What's the cheapest way to get 16, 24, 32, 48 or 96GB of VRAM for local AI right now?

The cheapest way to get each amount of VRAM for local AI, priced November 6, 2026: 16GB is an RX 9060 XT 16GB at about $620 new. 24GB is a used RTX 3090 at about $1,591. 32GB is a used AMD Instinct MI50 at $500 to $540 if you run Linux, or a new Radeon AI PRO R9700 from $2,000 if you don't. 48GB is two used RTX 3090s (about $3,180) or one Radeon PRO W7900 at $3,999. 96GB is a 128GB Ryzen AI Max+ 395 mini PC at $3,499.99, which gives its GPU up to 96GB. If you only need the memory for a project, renting is far cheaper: a 24GB RTX 3090 started at $0.10 an hour on our live table.

The cheapest way to each VRAM size, side by side

Prices are US dollars. "Street" prices are dated readings from price trackers and stores, not list prices, and they move every week. Rental prices are the cheapest single-GPU on-demand row for that card on our live cloud GPU table, read November 6, 2026 at 21:20 UTC; in brackets, a cheaper older or bigger card that was also listed.

Two things this table hides. First, memory speed: a mini PC's 96GB is shared system memory, much slower than a graphics card's, so big models answer more slowly. Second, software: the AMD and Intel picks work with llama.cpp and Ollama but not with every CUDA-only tool. Put your model into our VRAM calculator before you buy, so you know which row you actually need.

What's the cheapest 16GB card for local AI?

The RX 9060 XT 16GB, at about $620 new. That was the cheapest in-stock US price TechSpot found in October (published November 2, 2026), against AMD's $349 launch price. TechSpot still called it the cheapest way to get 16GB on a current-generation GPU.

• If you need CUDA: the RTX 5060 Ti 16GB was $790 in the same tracking. Most local-AI tools run first on NVIDIA, so the extra money buys you fewer setup problems. See what the RTX 5060 Ti 16GB can run.
• Intel's option: the Arc Pro B50 16GB was $639 at its lowest US new price on GPUPrix (updated November 4, 2026), up from a 12-month low of $378.
• Used data-center cards: a 16GB Tesla V100 PCIe sold for about $320 to $370 (American Compute, data through August 2026). It has no fans of its own and needs server-style airflow, so it isn't a drop-in desktop card.

Check that a listing says 16GB: both the RX 9060 XT and the RTX 5060 Ti also come in 8GB versions with the same name.

ASUS PRIME Radeon RX 9060 XT 16GB OC

AMD's RDNA 4 card with 16GB of GDDR6. Check that the listing says 16GB, not 8GB.

Best for: 1440p gaming with 16GB of VRAM for less money than the RTX 5060 Ti.

Specs, AI models it runs and rental prices →

Check price on Amazon

Today's price shows on Amazon

ASUS Prime GeForce RTX 5060 Ti 16GB

16GB of GDDR7 with DLSS 4 and Multi Frame Generation, per NVIDIA. It is the 16GB version, not the 8GB one, which is the one that matters for 14B models.

Best for: 1440p gaming, and local AI that needs 16GB of VRAM on a budget.

Specs, AI models it runs and rental prices →

Check price on Amazon

Today's price shows on Amazon

What's the cheapest way to get 24GB of VRAM?

A used RTX 3090, which averaged $1,591 on eBay US on November 6, 2026. ResalePrices tracked 325 listings with a fair asking range of $1,542 to $1,656, and says the average rose 27.1% in 90 days. At about $66 per GB it's not cheap anymore, but it's still the cheapest 24GB card that works with every CUDA tool, fits a normal desktop and has a big community behind it. Our used RTX 3090 buying guide covers what to check before you pay.

• Cheaper, with big caveats: a used Tesla P40 24GB sold for $210 to $250 (American Compute, data through August 2026). It's a 2016 data-center card with no fans and much slower than a 3090. It suits a tinkerer with a server case, not a first build.
• Modded, at your own risk: a memory-modded RTX 2080 Ti with 22GB was listed at about $500 on eBay in August, and a 20GB RTX 3080 at about $445. Both are hand-reworked boards with no factory warranty.
• New, with a warranty: Intel's Arc Pro B60 24GB had a lowest US new price of $1,000 on GPUPrix (updated November 4, 2026). Its 12-month low was $600, so it has risen fast. It's cheaper than a used 3090, but Intel's AI software support is narrower than NVIDIA's.
• Rent it: on November 6, 2026, an RTX 3090 rented from $0.10 an hour on demand on Vast.ai (32 offers, median $0.20) and $0.50 an hour on Runpod. At Vast's median, $1,591 buys roughly 7,900 hours, before you pay for power.

What's the cheapest 32GB GPU?

It depends on whether you run Linux. On Linux, a used AMD Instinct MI50 32GB is by far the cheapest: American Compute says used units "sold for about $500-540 over the last 12 months" (data through August 2026), about $16 per GB. The catch is support. AMD's current ROCm no longer supports the card, so it runs on community-built drivers and llama.cpp builds, and it needs server airflow. Our MI50 guide explains what works and what breaks.

