GPU SELECTION → · 2026-11-04 · 13 min read
Best GPU for video generation in 2026
A video model doesn’t bill you by the token. It bills you by the second of card time — so the right question is never “which GPU is fastest?” but “which is the smallest card that runs this model well, and what does one finished clip cost on it?” This guide answers that for the six open-weight video models people actually rent cards for in 2026 — Wan 2.2 in both sizes, HunyuanVideo-1.5, LTX-2, CogVideoX-1.5 and Mochi 1 — with verified prices checked Nov 4, 2026, and a cost per 5-second clip wherever the maker publishes a generation time. Where a maker publishes none, we price the model by the hour and say so, rather than invent a number.
The verdict
- HunyuanVideo-1.5 (step-distilled build, 480p) — RTX 4090 (24GB), RunPod Community $0.34/hr. Tencent reports the distilled build makes a ~5-second clip in 75 seconds on a single 4090 — about $0.007 a clip. The full 50-step base model runs on the same card at the same price, slower per clip; Tencent publishes no base timing.
- Wan 2.2 TI2V-5B (720p) — RTX 4090 (24GB), RunPod Community $0.34/hr. Wan’s own table measures 534.7 seconds and 22.9GB peak for a 5-second 720p clip on one 4090 — about $0.05 a clip. The default starting point for rented video work.
- Wan 2.2 T2V-A14B (720p) — A100 80GB, Thunder Compute $1.09/hr. Wan’s single-GPU run needs an 80GB-class card; its measured peak at 720p is 59.8GB. 2,735.7 seconds a clip on the A100 — about $0.83; on an H100 SXM at $2.69/hr, 1,041.5 seconds — about $0.78.
- CogVideoX-1.5-5B (1360×768) — A100 80GB, Thunder Compute $1.09/hr. The maker publishes ~1,000 seconds per 5-second video on one A100 at 50 steps — about $0.30 a clip; ~550 seconds on an H100 works out to ~$0.41.
- Mochi 1 (480p) — A100 80GB, Thunder Compute $1.09/hr. The official route needs ~60GB of VRAM on one card. Genmo publishes no generation time, so there is no honest cost per clip — priced per hour only.
- LTX-2 (video with sound) — A100 80GB, Thunder Compute $1.09/hr. Lightricks publishes neither a VRAM floor nor a generation time; the 80GB pick is our sizing from its checkpoint files (the bf16 transformer alone is a 43.3GB file), explained in the guide. Priced per hour only.
Prices for the cards in this guide
The board's current prices — checked nov 15, 2026.
TensorDockH100 SXM · 80GB 8-GPU minimum Best offer$18.00 / hr total Why video is different: you rent seconds, not tokens
Our language-model guides price work per token, because a language model’s cost really does scale with tokens in and out. A video model has no tokens to count. Its cost is set by two numbers: whether it fits — the VRAM it needs at the resolution and length you actually render — and how long a clip takes — the seconds of card time between pressing run and holding a finished file. Multiply the second number by the card’s hourly price and you have the only figure that matters: cost per finished clip. That is also why video guides are so easy to get wrong. Makers publish memory figures measured with different switches on — offloading to system RAM, tiling, reduced precision — and most publish no generation time at all. A “fits in 14GB” claim and a “runs well” claim can both be true of the same model while describing experiences an hour apart. So this guide holds a hard line: every timing below is published by the model’s own maker and labelled maker-reported; every price is from our board check of Nov 4, 2026; and where a maker publishes no timing — Mochi 1 and LTX-2 — the model is priced per hour, with the reason stated, instead of a clip cost built on air. One boundary before the tables: the newest Wan you can download is Wan 2.2. Wan-AI’s Hugging Face organisation, checked Nov 4, 2026, lists no Wan 2.5 or later weights — those versions are API-only, so if you want them you are buying API credits, not renting a card. Wan 2.1’s 14B is covered by its successor — Wan 2.2’s A14B is the same job, newer weights, same Apache 2.0 licence. (Wan’s own card notes the small Wan 2.1 1.3B makes a 5-second 480p clip on an RTX 4090 in about 4 minutes — fine for a first experiment, not a production pick.)
