARDWARE STORE

Full Catalog

Every GPU, CPU, memory module, server, rack, and support plan we carry — from a $249 Intel Arc card to a $6.25M GB300 NVL72 rack. Filter by category, vendor, and market tier.

18 products

Intelconsumer

Arc B580

The cheapest real GPU on the shelf — Intel's Battlemage value champion.

$249
Intelconsumer

Arc A770

First-gen Arc with 16GB — a cheap way to have more VRAM than a 4060.

$349
NVIDIAconsumer

RTX 5070

The affordable on-ramp to Blackwell — enough VRAM for quantized 7B–13B models.

$649
NVIDIAconsumer

RTX 5070 Ti

The mainstream Blackwell pick — most of the 5080's memory for less.

$899
NVIDIAconsumer

RTX 5080

The sensible sibling to the 5090 — still a capable local-inference card.

$1,199
AMDconsumer

RX 7900 XTX

AMD's consumer flagship — big VRAM per dollar if you're ROCm-friendly.

$1,339
NVIDIAconsumer

RTX 4090

Last generation's flagship — still a very real local-AI workhorse.

$1,999
NVIDIA

L4

The entry point. Cheap, low-power, and plenty for small inference jobs.

$2,800
NVIDIAconsumer

RTX 5090

The fastest consumer card NVIDIA makes — a hobbyist's on-ramp to local AI.

$2,999
NVIDIA

L40S

Ada Lovelace's do-everything card — fine-tuning, inference, and graphics.

$8,500
NVIDIA

A100 PCIe

The Ampere legend, still training real models for a fraction of Hopper cost.

$12,500
NVIDIAworkstation

RTX PRO 6000

96GB of workstation VRAM — the biggest single card you can put in a desktop.

$13,250
AMD

MI300X

AMD's answer to H100/H200 — more memory, a real second sourcing option.

$15,000
NVIDIA

A100 SXM4

NVLink-connected Ampere for real multi-GPU training pods.

$16,000
NVIDIA

H100 PCIe

Drop-in Hopper for standard servers — no SXM baseboard required.

$28,500
NVIDIA

H100 SXM5

The Hopper workhorse. Still the backbone of most training clusters.

$38,000
NVIDIA

H200 SXM5

H100's memory-hungry sibling — built for large-context inference.

$40,000
NVIDIA

B200 SXM6

Blackwell has arrived. The current frontier-training standard.

$45,000