Cheapest GPUs Per GB of VRAM: New and Used Prices

The cheapest VRAM in local AI right now is a used NVIDIA Tesla P40 at roughly $6-13 per gigabyte, but the cheapest VRAM most people should actually buy is a used RTX 3090 at about $29-42 per gigabyte — 24GB, CUDA, and fast enough to be worth owning. If you want a new card with a warranty, the floor is the RTX 5060 Ti 16GB at about $37/GB, and 32GB starts at the AMD Radeon AI Pro R9700 at about $42/GB. And the single cheapest gigabyte on this whole page is not a GPU at all: a 128GB Strix Halo mini-PC lists as low as $15.62/GB.

Prices collected in the first week of August 2026 for every new card below (lowest in-stock listings at Newegg), and from our mid-2026 used-market survey for the second-hand rows. The 128GB mini-PC prices are store prices checked in the same early-August window. This is a snapshot, not a quote — during the 2026 memory shortage one box on this page moved more than 100% in ten months. Re-check every number before you spend money, and treat the ranking as more durable than the absolute figures. If the "last updated" date at the bottom of this page is more than about six weeks old, use the $/GB column as an ordering and go get your own prices.

Every row below names where its price came from and when. Nothing here was measured on a bench; VRAM and bandwidth figures are manufacturer specifications, cross-checkable in TechPowerUp's GPU specification database.

How do you calculate dollars per gigabyte of VRAM?

Divide the price you would actually pay by the card's VRAM in gigabytes. That is the entire formula, and its simplicity is exactly why it misleads people.

A used RTX 3090 at $850 with 24GB is 850 / 24 = $35.42 per GB. An RTX 5090 at a $3,695 street price with 32GB is 3695 / 32 = $115.47 per GB. The 5090 costs 3.3x more per gigabyte — and it is also roughly twice as fast per token, has a warranty, draws its power from a modern connector, and holds a 32B model at a quantisation the 3090 cannot. None of that appears in the $/GB number.

So use this page the way it is built: $/GB tells you what capacity costs, and the bandwidth column tells you what that capacity is worth. Token generation on a local LLM is memory-bandwidth-bound — every token drags the active weights through memory once — so a gigabyte attached to 936 GB/s and a gigabyte attached to 256 GB/s are not the same product. We put both columns side by side for that reason, and there is a whole section below on the GB/s-per-dollar view that flips the ranking.

Two more definitions, so the tables are unambiguous:

• New price means the lowest in-stock retail listing found at a named retailer on a named date, not MSRP — MSRP has been fiction across most of 2026.
• Used price means the typical street band from our own market tracking, quoted as a range. Ranges are honest here; a single "median" would imply a precision the second-hand market does not have.

Which used GPUs give the most VRAM per dollar?

The used market wins on price per gigabyte and always will, because the cards are depreciating assets and the new market is being priced by a memory shortage. The 3090 is the well-known answer; the P40 is the answer people flinch at.

Read that table from the bottom up and the story is clearer than from the top. The used RTX 4090 is the worst value on it — the same 24GB as a 3090 for more than double the money, because it went out of production and the shortage caught it. If you are shopping used and someone tries to sell you a 4090 as the "sensible" upgrade from a 3090, they are quoting you 2.6x the price per gigabyte for roughly 20% more speed.

The 3090 and the used 7900 XTX are within a few dollars per gigabyte of each other, which makes that a genuine fork rather than a value question: 3090 for CUDA (ExLlamaV2, TensorRT-LLM, most fine-tuning tooling) and NVLink; 7900 XTX for slightly more bandwidth and an AMD stack that has actually matured. Our used GPU buying guide covers the part this page deliberately does not — how to check a second-hand card, how to spot a mining history, and what recourse you have when it dies.

The P40 needs its own paragraph, below.

Which new GPUs give the most VRAM per dollar?

Four of these rows are lowest in-stock Newegg listings from the first week of August 2026; the rest are dated street bands and are marked as such. The gap between the price column and the MSRP column is the memory shortage in one glance.

Caveat on the A770 row, stated plainly because it is the cheapest number in the table: that $279 figure comes from our Arc A770 guide and is the oldest price on this page. It is the row most likely to be stale, and a 2022 Alchemist card is also the row where "in stock" is least reliable. Verify it against a live listing before you let it decide anything. If it holds, 16GB at $17.44/GB is the best new-card ratio available; if it does not, the honest new-card floor for 16GB is the 5060 Ti at $37.50/GB.

