Nvidia Dgx-2 Price

Note on this topic: the DGX-2 is a discontinued enterprise AI training system, not consumer gaming hardware. I’m writing this honestly for that context, since anyone searching this price is almost certainly researching rather than shopping to buy new, and I’ll steer gamers and hobbyists toward what actually fits their needs.

What the Nvidia DGX-2 actually was

The DGX-2 launched in 2018 as Nvidia’s flagship AI training appliance. It packed 16 Tesla V100 GPUs (32GB each), NVSwitch interconnects, dual Xeon Platinum CPUs, 1.5TB of system RAM, and 30TB of NVMe storage into a chassis that drew up to 10kW and needed data center cooling and power infrastructure. Nvidia’s original list price was around $399,000. This was never a product aimed at individuals, let alone gamers. It was sold to enterprises, research labs, and cloud providers doing large-scale deep learning.

Nvidia discontinued the DGX-2 line years ago in favor of the DGX A100 and later DGX H100 systems. There is no current retail price because Nvidia isn’t selling new units anymore. If you’re seeing a “price” quoted anywhere today, it’s either a historical reference, a used/decommissioned unit on the secondary market, or a listing that’s effectively scrap value for parts.

What a used DGX-2 actually costs now

Secondary market units show up occasionally through enterprise liquidators, university surplus auctions, or B2B resellers. Prices vary wildly depending on condition, hours of use, and whether the Tesla V100 GPUs are still functional, since those cards degrade and the system’s value is almost entirely tied to them. Realistic used pricing has ranged from roughly $15,000 to $60,000 depending on configuration and seller, down from the original $399,000 list. That’s still a huge sum, and you’re buying seven-year-old enterprise hardware with no warranty, no driver support roadmap, and parts that are increasingly hard to source.

For context on how this compares to other Nvidia enterprise systems, here’s a rough picture:

Notice the DGX A100 actually had a lower list price than the DGX-2 despite being newer and faster per GPU. Nvidia moved to fewer, more powerful GPUs per node rather than packing in 16 at once, which also simplified cooling and interconnect design.

Who should actually buy one

Almost nobody reading a GPU hardware site needs to think seriously about this. If you’re a research institution or company with an existing data center, serious power and cooling budget, and a specific need for proven, supported multi-GPU training infrastructure, you’re better off looking at current-generation DGX systems or cloud GPU instances (AWS, Lambda, CoreWeave) where you pay by the hour instead of committing six figures to hardware that’s already two generations behind.

Buying a used DGX-2 only makes sense in narrow cases: you need the specific NVSwitch all-to-all GPU topology for a legacy workload, you’ve already got the power and cooling infrastructure sitting idle, and you’re getting it at genuine scrap-adjacent pricing with GPUs you can verify are healthy. Otherwise the economics don’t work. You’re paying serious money for V100s that are slower and have less memory than current consumer and prosumer options, in a chassis that costs a fortune to run.

What to buy instead if you’re doing AI work at home

If you landed on this page because you’re interested in AI/ML work and wondered if enterprise gear is the move, it almost certainly isn’t, unless you’re running a company. For local model training and inference at a hobbyist or small studio scale, a single high-VRAM consumer or prosumer GPU goes a lot further per dollar. Something like an RTX 4090 graphics card gives you 24GB of VRAM and strong tensor core performance for a tiny fraction of DGX-2 money, and it also plays games, which the DGX-2 cannot do at all (it has no display output pipeline suited for that and no reason to have one).

If your workloads need more VRAM than a single consumer card offers, look at workstation cards instead of enterprise appliances. An RTX 6000 Ada workstation GPU gets you 48GB on a single card you can put in a normal workstation chassis with standard cooling and power. Two of those in a workstation will outperform a chunk of what the DGX-2 offered, for a cost that’s still enormous but at least measured in tens of thousands, not hundreds.

And if you just want to experiment with local LLMs or Stable Diffusion without committing to new hardware, renting cloud GPU time is the honest answer for most people. You get access to current-gen hardware (H100s, A100s) by the hour, with no upfront cost, no cooling problem, and no risk of buying a dead unit.

For PC gaming specifically, none of this matters. A DGX-2 is wildly inappropriate as a gaming machine, both because it has no gaming-oriented display path and because the money buys you dozens of actual gaming PCs instead. If you’re after frame rates, a single current RTX 4080 graphics card will outperform any individual V100 in the DGX-2 for gaming workloads, since the V100 architecture wasn’t designed with gaming rasterization or ray tracing performance as a priority at all.

FAQ

Can I buy a new DGX-2 from Nvidia today?

No. Nvidia discontinued the DGX-2 and it’s not part of their current lineup. Current systems are the DGX A100 and DGX H100.

Is a used DGX-2 a good deal for AI hobbyists?

No. Even at discounted used prices, the power draw, cooling requirements, and aging V100 GPUs make it a poor fit compared to a single modern GPU or renting cloud compute.

Can a DGX-2 be used for gaming?

Not practically. It’s built for headless AI training workloads, not display output or gaming drivers, and the cost makes no sense next to a dedicated gaming GPU.

Why did the DGX A100 cost less than the DGX-2 despite being newer?

Nvidia shifted to fewer, more powerful GPUs (8x A100 instead of 16x V100) with a more efficient design, which lowered both hardware and infrastructure costs per system.