The RTX 5090 and RTX 5080 share a name, a generation, and almost nothing else that matters for local AI. One card has 32GB of VRAM and 1,792 GB/s of memory bandwidth; the other has 16GB and 960 GB/s. That single gap decides which models you can load at all — and in August 2026, the street-price gap between them has stretched to roughly $2,700, which changes who should even be having this debate.

Short version: the 5080 is a fast 16GB card in a market where 16GB is the most crowded tier NVIDIA sells. The 5090 is the only consumer GeForce card that breaks the 16GB ceiling, and it charges accordingly. Whether the upgrade is worth it depends almost entirely on whether the models you want live above or below the 16GB line.

The spec sheet, and what August 2026 did to the prices

	RTX 5080	RTX 5090
VRAM	16GB GDDR7	32GB GDDR7
Memory bandwidth	960 GB/s	1,792 GB/s
Memory bus	256-bit	512-bit
Board power	360W	575W
MSRP	$999	$1,999
Street price (Aug 2026)	~$1,600 average; Amazon listings from ~$1,249	~$4,329 on Amazon; AIB cards $2,900–$5,000+

MSRPs stopped being real prices months ago. The GDDR7 shortage we covered in the RTX 5090 price-hike analysis has only gotten worse: US price trackers put the RTX 5080 at about $1,600 average in early August 2026 after a late-July hike, with the cheapest Amazon listings around $1,249 and premium models like the ASUS ROG Astral at $2,099 — above the 5090’s own MSRP. The 5090 side is uglier. The Founders Edition nominally holds $1,999 but is nearly never in stock; Amazon street pricing sits around $4,329, and flagship AIB cards have crossed $5,000.

So the real-world question isn’t “$999 vs $1,999.” It’s roughly $1,300–$1,600 vs $3,000–$4,300+. Keep that ratio in mind for everything below.

The 16GB line: what each card can actually load

VRAM decides whether a model runs at all. Every weight has to sit in GPU memory, plus a few gigabytes for the KV cache. Overflow even slightly and your runtime spills layers to system RAM over PCIe, where generation speed collapses to single digits. Here’s where current models land against 16GB and 32GB, using file sizes we’ve verified in our individual model guides:

Model	Quant	Weights	RTX 5080 (16GB)	RTX 5090 (32GB)
Llama 3.1 8B	Q4_K_M	~4.9GB	Fits, fast	Fits, faster
Qwen3 14B	Q4_K_M	~9GB	Fits	Fits
Gemma 4 26B-A4B (QAT)	Q4_0	~15GB	Tight fit, short context	Comfortable
Qwen3.6-27B	Q4_K_M	~16.5GB	Spills — drops to ~31 tok/s	Fits with headroom
Gemma 4 31B (QAT)	Q4_0	~18GB	Does not fit	Fits with room for context
Qwen3.6 35B-A3B	Q4_K_M	~20GB	Does not fit	Fits, very fast (MoE)
Llama 3.3 70B	Q4_K_M	~40–42GB	Does not fit	Does not fit either
Llama 4 Scout	Q4	~67GB	Does not fit	Does not fit

Three things worth calling out.

The 27B–35B class is the whole argument for the 5090. Qwen3.6-27B at Q4_K_M needs about 16.5GB before you allocate a single token of context — it’s over the 5080’s limit on weights alone. Gemma 4’s 31B dense model at QAT Q4 wants ~18GB. Qwen3.6 35B-A3B wants ~20GB. This tier is where 2026’s best local models live, and the 5080 can’t hold any of them at standard Q4.

70B changes nothing between these cards. A dense 70B at Q4_K_M is a 40GB+ file. It doesn’t fit the 5080, and it doesn’t fit the 5090 either — you’d need Q3 quantization with quality loss, or you’re spilling to system RAM on both cards. If a circulating comparison claims the 5090 “unlocks 70B,” check whether it means fully in VRAM (it doesn’t) or partially offloaded (slow on either card). And despite what some Reddit threads assume, Llama 4 Scout is no small-model option — its 17B active parameters ride on a 109B total MoE, a ~67GB file at Q4 per our Scout VRAM guide.

The 5080 has escape hatches, with costs. Google’s QAT checkpoints put Gemma 4 26B-A4B in about 15GB — that one genuinely fits a 16GB card, with a shortened context budget. And Qwen3.6-27B at Q3_K_M (~13GB) fits, trading measurable quality. The 5080 isn’t locked out of the mid tier; it’s locked out of the mid tier at full Q4 quality with room to breathe.

