Specialist NVIDIA reseller

The fundamentals of compute.

Pre-owned NVIDIA GPUs and turnkey inference rigs, prepared the way a lab would prepare them: inspected, load-tested, warranted, and packed with the documents international freight actually requires.

See available cards Talk to the desk
Studio photograph of three pre-owned NVIDIA Founders Edition graphics cards on a charcoal surface
Photographed, tested, and graded before it is listed

Load-tested

Sustained workload, thermal logs, and a clean error pass before anything ships.

30-day warranty

Functional coverage from delivery. Dead-on-arrival cards are replaced, not argued.

US & export

Domestic ground plus international air, with commercial invoices and packing lists.

Honest grade

Used hardware, described plainly. No “as-is lottery,” no silent mining damage.

Inventory

Pre-owned NVIDIA GPUs

Ampere and Ada cards chosen for inference, fine-tuning, and small production — not leftover mining boards with the fans already dying. Quotes follow the live market. Stock turns quickly.

RTX 3070 Founders Edition style card, dual-fan, studio photograph
8 GB GDDR6 220 W Ampere

GeForce RTX 3070

A compact inference card for 7B–13B models and lighter fine-tunes. Easy to place, easy to cool, still useful in 2026.

Quoted to current market Inquire
RTX 3080 Founders Edition style card, studio photograph
10 GB GDDR6X 320 W Ampere

GeForce RTX 3080

More bandwidth for batch inference and concurrent sessions. The step up when 8 GB starts clipping context or queue depth.

10 GB and 12 GB when found Inquire
RTX 3090 Founders Edition style card, larger triple-slot cooler, studio photograph
24 GB GDDR6X 350 W Ampere

GeForce RTX 3090

The local-AI standard. Enough memory for 30B-class models, longer context, or pairing into a multi-GPU rig.

Most requested card Inquire

Also sourced on request. RTX 3060 Ti, 4060 / 4060 Ti 16 GB, and occasional 3090 Ti. Tell us the VRAM floor and we will answer with what is actually on the bench.

Ask for a card

Systems

Turnkey AI rigs

Multi-GPU machines assembled, burned in, and shipped as a unit. Power, cooling, and cable dress are part of the job — not an afterthought you inherit.

Open-frame 4-GPU workstation on a workbench, tidy sleeved cables
4× GPU Inference Burned in

Four-card workstation

A compact inference server for a small team or a production pilot. Sized PSU, open frame or short chassis, documented thermals.

Typical build: 4× 3090 or mixed Configure
High-density 6-GPU open server chassis with dressed power cables
6× GPU Training / serve High density

Six-card server

High-density training and serving. Redundant cooling, dual PSU when the load wants it, and a packing plan that survives freight.

8× and 10× built to spec Configure

How we work

Prepared before it leaves the bench

Used compute is only a bargain if someone has already done the unglamorous work. That is the whole business.

01

Intake

Physical inspection, cooler and pad condition, VBIOS, power delivery, and a written note on prior use when we know it.

02

Load-test

Sustained workload on a test bench. We keep thermal and error logs. Cards that throttle oddly or throw ECC noise do not ship.

03

Grade

Condition described in plain language — cosmetic marks, cooler type, warranty start. You buy a known card, not a photograph.

04

Pack and clear

Foam, anti-static, commercial invoice, packing list, and an HS classification that matches the box. Then a tracking number.

Graphics card on an open test bench with a thermal graph on a small monitor
Burn-in on the bench
GPU packed in custom foam inside an export carton with a packing slip
Packed for freight

After the sale

Warranty, freight, and paper

Hardware that cannot be replaced quickly is not a good deal. Neither is a box that dies in customs.

Coverage

Thirty-day functional warranty

If the card or rig fails to perform as described, we replace it or make it right. Cosmetic wear on used metal is disclosed, not warrantied away.

  • Dead-on-arrival replacement, not store credit games
  • Rigs covered as a system, not as a pile of parts
  • We do not sell untested “as-is” lots on this desk

Dispatch

Shipped as if it has to clear

Domestic US plus international. We prepare the commercial invoice, packing list, and classification. Track the carton after it leaves.

  • FedEx tracking once the label is live
  • UK / EU paper prepared for the lane you are actually using
  • Dimensional weight and box called out before you pay freight

Open shipment tracking  ·  Classification desk

Questions

Straight answers

No. They are pre-owned, inspected, and load-tested. We describe cooler type, cosmetic grade, and anything that affects thermals or fit. If a card is effectively unused, we will say so — we will not dress used metal as retail-new.
Used 30-series moves with the week. A frozen number on a homepage goes stale and starts lying. Ask for a quote; we answer against current inventory and the market that day, not a page from last quarter.
Yes. We regularly prepare export cartons with a commercial invoice, packing list, and HS classification. Tell us the destination and whether you want DAP or to handle the import yourself.
For local inference, 24 GB of GDDR6X is still the value argument in 2026. Newer cards are faster per watt; they are rarely cheaper per gigabyte. If your constraint is VRAM, this generation remains the honest buy.
Yes, on request. We will not force six cards into a chassis that cannot cool them. Send the workload, the room’s power, and any board preference — we will answer with a build that can actually run.

The desk

Request a quote

Tell us the card or the rig, how many, and where it needs to land. We reply with what is on the bench, a market-aligned number, and a freight sketch — not a newsletter.

Prefer mail directly? hello@primaryvowels.com