TL;DR: The RTX 5080 is the fastest 16GB card you can buy — 960 GB/s of GDDR7 bandwidth, more than a used RTX 3090’s 936. But at $1,580 street in September 2026, it costs roughly 50% more than that 3090, which brings 24GB and runs a whole model class the 5080 can’t load. Buy it for speed on models that fit; don’t buy it to run 27B models.

	RTX 5080 (new)	Used RTX 3090	RTX 5060 Ti 16GB
Best for	Fastest decode on ≤20B models, gaming + AI dual use	Cheapest ticket into the 27B–35B class	Cheapest new 16GB card
Price (Sep 2026)	~$1,580–$1,700 new	~$1,000–$1,050 eBay	~$570 lowest, ~$805 median
VRAM / bandwidth	16GB / 960 GB/s	24GB / 936 GB/s	16GB / 448 GB/s
The catch	16GB ceiling at a 24GB price	2020 silicon, no warranty, 350W	Half the 5080’s speed

Honest take: if your goal is running the best local model per dollar, a used RTX 3090 is still the answer — $550 less and it loads Qwen3.8-27B, which the 5080 cannot. Buy the 5080 only if you want one new card with a warranty that games at 4K and tears through 8B–20B models faster than anything else in its VRAM class.

The RTX 5080 has spent 2026 in an odd spot for local AI. On paper it’s the best 16GB card ever made. In practice, the GDDR7 shortage has pushed its street price into territory where it competes with cards that carry 8GB more memory — and for local LLMs, capacity decides what you can run before bandwidth decides how fast. Before committing, run your target model and context through our VRAM calculator — the fit math below is the whole decision.

The spec sheet really is the best 16GB has

Verified against NVIDIA’s product page and Wccftech’s spec roundup:

Spec	RTX 5080	Why it matters for LLMs
VRAM	16GB GDDR7 (30 Gbps)	Sets the hard ceiling on model + context
Memory bandwidth	960 GB/s (256-bit)	Decode speed on dense models scales with this
CUDA cores	10,752 (Blackwell)	Prompt processing, batch work
TBP	360W	~$0.043/hour under load at $0.12/kWh
MSRP	$999	Fiction since launch — see below

That 960 GB/s figure is the headline. It beats the RTX 5070 Ti (896 GB/s), the RTX 4080 Super (736), the 5060 Ti (448) — and, narrowly, the used RTX 3090 (936). Among every card with 16GB or less, nothing new or used moves memory faster, and memory speed is what token generation is bottlenecked on for dense models. Our 50-series model guide covers the whole lineup; this page is about whether the fastest 16GB card makes sense when 16GB is the constraint.

The September 2026 price is the problem

The $999 MSRP has not been a real number all year. As of September 5, 2026, GPU PRIX tracks the lowest new US listing at $1,580, and videocardprices.com’s September tracker shows Newegg at $1,579.99, Amazon at $1,599.99, and Best Buy at $1,699.99, with used eBay units around $1,399.99. That’s roughly 58% over MSRP, and the driver is the same GDDR7/DRAM crunch that has reshaped every GPU budget this year — memory now dominates the bill of materials, and analysts don’t expect relief before late 2027.

Meanwhile a used RTX 3090 sells for about $1,050 per BestValueGPU’s September tracker (recent eBay sales near $1,000 per GPUDojo; active asking prices run higher, $1,287–$1,411 per ResalePrices — patient buyers pay sold-listing prices, not asking prices). The 5080 also sits close enough to the RTX 5090’s orbit to sting: yes it’s less than half the 5090’s street price, but the 32GB vs 16GB question is a different article. The fight that matters at this price is 16GB-new versus 24GB-used.

What fits in 16GB — the ceiling didn’t move

Speed is irrelevant for a model you can’t load. Here’s the September 2026 fit map, using GGUF sizes we’ve verified across the 16GB tier guide:

Model	Quant / file size	Fits 16GB?
Qwen3-14B	Q4_K_M, ~9GB	Yes, with real context headroom
GPT-OSS 20B	MXFP4, ~12–13GB	Yes
Codestral 2 22B	Q4_K_M, 13.3GB	Yes
Devstral Small 2 24B	Q4_K_M, 14.7GB	Only with KV-cache quantization (~17GB at 8K otherwise)
Gemma 4 26B-A4B QAT	~15GB	Yes, barely — short context
Qwen3.8-27B	Q4_K_M, 16.8GB	No
Qwen3.6-35B-A3B	Q4_K_M, ~21GB	No

The pattern: everything that fits a $570 RTX 5060 Ti 16GB fits the $1,580 RTX 5080, and nothing more. The extra $1,000 buys speed on the same list, not a longer list. That’s the entire buying decision compressed into one row of the table — the 27B line sits at 16.8GB, and Qwen3.8-27B’s own release math puts even the smallest 4-bit quant (IQ4_XS, 15.4GB) at 0.6GB of headroom — less than the KV cache costs at any context length you’d actually use.