| GPU | VRAM | Bandwidth | Lowest verified price | Verdict |
|---|
| RTX 4090 | 24GB | 1,008 GB/s | $0.34/hr — RunPod Community | The small-model card: Wan 2.2’s 5B and distilled HunyuanVideo-1.5 live here, with maker timings measured on this very card. Nothing bigger fits honestly. |
| RTX 5090 | 32GB | 1,792 GB/s | $0.69/hr — RunPod Community | More speed and headroom for the 5B-class models; still too small for the 14B and 19B builds without heavy offloading. |
| L40S | 48GB | 864 GB/s | $0.79/hr — RunPod Community | The offload tier: Mochi 1’s 42GB diffusers route and LTX-2’s FP8 checkpoint fit here — workable, not quick. |
| A100 80GB | 80GB | 2,039 GB/s | $1.09/hr — Thunder Compute | The workhorse of this guide: every model here runs on one, at the cheapest verified 80GB price on the board. |
| H100 SXM | 80GB | 3,350 GB/s | $2.25/hr — TensorDock | The same memory, roughly 2–3× the A100’s speed on the makers’ own tables. Buys turnaround, not cheaper clips. |
| RTX PRO 6000 | 96GB | 1,792 GB/s | $1.69/hr — RunPod Community | The most single-card memory under $2/hr: room for the big builds plus an upscaler in the same session. |
What each model needs — from the makers’ own cards
“Runs it well” in this guide means the full model at the resolution people actually use, on one card, without shrinking the job to fit — not the maker’s absolute minimum with every memory-saving switch on. The minimums are real, and they are in this table, next to the cards we would actually rent:
| Model | Size | Licence | The maker’s memory figure | Smallest card that runs it well |
|---|
| Wan 2.2 TI2V-5B | 5B, dense | Apache 2.0 | “At least 24GB VRAM” with offloading (card); measured peak 22.9GB at 720p | RTX 4090 |
| HunyuanVideo-1.5 | 8.3B | Tencent Hunyuan Community licence | 14GB minimum with model offloading (card) | RTX 4090 |
| Wan 2.2 T2V-A14B | 27B total, ~14B active (MoE) | Apache 2.0 | Single-GPU run needs an 80GB-class card (card); measured peak 59.8GB at 720p with offload | A100 80GB |
| CogVideoX-1.5-5B | 5B | CogVideoX licence (code Apache 2.0) | 76GB for the full BF16 route; the diffusers route runs “from 10GB” with every memory optimisation on (maker’s README) | A100 80GB |
| Mochi 1 | 10B | Apache 2.0 | ~60GB VRAM for the official single-GPU route; the maker recommends at least one H100; its diffusers example needs 42GB with offloading and tiling | A100 80GB |
| LTX-2 | 19B | LTX-2 Community License | No VRAM floor published. The bf16 transformer file alone is 43.3GB; the FP8 checkpoint is 27.1GB (repository file listing) | A100 80GB — our sizing, see below |
The pattern to notice: the honest dividing line in video is not 24GB versus 48GB — it is 24GB versus 80GB. The two small models (Wan’s 5B and HunyuanVideo-1.5) are built to live on a 4090-class card, and their makers prove it with timings measured on that card. Everything bigger either states an 80GB-class requirement outright (Wan’s A14B, CogVideoX’s full-precision route, Mochi’s official route) or, like LTX-2, ships checkpoint files so large that an 80GB card is the sensible planning unit. The 48GB cards (L40S, RTX A6000) earn their place in exactly one way: the offload routes — Mochi’s 42GB diffusers example fits there, and LTX-2’s FP8 checkpoint is the lighter way into that model. On LTX-2’s sizing, plainly: Lightricks publishes no minimum VRAM and no speed on the LTX-2 card, and its own example pipeline offloads parts of the model to the CPU. What can be checked is the weight of the download — the bf16 transformer checkpoint is a 43.3GB file, the FP8 build 27.1GB, the NVFP4 build 20.0GB — and the text encoder those files pair with is heavy in its own right. We would not plan a production run on less than an 80GB card for the bf16 build, and we label that as our sizing from the published files, not a maker specification. Lightricks has also since shipped LTX-2.3 (the card flags it as the newer version); every figure here is for the LTX-2 checkpoints listed.