What the spill actually looks like (and the fix)

If you try the tight fits anyway, this is the failure mode. Pull Qwen3.6-27B Q4_K_M on a 16GB card with Ollama v0.30.x and check placement:

$ ollama run qwen3.6:27b "Explain KV cache in one paragraph." 
# ...noticeably sluggish generation...
$ ollama ps
NAME           ID            SIZE     PROCESSOR         UNTIL
qwen3.6:27b    a1b2c3d4e5f6  19 GB    18%/82% CPU/GPU   4 minutes from now

That 18%/82% CPU/GPU split is the tell: the model plus KV cache didn’t fit in 16GB, so Ollama parked part of it in system RAM. When we benchmarked this exact configuration for the 27B guide, generation ran at roughly 31 tok/s on a 960 GB/s-class card — versus ~40 tok/s fully resident on a 24GB RTX 3090 with far less bandwidth. PCIe becomes the bottleneck; the 5080’s bandwidth advantage evaporates.

The fix on a 5080 is one of: drop to Q3_K_M (ollama run qwen3.6:27b-q3_K_M), switch to the Gemma 4 26B-A4B QAT build, or cap context (num_ctx 8192) to shrink the KV cache until ollama ps reports 100% GPU. On a 5090, you skip this entire paragraph — 27B Q4 plus long context is resident with ~13GB to spare.

Speed where both cards can play: the bandwidth math

For models that fit both cards, LLM decode speed is memory-bandwidth-bound, and the ceiling is easy to estimate:

theoretical max tok/s ≈ memory bandwidth ÷ model weight size (GB)

An 8B Q4_K_M model is about 4.9GB. The 5080’s 960 GB/s gives a ~196 tok/s ceiling; the 5090’s 1,792 GB/s gives ~366. Real measurements land well under the ceiling but preserve the ratio:

Model	RTX 5080 (measured)	RTX 5090 (measured)
Llama 3 / 3.1 / 3.2 8B, Q4_K_M	~103 tok/s (llama.cpp), up to ~132 (tuned)	~142 tok/s (Ollama), up to ~213 (tuned)
Gemma 4 (mid-size), Q4	~47 tok/s at 11.2GB VRAM	~241 tok/s (26B-A4B MoE)
14B, Q4_K_M	~58 tok/s	~90–110 tok/s (bandwidth-scaled)

The 5090 runs the same small models roughly 1.5–1.9× faster, tracking its 1.87× bandwidth edge. Whether that matters depends on what you’re doing. Human reading speed is ~7–10 tok/s, so a 5080 pushing 103 tok/s on an 8B model is already 10× faster than you can read — for interactive chat, the 5090’s extra speed is imperceptible. Where it pays is agentic and batch work: coding agents that make dozens of model calls per task, RAG pipelines, or serving a local BYOK backend for Cline or Continue.dev where total tokens per session run into the hundreds of thousands. There, 1.9× throughput is 1.9× less waiting, every time.

One more 5090-only trick: 32GB lets you keep two mid-size models resident at once — a 14B coder plus an 8B chat model, or A/B eval pairs — without the constant model-swapping reload penalty a 16GB card forces.

Price per performance: the part that hurts

At August 2026 street prices, the 5090 costs about 2.7× the 5080 for 1.87× the bandwidth and 2× the VRAM. Against its own MSRP it’s marked up over 100%. So before paying it, check the alternatives:

• Used RTX 3090, ~$1,050. Still the value answer for the 24GB tier: 936 GB/s and enough VRAM for the entire 27B–35B class the 5080 can’t hold — Qwen3.6 35B-A3B at ~107 tok/s. It’s slower than a 5090 and carries no warranty, but it costs a quarter of one and less than a new 5080 while running strictly more models.
• Used RTX 4090, ~$2,268–$2,500. 24GB and ~127 tok/s on 8B models. The oft-repeated “two used 4090s for $2,600 total” plan is a fantasy at August 2026 prices — a pair runs $4,500–$5,000, more than a single 5090 at Amazon pricing, plus you’re managing multi-GPU inference at 900W.
• Rent the 32GB tier first. An RTX 5090 on RunPod costs well under a dollar an hour. If you’re unsure whether your workload actually needs 30B-class models, a $20 weekend of renting answers it before you commit $4,000.

For the power-bill math: at $0.12/kWh, the 5090’s 575W board power costs about $0.069/hour under load versus $0.043 for the 5080’s 360W — noise, not signal, next to the purchase-price gap, though the 5090 will also want a bigger PSU (1,000W+ recommended system total).