Measured speeds: where the 5080 earns its keep

On models that fit, the numbers are excellent:

• 8B class: 132 tok/s, measured in LocalAIMaster’s 5090-vs-5080 testing (the 5090 hits 213 — 60% faster, for ~3× the price).
• 14B class at Q4: ~58 tok/s per ModelFit’s RTX 5080 page, which calls it the fastest 16GB card outright.
• Gemma 4 at Q4_K_M: 47 tok/s steady-state in Markaicode’s Ollama benchmark (Ollama v0.5.7; their run logged 11.2GB of VRAM at Ollama’s short 512-token default context — budget more VRAM as context grows).
• GPT-OSS 20B: no direct 5080 measurement we could verify, but the same MoE was clocked at 139.93 tok/s on an RTX 4080 16GB (717 GB/s). Scaling by the 5080’s 34% bandwidth advantage puts it in the ~160–185 tok/s range — treat that as a bandwidth-derived estimate, not a measurement. Our GPT-OSS 20B guide has the model’s context trap.

For coding-agent workloads — Cline or Continue.dev pointed at a local backend (setups covered on aicoderscope.com) — those 14B–22B speeds are past the threshold where the model, not the wait, is your bottleneck. Micro Center’s 5080 LLM benchmark write-up lands on the same conclusion: quantization choice, not compute, decides what this card can do.

The 27B trap: what spilling actually looks like

The most common mistake people make with this card: pulling Qwen3.8-27B Q4_K_M anyway, because 16.8GB sounds close enough to 16GB. It isn’t, and the failure is silent. Ollama won’t error — it splits the model between VRAM and system RAM, and decode speed falls off a cliff. Check what actually happened:

$ ollama ps
NAME              ID            SIZE     PROCESSOR          UNTIL
qwen3.8:27b       a1c2e3f4      18 GB    22%/78% CPU/GPU    4 minutes from now

Anything other than 100% GPU in that column means you’re decoding through your CPU’s memory bus, and a card with 960 GB/s of bandwidth is idling while DDR5 at a tenth of that does the work. Community reports of 27B runs on the 5080 in the low teens of tok/s are this mechanism — partial offload, not the card’s speed. The fix is not a flag; it’s picking a model from the fit table, stepping up to 24GB, or renting. If you’re determined to squeeze the tier’s biggest files (Devstral, Gemma 4 26B QAT), two service-level settings buy real room: OLLAMA_FLASH_ATTENTION=1 and OLLAMA_KV_CACHE_TYPE=q8_0 roughly halve KV-cache VRAM at near-zero quality cost, and capping num_ctx at 32768 keeps the cache from eating what’s left. Full recipe in the tokens-per-second tuning guide.

That’s also the honest way to read the 5080’s long-context appeal: with a ~13GB model file, quantized KV cache, and 960 GB/s to keep decode usable as the cache grows, this card sustains 30K-plus contexts on 20B-class models more gracefully than any other 16GB card. What it cannot do is hold 27B weights and a working cache at once.

RTX 5080 vs used RTX 3090: the $550 question

	RTX 5080 ($1,580 new)	Used RTX 3090 (~$1,050)
VRAM	16GB	24GB
Bandwidth	960 GB/s	936 GB/s
Qwen3.8-27B Q4_K_M	Spills — unusable	~41 tok/s, fully resident
8B–14B models	Fastest in class	~3% slower on paper
Warranty / condition	New, 2–3 years	None, 2020 silicon
Power	360W	350W

On bandwidth the cards are effectively twins — 2.5% apart. On capacity they’re a class apart: the 3090 runs the 27B–35B tier that current open-weight releases keep landing in, with context to spare. The 3090 has held this crown through three GPU generations for exactly this reason, and our full buyer’s guide by budget reaches the same verdict at the $1,000 tier. Paying more for the 5080 is paying a premium to run smaller models faster — defensible for some buyers, but you should know that’s the trade.