What a finished clip costs — maker-reported times only
Every row is arithmetic on a timing the model’s maker publishes, at a card price from our board check of Nov 4, 2026: clip cost = seconds per clip ÷ 3,600 × $/hr. The clips are the makers’ own reference jobs — a 5-second clip at each model’s standard settings — so the rows compare fairly with each other, and none of them is a benchmark we ran:
| Model + setting | Card | Time per clip (maker-reported) | Card price | Cost per clip |
|---|
| HunyuanVideo-1.5, step-distilled, 480p image-to-video | RTX 4090 | 75s | $0.34/hr | $0.007 |
| Wan 2.2 TI2V-5B, 720p text-to-video | RTX 4090 | 534.7s | $0.34/hr | $0.05 |
| Wan 2.2 T2V-A14B, 480p | A100 80GB | 785.7s | $1.09/hr | $0.24 |
| Wan 2.2 T2V-A14B, 480p | H100 SXM | 326.9s | $2.69/hr | $0.24 |
| CogVideoX-1.5-5B, 1360×768, 50 steps | A100 80GB | ~1,000s | $1.09/hr | $0.30 |
| CogVideoX-1.5-5B, 1360×768, 50 steps | H100 SXM | ~550s | $2.69/hr | $0.41 |
| Wan 2.2 T2V-A14B, 720p | H100 SXM | 1,041.5s | $2.69/hr | $0.78 |
| Wan 2.2 T2V-A14B, 720p | A100 80GB | 2,735.7s | $1.09/hr | $0.83 |
| Mochi 1, 480p | A100 80GB | Not published by the maker | $1.09/hr | Priced per hour |
| LTX-2, two-stage pipeline | A100 80GB | Not published by the maker | $1.09/hr | Priced per hour |
Three things this table teaches. First, the cheap card wins more often than you’d expect: CogVideoX is faster on an H100 but cheaper per clip on an A100, because the A100’s hourly price falls further than its speed does. Second, for Wan’s big model the two 80GB cards tie at 480p ($0.24 either way) and nearly tie at 720p — so the H100’s real product is not cheaper clips, it is time: about 3.5 clips an hour against the A100’s 1.3. Third, the distilled small models are in a different economy altogether — at 75 seconds a clip, a 4090 turns out roughly 48 HunyuanVideo clips an hour for $0.34. And one honest multiplier applies to every row: these are costs per generation. Nobody ships the first roll of a video model, so budget two to three generations per clip you keep — the ranking doesn’t change, the budget does. The models, one by one: Wan 2.2 TI2V-5B is the default starting point — Alibaba’s 5B does text- and image-to-video at 720p and 24fps under Apache 2.0, and Wan calls its 4090 result “under 9 minutes” unoptimized on a consumer card. HunyuanVideo-1.5 generates at 480p and 720p with super-resolution to 1080p, and its 75-second figure belongs to the step-distilled build, not the full 50-step base model. Wan 2.2 T2V-A14B is the quality pick — and its measured 59.8GB peak is why claims that it needs a 96GB card are wrong on the maker’s own numbers; a matched image-to-video twin (I2V-A14B) sits under the same licence. CogVideoX-1.5-5B, out since November 2024, renders 1360×768 at 5 or 10 seconds and publishes the most complete speed table of any model here. Mochi 1 is a 480p preview (848×480) whose maker recommends an H100 and publishes no timing at all. LTX-2 is the only model here that makes sound with its picture — a 19B two-stage pipeline with bf16, FP8 and NVFP4 checkpoints plus a distilled build that runs in 8 steps — and the least documented: treat your first rented hour as its benchmark.