Three things worth pulling out:

• The RTX 5090 is a terrible capacity buy and a fine speed buy. At $115.47/GB it is roughly 3x the price per gigabyte of the R9700 for the same 32GB — but it also carries 1,792 GB/s against the R9700's 640, which is the widest bandwidth gap in the table. Nobody should buy a 5090 for VRAM. People buy it because those 32GB are the fastest 32GB in existence.
• The R9700 is the cheapest new 32GB card, full stop. $1,299 list, in-stock listings around $1,350-1,380 at Newegg in early August 2026. The trade is prompt processing: our R9700 review documents prefill running 2.6-3.4x slower than a 5090 on published benchmarks. Generation speed is close; ingesting a long document is not.
• Every MSRP in that table is below every street price. That is not a rounding error, it is the market. Our GPU prices and the memory shortage breakdown covers why — IDC's forecast that AI datacenters may consume up to ~70% of world memory output in 2026 is the root cause, and it is why NVIDIA raised the DGX Spark's MSRP mid-cycle rather than lowering it.

Why does a 128GB mini PC beat every GPU on price per GB?

Because unified-memory boxes buy LPDDR5X, and GPUs buy GDDR — and in a shortage the cheap memory stays comparatively cheaper. A 128GB Strix Halo mini-PC at its list price works out to $15.62 per gigabyte, which undercuts every card in the used table except the P40.

The most important column in that table is "stock." The two cheapest rows are boxes you could not buy in early August 2026: every GMKtec EVO-X2 configuration showed unavailable at that attractive $1,999.99, and the Framework Desktop was out of stock at $3,449. The cheapest gigabyte you could actually put on a credit card was the Minisforum MS-S1 Max at $28.43/GB — fractionally below the bottom of the used-3090 band ($29.17), for 5.3x the capacity at 27% of the bandwidth. Our best Strix Halo mini PC roundup has the box-by-box detail and the stock situation.

Then look at the last row and notice what $41.66/GB buys. The Mac Studio M3 Ultra is the most expensive gigabyte in this table and by a wide margin the best one — 819 GB/s of Apple-published bandwidth against 256 GB/s on every Strix Halo box. You are paying about 46% more per gigabyte for 3.2x the bandwidth. That is the trade the $/GB column cannot show you, and it is the reason we wrote the next section.

If you are choosing between these three platforms specifically, we broke the comparison out: DGX Spark vs Strix Halo vs Mac Studio.

What is the cheapest way to reach 16GB, 24GB, 32GB or 128GB?

Most people arrive at this page with a capacity in mind, not a card. Here is the table cut that way.

The 48GB row deserves a flag: two 3090s hold the same $29-42/GB as one, which is why dual-3090 is still the standard 70B answer, but the power and cooling are a real project rather than a footnote. 700W of sustained inference heat in one case is not a thing you solve with a better fan curve.

Once you have picked a capacity, the useful next question is what actually fits in it — our 24GB VRAM model picks and 16GB picks answer that per tier.

Why is the cheapest dollar per gigabyte usually the wrong buy?

Because you are not buying gigabytes, you are buying gigabytes per second. Flip the table to bandwidth-per-dollar and the ranking rearranges hard.

Using each row's midpoint price (arithmetic shown so you can redo it with your own numbers):

Now the P40 tops both tables — and it is still the row we would talk most people out of. This is where a pure ratio stops being a recommendation. From our Tesla P40 guide, the specifics: it is a September 2016 Pascal card with no Tensor Cores at all, FP16 runs at roughly 1/64 the rate of FP32 (about 0.18 TFLOPS), Flash Attention requires Ampere or newer so the P40 simply cannot use it, and prompt processing is correspondingly slow. It ships passively cooled with no fan and no video output, and it takes an EPS-style 8-pin rather than a PCIe connector — so budget a shroud, a fan and an adapter on top of the card. Inside its lane (GGUF Q4/Q5 through llama.cpp or Ollama, short prompts, patience) it is a genuinely capable 24GB card for the price of a budget gaming GPU. Step outside that lane and the numbers collapse.

The unified-memory boxes are the mirror image of the same lesson. At 0.07-0.09 GB/s per dollar they are the worst bandwidth value on the page — and they are still the only way to load a 120B model on a desk. That is not a contradiction; it is the point. Capacity you cannot buy any other way is worth paying a bad ratio for; capacity you could have had faster for the same money is not. Our memory bandwidth explainer works through why generation speed tracks GB/s so closely.

The practical rule that falls out of both tables: pick the smallest capacity that holds the models you actually run, then buy the most bandwidth available at that capacity. Capacity is a hard wall — under it nothing runs at all. Bandwidth is a dial. Optimise the wall first, then spend everything left on the dial. Where each current card lands overall is in our best GPUs for AI ranking.

When were these prices collected and how fast do they move?

New-card street prices: first week of August 2026, lowest in-stock listings at Newegg. Mini-PC prices: vendor stores, same week. Used bands: our mid-2026 market tracking. The GX10, DGX Spark and RTX 5090 Founders figures come from our mid-July 2026 survey.