If you want a new card with drivers and a warranty:

• Radeon AI PRO R9700 32GB: lowest US new price $2,000 on GPUPrix (updated November 4, 2026), against a 12-month median of $1,460. It's the cheapest new 32GB card we found.
• RTX 5090 32GB: NVIDIA's own page still says "Starting at $1999", but the cheapest in-stock US card TechSpot found in October was $9,400. Our GPU price check explains why.
• Rent an RTX 5090 instead: from $0.44 an hour on demand on Vast.ai (55 offers, median $0.60) and $1.19 on Runpod, November 6, 2026. Older 32GB V100s rented from about $0.07 an hour on Vast; to own one, see the Tesla V100 adapter build.

What's the cheapest way to get 48GB of VRAM?

Two used RTX 3090s, about $3,180 at the November 6 average. llama.cpp, Ollama and vLLM can split one model across two cards, so two 24GB cards hold models that need about 48GB. You need a motherboard with two usable slots, room for two big cards and a power supply that can feed both. Our multi-GPU guide covers the setup.

One card with 48GB: the Radeon PRO W7900 at $3,999. That was the lowest US new price on GPUPrix (B&H Photo, updated November 3, 2026). It costs about $820 more than two used 3090s, but it's one card in one slot, with a warranty. It is now less than half of TechSpot's October street price for an RTX 5090, with 16GB more memory. The trade-off is software: it runs on AMD's ROCm stack, not CUDA (see AMD vs NVIDIA for local AI).

• NVIDIA's 48GB card: the RTX PRO 5000 Blackwell 48GB had a lowest US new price of $7,985 on GPUPrix (updated November 4, 2026), twice the W7900.
• Linux tinkerers: two used MI50s give you 64GB for about $1,000 to $1,080 at American Compute's range, with the same support caveats as above.
• Rent 48GB: an RTX A6000 rented from $0.31 an hour on demand on Vast.ai (median $0.40) and $0.59 on Runpod, November 6, 2026. An 80GB A100 PCIe was even cheaper on Vast that hour, from $0.27.

What's the cheapest way to get 96GB of VRAM?

A 128GB Ryzen AI Max+ 395 mini PC, at $3,499.99. That was GMKtec's sale price for its EVO-X2 with 128GB and 1TB on November 6, 2026 ($3,999.99 regular). AMD's Variable Graphics Memory lets the built-in GPU use up to 96GB of that memory. It's the cheapest 96GB by a wide margin, about $36 per GB, and it's a whole computer, not just a card.

The catch is speed. The memory is shared system memory, much slower than a graphics card's, so large models generate text slowly. Our Strix Halo VRAM guide and our Mac Studio vs DGX Spark vs Strix Halo comparison show what that means in practice.

The other 96GB options, cheapest first:

• Mac Studio with M5 Ultra: Apple's store says "Buy from $5499", and the base M5 Ultra model comes with 96GB of unified memory and 1TB (read November 6, 2026). Faster memory than the AMD box, at a higher price.
• NVIDIA DGX Spark (128GB): $4,699 at NVIDIA's last announced price (February 25, 2026); HotHardware reported about $6,995 in October.
• Four used RTX 3090s: about $6,360 for the cards alone at the November 6 average, plus a board, case and power supply that can run four.
• RTX PRO 6000 Blackwell 96GB: lowest US new price $17,063 on GPUPrix (updated November 4, 2026), after NVIDIA raised its marketplace price to $16,000 in August. It's the fastest single 96GB card here, and the most expensive. See is the RTX PRO 6000 worth it.

Renting 96GB: on November 6, 2026, an RTX PRO 6000 rented from $1.07 an hour on demand on Vast.ai (the lower-power Max-Q version, 13 offers; the Server Edition started at $1.33, 24 offers, median $1.63) and from $1.64 on Runpod (Max-Q; $2.49 for the standard card). At the Server Edition's $1.63 median, the $17,063 card buys about 10,500 hours of rental. One GB10 machine (the DGX Spark's chip, 128GB) was listed on Vast at $0.28 an hour, a cheap way to try a big model on unified memory first.

Should you buy or rent?

Rent if you'll use the memory for a project, a few evenings a week, or to find out which tier you really need. Buy if you'll run models most days for a year or more, or if your data can't leave your machine. A rough guide from the prices above:

• 24GB: a used 3090 at $1,591 equals about 7,900 hours at Vast's median 3090 rate. At four hours a day, that's more than five years.
• 96GB: an RTX PRO 6000 at $17,063 equals about 10,500 hours at Vast's $1.63 median for the Server Edition. The $3,499.99 mini PC pays back far sooner, if its speed is enough for you.

Both ignore electricity, resale value and setup time. Rental prices change hourly; the cloud GPU page has today's, and cost to run shows the cheapest card for a specific model.

What these prices don't tell you

• Prices move weekly. TechSpot found US graphics card prices up 35% in three months to October. Every price here carries its date; check again before you pay.
• Trackers aren't checkout pages. GPUPrix and ResalePrices show the lowest listing or average they saw. A specific card can cost more, and used listings vary in condition.
• More VRAM isn't more speed. A mini PC's 96GB holds a big model but answers more slowly than a 96GB graphics card. If speed matters, check benchmarks for your model.
• Software matters. AMD and Intel cards run llama.cpp and Ollama well, but some tools still only support CUDA. The MI50 depends on community drivers.

Next: how much VRAM different models need, the best GPU for local AI and prebuilt vs DIY local-AI PCs.