Best Local AI GPU Prices Right Now: RTX 5090, 4090, 3090 and More
A price-per-GB and price-per-bandwidth breakdown of the best and worst GPUs for local AI right now, from RTX 5090s to V100s.
The best value GPUs for local AI right now are older cards, not new ones
Right now the RTX 3090 and the modded 2080 Ti (22GB) offer the best combination of price per gigabyte of VRAM and price per gigabyte-per-second of bandwidth among mainstream options, while the RTX 5090 and RTX 4090 sit at the most expensive end of that same scale despite being the fastest cards available. Used GPU prices for local AI have spiked hard in recent months, driven partly by demand from China and partly by general FOMO, which means the “best” card depends less on raw speed and more on what you’re paying per gigabyte of memory and per unit of memory bandwidth.
TL;DR
- The RTX 5090 is currently the most expensive card on a price-per-VRAM-GB basis, with used units going for around $5,400, a premium of roughly $3,400 over its MSRP.
- The RTX 4090 is following the same trajectory, with used prices now sitting above its original MSRP even though it launched years ago.
- The RTX 3090 remains one of the strongest value picks, trading below its original MSRP while still offering 24GB of VRAM and nearly a terabyte per second of bandwidth.
- A modded 22GB RTX 2080 Ti and the 12GB RTX 3060 both post attractive price-per-GB and price-per-bandwidth numbers, making them viable budget entries.
- Enterprise castoffs like the V100 (32GB) can be found used for a small fraction of their original list price, though they need external cooling and aren’t plug-and-play.
- AMD’s Radeon Pro W7900 and 7900 XTX offer competitive bandwidth-per-dollar, and Apple/AMD/Nvidia AIO systems like the DGX Spark trade raw bandwidth for large unified memory pools.
- Avoid the RTX 4060 Ti, which still commands high used prices despite mediocre memory bandwidth that drew criticism even at launch.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Why do price-per-GB and price-per-bandwidth matter more than raw specs?
Two numbers tell you almost everything you need to know about whether a GPU is a good deal for running local AI models: dollars per gigabyte of VRAM, and dollars per gigabyte-per-second of memory bandwidth. VRAM capacity determines whether a model fits on the card at all. Bandwidth determines how fast tokens actually generate once it does. A card can have huge capacity and still be sluggish if the bandwidth is thin, and a card can be blazing fast but useless if the model doesn’t fit in memory.
Looking at both numbers together separates genuinely good deals from cards that are simply popular. A 5090 delivers 32GB of VRAM and roughly 1.792 terabytes per second of bandwidth, which is serious performance, but at current used prices near $5,400 that works out to about $168.75 per gigabyte of VRAM and roughly $3.00 per gigabyte-per-second of bandwidth. That’s the highest price-per-capacity of any card in this comparison, RTX 6000-class and datacenter cards aside.
How do the RTX 5090, 4090, and 3090 actually compare?
The 5090, 4090, and 3090 sit on a clear spectrum from expensive-and-fast to cheap-and-still-capable.
The 5090 launched with a high MSRP already, and used prices have since blown past that by around $3,400, landing near $5,400 for the cheapest buy-it-now listings. Auction prices might dip lower occasionally, but the trend is upward, fueled by strong demand both domestically and from overseas buyers.
The 4090 launched at $1,599 with roughly one terabyte per second of bandwidth and 24GB of VRAM. Used units are now going for around $1,250, but this is misleading because it’s actually creeping toward and past MSRP again as demand catches up to the 4090 the way it already caught up to the 5090. Price per gigabyte of VRAM sits around $118.75, and price per gigabyte-per-second of bandwidth is about $2.83, both high relative to older cards.
The 3090, by contrast, has 24GB of VRAM and about 936GB/s of bandwidth, and used prices (around $1,300) are actually running roughly $199 below its original $1,499 MSRP. That works out to about $54.17 per gigabyte of VRAM and around $1.39 per gigabyte-per-second of bandwidth, half the per-GB cost of a 4090 and about a third that of a 5090. The caveat is that many used 3090s haven’t been repasted, and doing that maintenance (fresh thermal paste or pads) should be factored into the purchase if you want it running at sane temperatures under sustained load.
Are older or enterprise cards worth considering?
Yes, with caveats. A few older and enterprise-class cards post genuinely strong numbers once you account for their rock-bottom used prices:
- Modded 22GB RTX 2080 Ti: These are 2080 Tis with extra memory soldered on, offering 616GB/s of bandwidth. They launched around $999 and now sell used for about $575, undercutting MSRP by $424. That comes out to roughly $26.14 per gigabyte of VRAM and about 93 cents per gigabyte-per-second of bandwidth, genuinely attractive if you’re comfortable with a modified card.