How fast do they move? Fast enough that this section exists. Two documented examples from our own tracking, both concerning the same hardware:

• The Beelink GTR9 Pro was described by ServeTheHome's October 2025 review as often selling for "$1999 or less." It listed at $4,349 on Beelink's own store in early August 2026 — more than 100% higher in about ten months, with no hardware change.
• NVIDIA raised the DGX Spark Founders Edition MSRP from $3,999 to $4,699 in late February 2026, explicitly citing the DRAM/NAND shortage. Vendors do not usually raise MSRP mid-generation. This one did.

So: we re-check this table roughly monthly, and the modified date at the top and bottom of the page is the one to trust. If it is stale by more than about six weeks, the safe move is to keep the ordering, discard the absolute numbers, and pull live prices yourself. TechPowerUp's database will confirm any VRAM and bandwidth figure here in seconds; the prices you have to go and look up.

One structural note about used prices, since it is the question we get most: we quote bands rather than a single median because the second-hand market genuinely is a band. A 3090 from a gamer with the box, at $1,000, and a 3090 from a liquidated mining rig, at $700, are different products wearing the same name — and the $/GB spread between them is $12 per gigabyte, larger than the entire difference between several rows in the new-card table. What separates the two is inspection, not arithmetic, and that is the used GPU buying guide's job rather than this page's.

FAQ

What is the cheapest GPU with 24GB of VRAM?

The Tesla P40, at roughly $150-300 used — about $6.25-12.50 per gigabyte, the lowest ratio on this page by a wide margin. It is a 2016 Pascal card with no Tensor Cores, no Flash Attention support, unusable FP16, no fan and no video output, so budget for a cooling shroud and an EPS-to-PCIe adapter and expect slow prompt processing. If you want 24GB that is fast as well as cheap, the used RTX 3090 at $700-1,000 ($29-42/GB) is the card almost everyone should buy instead.

What is the cheapest new 32GB GPU?

The AMD Radeon AI Pro R9700 at $1,299 list — in-stock Newegg listings were around $1,350-1,380 in early August 2026, which works out to roughly $42-43 per gigabyte. The only other 32GB consumer-class option is the RTX 5090 at a $3,695 street price, or $115.47 per gigabyte. The R9700 gives up prompt-processing speed (2.6-3.4x slower than a 5090 on published benchmarks) rather than capacity.

Is a used RTX 3090 still the best value for local AI?

For a single card with CUDA, yes — $700-1,000 for 24GB at 936 GB/s is a combination nothing new comes close to on price per gigabyte. The used RX 7900 XTX is within a few dollars per GB with slightly more bandwidth if you can work in ROCm. What has changed is the ceiling above it: a 128GB Strix Halo box now reaches the same $28-30 per gigabyte at 5.3x the capacity, so "best value" depends entirely on whether your models fit in 24GB.

Why is a 128GB mini PC cheaper per GB than a graphics card?

Because it is buying LPDDR5X system memory rather than GDDR6/GDDR7 attached to a wide GPU bus, and LPDDR5X is both cheaper per gigabyte and much slower. A 128GB Strix Halo box runs at 256 GB/s; a used RTX 3090 runs at 936 GB/s. You are getting 5.3x the capacity at roughly 27% of the bandwidth, which is exactly the right trade for large mixture-of-experts models and exactly the wrong one for anything that would have fitted in 24GB.

Should I wait for prices to fall?

Nothing in the current reporting points at a near-term correction — memory contracts are the bottleneck and the AI buildout writing them shows no reported sign of slowing. The one dated thing that could reset the mid-range is the delayed RTX 50 Super refresh, reportedly on hold over 3GB GDDR7 pricing with CES 2027 the most-cited window. That is a leak-driven timeline, not a plan. Buy for the models you need to run this quarter; the memory shortage explainer has the full picture.

Does dollars-per-GB matter more than tokens-per-second?

Only up to the point where the model fits. Capacity is a hard wall — below it the model does not load at any speed — so getting over the wall is the first purchase decision and $/GB is the right lens for it. Above the wall, bandwidth decides everything about how the machine feels, and $/GB actively misleads: the cheapest gigabytes on this page (P40, Strix Halo) are attached to the slowest memory. Pick the capacity first, then buy the most GB/s available at that capacity.

Sources

• Prices for new cards and mini-PCs: lowest in-stock listings at Newegg and vendor stores, checked in the first week of August 2026, as recorded in our RTX 5060 Ti 16GB guide, RX 9070 XT guide, Radeon AI Pro R9700 review and Strix Halo mini-PC roundup
• GB10 and RTX 5090 Founders pricing: our mid-July 2026 market survey in GPU prices and the memory shortage
• Used-market bands: RTX 3090 guide, RX 7900 XTX guide and Tesla P40 guide, mid-2026
• TechPowerUp GPU specification database — VRAM capacity, memory type and bandwidth for every card listed
• Apple Mac Studio technical specifications — M3 Ultra 819 GB/s memory bandwidth and 96GB configuration
• ServeTheHome, Beelink GTR9 Pro review (October 2025) — the "$1999 or less" figure used in the price-movement example