Image generation splits the same way

The pattern repeats in ComfyUI. FLUX.1-dev at FP8 needs about 12GB — both cards run it, and both benefit from Blackwell’s NVFP4 path in ComfyUI. The full BF16 checkpoint is a ~24GB file: 5090 only. On a 5080 you’ll run quantized image models happily and never touch full-precision checkpoints; video models like Wan land the same way, with the heavier variants needing more than 16GB.

Honest take

Buy the RTX 5090 if your target models live above 16GB and you’ll use its speed — Qwen3.6-27B/35B-class LLMs at full Q4 with long context, BF16 image checkpoints, multi-model setups, or high-volume agentic coding. It’s the only new consumer card that does this, and at ~$3,000–$4,300 street you’re paying monopoly rent for that fact. It is genuinely excellent; it is nowhere near fairly priced.

Skip the RTX 5080 for local AI at August 2026 prices. That’s not because it’s slow — 103+ tok/s on 8B models is the fastest 16GB experience you can buy — but because ~$1,600 is 24GB-money in a market where a used 3090 costs $1,050 and holds strictly more models. The 5080 makes sense only if you also game heavily, want a warranty, and are certain you’ll stay in the 8B–14B tier (plus QAT builds) for the card’s lifetime. If you’re VRAM-limited and the 5090’s price is absurd to you — it should be — the used 24GB tier or renting 32GB by the hour covers the gap for a fraction of the cost.

The 32GB vs 16GB question, answered plainly: yes, it changes which models you can run — the entire 2026 mid-tier sits between those two numbers. Whether that’s worth $2,700 extra is a question about your workload, not the silicon.

FAQ

Can the RTX 5090 run a 70B model? Not fully in VRAM. Llama 3.3 70B at Q4_K_M needs ~40–42GB against the 5090’s 32GB. You’d run Q3 with quality loss or offload layers to system RAM at a large speed penalty. For dense 70B+, the answer is a used dual-GPU rig, a Mac with enough unified memory, or the cloud.

Is the RTX 5080 good for Stable Diffusion and FLUX? Yes — SDXL and FLUX.1 FP8 (~12GB) run comfortably. What 16GB rules out is full-precision BF16 checkpoints (~24GB) and the largest video models.

Does the 5090’s 1.9× bandwidth matter for chat? Barely. Both cards generate far past reading speed on models that fit both. The speed premium pays off in agentic workflows, batch processing, and multi-call coding assistants — not in a chat window.

Should I wait for a 5080 Ti or Super with more VRAM? Supply reporting through 2026 points the other way: the 3GB-module GDDR7 shortage pushed RTX 50 SUPER variants to late 2026 at the earliest, and NVIDIA has thinned the existing lineup rather than expanded it. Buying decisions made on rumored VRAM bumps have lost all year.

Is a used RTX 3090 really better than a new RTX 5080 for LLMs? For LLMs specifically, mostly yes: 24GB > 16GB decides model access, and the 3090’s 936 GB/s is close to the 5080’s 960. The 5080 wins on warranty, power draw, FP4 support, and gaming. If the machine is a dedicated inference box, the $1,050 used 3090 is the better buy; we run the full comparison in the used RTX 3090 value analysis.

Sources

• RTX 5080 price tracker, US average ~$1,600 — GPU Price History
• RTX 5080 price tracker, August 2026 — videocardprices.com
• Gaming GPU prices 2026: RTX 5090 tops $5,000, ~$4,329 on Amazon — Tech Insider
• NVIDIA flagship price hikes at Best Buy; RTX 5080 above RTX 5090 MSRP — The FPS Review, Jul 30 2026
• RTX 5090 price tracker: MSRP vs actual — StockMaid
• llama.cpp CUDA performance: RTX 5080, Llama 3.1 8B Q4_K_M at 103 tok/s — ggml-org GitHub discussion #15013
• Gemma 4 on RTX 5080: 47 tok/s, 11.2GB VRAM — Markaicode benchmarks
• RTX 5080 16GB local LLM results, 14B at ~58 tok/s — ModelFit
• RTX 5080 accelerator results — LocalScore
• RTX 5090 vs 5080 for local AI: 213 vs 132 tok/s on 8B — Local AI Master
• Used RTX 4090 fair asking range ~$2,500 — ResalePrices
• Used RTX 4090 eBay pricing ~$2,268, August 2026 — Best Value GPU
• GeForce RTX 5090 specifications: 32GB GDDR7, 575W — NVIDIA
• GeForce RTX 5080 specifications: 16GB GDDR7, 360W — NVIDIA

Last updated August 6, 2026. GPU prices in 2026 move weekly; verify current listings before purchasing.