- RTX 3060 (12GB): A long-time favorite for good reason. It launched at $329, and used units run about $300 today, still under MSRP. That’s about $25 per gigabyte of VRAM and 83 cents per gigabyte-per-second of bandwidth, making it one of the cheapest entry points into local AI, especially if you stack multiple units.
- Nvidia V100 (32GB): An enterprise-class Volta-generation card that launched around $11,500 and now sells used for about $625, a drop of nearly $10,900. With 900GB/s of bandwidth, it works out to roughly $19.53 per gigabyte of VRAM and about 69 cents per gigabyte-per-second of bandwidth. The catch is it has no built-in fans and needs a blower solution moving serious airflow, plus it’s loud. The 32GB variant is worth paying up for over the 16GB version since 32GB opens up meaningfully more headroom for models.
- Tesla P100/P4 (24GB variants): Around 346GB/s of bandwidth, these launched near $6,000 and now go for roughly $300, about $12.50 per gigabyte of VRAM and 87 cents per gigabyte-per-second of bandwidth. Still, the V100 is generally the better buy at a similar price tier because of its higher bandwidth.
- CMP 170HX: A repurposed, memory-cut mining variant of an Ampere-class card that can reportedly be found in versions up to 64GB, with bandwidth around 1.5 terabytes per second. These launched at $4,299 and now go for around $2,700, but there are reports of reliability issues, and this is really only a good idea for buyers comfortable troubleshooting hardware with uncertain warranty coverage.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Which new GPUs and AIO systems are worth buying today?
On the AMD retail side, the RX 9060 XT (16GB, 320GB/s bandwidth) launched at a low $349 MSRP and now trades around $500, still a reasonable performer for text generation workloads even though it’s now well above launch price. The 7900 XTX (24GB, 960GB/s) launched at $899 and is holding almost exactly at that price used, working out to about $37.50 per gigabyte of VRAM and 94 cents per gigabyte-per-second of bandwidth, solid for a 24GB card. The Radeon Pro W7900 (32GB, 640GB/s) launched at $1,299 and now sells for about $1,675 used, roughly $52.34 per gigabyte of VRAM and $2.62 per gigabyte-per-second of bandwidth, a premium option but one with a large VRAM footprint that some users report performing well in multi-GPU setups.
Among AIO/unified-memory systems, the AMD Strix Halo class (128GB, 256GB/s) launched near $1,999 and now runs about $3,350 used, and the Nvidia DGX Spark (128GB, 273GB/s) launched at $3,999 and sells for roughly $5,000 now. Both trade raw bandwidth for large memory pools, which matters for fitting bigger models but limits how fast tokens actually generate compared to discrete GPUs, and the Spark in particular is built around FP4 quantization rather than full-precision inference.
Which GPU should you avoid?
The RTX 4060 Ti stands out as the clearest cautionary example. It launched around $500 with only about 288GB/s of memory bandwidth, drawing criticism at launch for being underpowered relative to its price. Used units are still commanding around $600 today, roughly $101 over its original MSRP, despite the underlying bandwidth problem never having gone away. At about $37.50 per gigabyte of VRAM and $28 per gigabyte-per-second of bandwidth, there are better options at similar or lower price points.
Frequently Asked Questions
What’s the best cheap GPU for local AI right now?
The RTX 3090 (24GB) offers the strongest overall value, trading below its original MSRP while delivering solid bandwidth. The RTX 3060 (12GB) and modded 22GB RTX 2080 Ti are cheaper entry points with good price-per-GB numbers, especially if you plan to run multiple cards.
Is the RTX 5090 worth buying for local AI?
It’s the fastest card in this comparison but also the most expensive per gigabyte of VRAM and per unit of bandwidth, with used prices running thousands over MSRP. It makes sense mainly if you need the absolute fastest single-card performance and can absorb the premium.
Why are used GPU prices for AI rising so fast?
Demand for local AI compute has spiked, and cards like the RTX 5090 and 4090 are being bought up domestically and exported overseas, pushing used prices above original MSRP in many cases.
Are enterprise GPUs like the V100 practical for home use?
They can be, since used prices have collapsed relative to original list prices, but they typically lack built-in cooling and require a blower setup to manage airflow and noise, which adds cost and complexity.
Should I buy a used RTX 4060 Ti?
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Generally no. Its memory bandwidth was criticized at launch and used prices haven’t come down enough to justify it compared to alternatives like the RTX 3060 or RTX 3090.