Who should actually buy the 5080

• The dual-use buyer. You want one new card that handles 4K gaming, DLSS 4, and local AI. The 3090 loses this comparison badly — the 5080 is ~2 generations newer in raster and features, and it’s the best gaming card that also happens to top the 16GB AI class.
• The warranty-required buyer. No used market, corporate purchase, or you’ve been burned by a mined-on card. The 5080 is the fastest 16GB you can buy with a receipt.
• The 20B-and-under speed maximalist. Your daily models are GPT-OSS 20B, Codestral 2, Qwen3-14B — and you want the lowest latency per token available under $2,000. Nothing at 16GB touches it.

Everyone else: buy the used 3090, or spend $570 on a RTX 5060 Ti 16GB that runs the identical model list at half speed for a third of the price. Whichever card you land on, the software side is the same — aifoss.dev covers the Ollama + Open WebUI stack that all three run. And if you’re not sure the 20B class is enough for your workload, rent before you buy — an hour on a 24GB cloud GPU via RunPod running Qwen3.8-27B tells you whether the bigger class matters for your use case before you commit $1,000+. The rent-vs-buy math is its own article.

Recommended Gear

• RTX 5080 16GB — the fastest 16GB card for local AI; ~$1,580–$1,700 new as of Sep 2026
• Used RTX 3090 24GB — $550 less, loads the 27B class the 5080 can’t; ~$1,000–$1,050 on eBay
• RTX 5060 Ti 16GB — same 16GB model list at half the speed, from ~$570

FAQ

Is the RTX 5080 good for local LLMs? For models that fit in 16GB — the 8B to 22B class — it’s the fastest option, period: 132 tok/s on 8B models, ~58 tok/s on 14B at Q4, 47 tok/s on Gemma 4. The limitation is the list of models, not their speed.

Can the RTX 5080 run Qwen3.8-27B? Not resident in VRAM. Q4_K_M is 16.8GB against a ~15.5GB usable budget, and the smallest 4-bit (15.4GB) leaves no room for KV cache. It will “run” via CPU spillover at low-teens tok/s — check ollama ps for the CPU/GPU split. A 24GB card runs it at ~41 tok/s.

RTX 5080 or used RTX 3090 for AI? Same bandwidth (960 vs 936 GB/s), different ceilings (16 vs 24GB), $550 apart in the wrong direction. For AI alone, the 3090. For gaming plus AI on a warranty, the 5080.

Why is the RTX 5080 so expensive in late 2026? GDDR7 supply. The lowest US street price was $1,580 on September 5, 2026 against a $999 MSRP, and trackers show the premium persisting since spring. Analysts tie relief to new memory capacity that doesn’t arrive before late 2027.

Sources

• GeForce RTX 5080 official specs — NVIDIA
• NVIDIA RTX 5080 full specs and price roundup — Wccftech
• GeForce RTX 5080 price history, lowest US listing $1,580 (Sep 5, 2026) — GPU PRIX
• RTX 5080 price tracker, September 2026 retailer listings — videocardprices.com
• RTX 3090 price tracker US, September 2026 — Best Value GPU
• RTX 3090 24GB used price and history — GPUDojo
• RTX 3090 used fair asking range — ResalePrices
• RTX 5090 vs 5080 for local AI, measured tok/s — LocalAIMaster
• RTX 5080 16GB local LLM benchmarks — ModelFit
• Gemma 4 on RTX 5080: 47 tok/s Ollama throughput benchmark — Markaicode
• Comparing LLM performance on a 16GB VRAM GPU (GPT-OSS 20B, 139.93 tok/s on RTX 4080) — Rost Glukhov
• Benchmarking AI on an RTX 5080 — Micro Center
• Devstral-Small-2-24B GGUF sizes (Q4_K_M 14.7GB) — Bartowski on Hugging Face
• Devstral Small 2 hardware requirements (~17GB at 8K context) — Hardwarepedia
• Qwen3.8-27B VRAM requirements — OrcaRouter
• RTX 5080 costs 39% more than 5070 Ti, runs 17% faster — TechTimes

Last updated September 7, 2026. Prices and specs change weekly in this market; verify current rates before purchasing. The GPT-OSS 20B RTX 5080 figure is a bandwidth-derived estimate from a measured RTX 4080 result, flagged as such above.