Two used RTX 3090s or one RTX 5090 for local AI: which should you buy?

Buy two used RTX 3090s if you want to run 70B-class models; buy one RTX 5090 if everything you run fits in 32GB. Two 3090s give you 48GB of VRAM for about $2,000 (hardware-corner.net lists a $1,000 average US market value per card, read Nov 3, 2026), enough for Llama 3.3 70B at Q4, which no single 32GB card can hold. One 5090 runs models that fit in 32GB about twice as fast, at 575W instead of 700W for the pair, with none of the multi-GPU setup. The catch is its price: NVIDIA's Founders Edition lists at $1,999 but was out of stock on NVIDIA's US store on Nov 1, 2026, and hardware-corner.net's average US market value for a 5090 was $3,699 on Nov 3, 2026.

Which is cheaper per GB of VRAM?

Two used 3090s, by a wide margin. VRAM is what decides which models load, so cost per GB is the number that matters most for local AI.

hardware-corner.net's averages carry no date of their own, so treat them as a guide and check recently sold listings before you pay. The used 3090 price is for one card in unknown condition; our used RTX 3090 buying guide has the checks to run on each one.

What can 48GB run that 32GB can't?

A dense 70B model at a normal 4-bit quantization. That's the whole case for the dual-3090 build. By our VRAM calculator's formula, Llama 3.3 70B at Q4_K_M needs about 40GB for its weights, plus about 2.5GB of KV cache at 8K context and under 1GB of overhead: roughly 43GB. That fits in two 24GB cards with a little room to spare, and doesn't fit in 32GB. hardware-corner.net's 3090 FAQ puts a 70B model at "48–80 GB VRAM depending on quantization."

On a single 5090 you can still load a 70B, but only at about 3-bit (IQ3_XS comes to around 30GB at 8K context in our formula), which leaves almost no room for longer context and costs more quality than Q4. For models up to about 32B, both setups work: Qwen3 32B at Q4_K_M is about 18.5GB of weights, so it fits on one 3090 and on the 5090 with plenty of room for long context. Check your exact model on /cost-to-run/llama-3.3-70b or /cost-to-run/qwen3-32b.

If the biggest models you want are 32B or smaller, the 48GB buys you nothing you'll use, and the 5090's speed is the better deal.

Are two 3090s faster than one 5090?

No. Two cards give you more memory, not more speed for a single chat. hardware-corner.net's relative index puts one RTX 3090 at 51% of an RTX 5090's token-generation speed; on its Qwen3 14B Q4 test the 5090 is listed at 123.8 tokens a second at 4K context and the 3090 at 70.0. A model split across two 3090s still runs at roughly 3090 speed, because of how the split works.

The cleanest like-for-like numbers come from llama.cpp's own community CUDA scoreboard: llama-bench runs of one pinned model on a single GPU. We don't run our own benchmarks, so these are other people's measurements.

Two things fall out of that table. Writing a reply tracks memory bandwidth: the 5090 has 1.91 times the 3090's bandwidth (1,792 GB/s against 936 GB/s) and writes 1.86 times as fast. Reading a prompt tracks compute: there the 5090 is 2.69 times faster, far more than bandwidth predicts. If you feed it long documents, whole codebases or agent loops, that prefill gap is the strongest single reason to pick the 5090.

• llama.cpp (and the tools built on it, such as Ollama and LM Studio): the default --split-mode layer "split layers and KV across GPUs (pipelined)", per its README. The cards take turns, so a single chat runs at about the speed of one card. --tensor-split 1,1 sets how much goes on each card. The row and tensor modes run the cards in parallel; llama.cpp marks tensor as EXPERIMENTAL.
• vLLM: its docs say that if "the model is too large for a single GPU but fits on a single node with multiple GPUs, use tensor parallelism", with --tensor-parallel-size 2 for two cards. That runs both GPUs on each step and suits serving several requests at once, but vLLM takes more setup than Ollama.

The 3090 supports an NVLink bridge (NVIDIA lists "NVIDIA NVLink (SLI-Ready)" in its spec table); llama.cpp's options above work over the normal PCIe slots. If you value speed on models that fit in 32GB, one 5090 is the faster machine.

How much power and PSU do two 3090s need?

More than one 5090. NVIDIA lists 350W of graphics card power for each RTX 3090, so 700W for the pair, against 575W for an RTX 5090. NVIDIA's required system power is 750W for a single 3090 and 1000W for a 5090. For two 3090s, hardware-corner.net's FAQ says "Multi-RTX 3090 setups commonly require power supplies in the 1200 W to 1600 W range."

Should I rent first?

Yes, if you're not sure which you need, or you'd only run the card a few hours a week. When our tracker read prices at 15:20 UTC on Nov 3, 2026, a single 24GB RTX 3090 rented on Vast.ai from $0.11 an hour on-demand and a 32GB RTX 5090 from $0.45 an hour; on Runpod the same cards were $0.50 and $1.19 an hour. An evening on each tells you whether your model fits and whether the speed is enough. Prices change hourly: /cloud-gpu shows today's, and our rent vs buy break-even works out when buying starts to pay.

Prices were read on hardware-corner.net on Nov 3, 2026 (its averages carry no date) and on NVIDIA's US marketplace on Nov 1, 2026. Specs and power figures are from NVIDIA's own pages, and the multi-GPU behavior from llama.cpp's and vLLM's own documentation, read Nov 3, 2026. VRAM figures come from our calculator's formula, not a test we ran; we haven't built this machine ourselves. Rental prices are from our tracker of Vast.ai and Runpod at 15:20 UTC on Nov 3, 2026. The Vast.ai buttons are referral links, and we may earn a fee if you rent through them; the advice above often says to buy instead.

For every other route through the VRAM problem, from a 12GB card to workstation GPUs, see the best GPU for local AI. For the single-card case on its own, see is the RTX 5090 worth it for local AI, and for big MoE models that run on one 24GB card plus system RAM, see what --n-cpu-moe actually does.