The counterweight: when renting is the wrong answer
For a handful of clips, a hosted video API beats every row in this guide: no download, no stack, a finished clip in minutes, and somebody else’s queue to wait in. Renting wins when the volume is steady, when the footage can’t leave your own machine, or when you need the model itself — fine-tunes, LoRAs, a fixed-seed pipeline an API won’t expose. And a note on the other big guide to this question: Spheron’s “Best GPU for Video Generation” coverage, read Nov 4, 2026, puts Wan 2.2 at 720p in a 65–80GB bracket with an H100 as the floor, and builds its clip costs on Spheron’s own stated rates. Wan’s published table measures 59.8GB peak on one GPU, and at a verified $1.09/hr the A100 run in the table above costs $0.83 a clip. Dated, checkable prices change the answer — which is rather the point of this page.
Steps
1. Pick the model by the clip, not the headlineNeed 720p cheaply and in volume: Wan’s 5B or distilled HunyuanVideo-1.5 on a 4090. Need the best open quality at 720p: Wan’s A14B on an 80GB card. Need sound with the picture: LTX-2, priced by the hour. Need it under Apache 2.0 for a commercial pipeline: Wan or Mochi — and read the licence box below before choosing the others.
2. Rent the smallest card that runs it well — for one hour firstThe makers’ timings are their settings: their prompts, their step counts, their reference clips. Your first rented hour is for timing your prompts at your settings. Your seconds-per-clip × the card’s $/hr ÷ 3,600 is the only cost figure worth budgeting from after that.
3. Scale by the arithmetic, not by feelClips per hour is 3,600 ÷ your seconds per clip. If the queue outgrows one A100, check the makers’ tables before adding a second card: on Wan’s own numbers the H100 is roughly 2.6× faster at 720p for roughly 2.5× the single-card price — a wash on cost per clip, a clear win on turnaround.
Alternatives
Watch out — the $2.25 H100 is a fleet priceTensorDock’s $2.25/GPU/hr for an H100 SXM — the cheapest H100 figure on our board — carries an 8-GPU minimum: $18.00/hr before you have made a single frame. Every single-card H100 figure in this guide uses RunPod Community’s $2.69/hr, the cheapest H100 you can actually rent one of.
Watch out — minimum-VRAM figures are offload figuresHunyuanVideo-1.5’s 14GB and CogVideoX’s “from 10GB” both need model offloading switched on, and the makers’ headline timings were not measured at those minimums. Renting the smallest card a model can technically start on, and expecting the headline speed, is how an hour’s budget becomes an afternoon.
Watch out — three of these licences are not ApacheHunyuanVideo-1.5 is under the Tencent Hunyuan Community licence (its repository is flagged EU-disallowed on Hugging Face), LTX-2 under the LTX-2 Community License Agreement, and CogVideoX-1.5-5B’s weights under the CogVideoX licence. Wan 2.2 and Mochi 1 are Apache 2.0. Read the licence that covers your model before commercial work.
Watch out — per-clip cost is per generation, not per keeperEvery figure in the cost table buys one roll of the model. At the two to three generations a usable clip usually takes, multiply accordingly — on the 4090 rows that is still cents; on the A14B at 720p it is the difference between $0.83 and about $2.50 per kept clip.
Good to know — you do not need a 96GB card for 720p WanWan’s own table peaks at 59.8GB for a 5-second 720p A14B clip on one GPU, so an 80GB card runs it with headroom to spare. Bigger cards buy speed and batch room on this model, not permission to run it.
The licences, in plain words
Wan 2.2 (both the A14B and the 5B) and Mochi 1 are Apache 2.0 — the permissive licence most commercial pipelines are built on. HunyuanVideo-1.5 is under the Tencent Hunyuan Community licence; Hugging Face’s metadata names that licence and flags the repository as disallowed in the EU. LTX-2 is under the LTX-2 Community License Agreement. CogVideoX-1.5-5B: its code is Apache 2.0, its weights are under the separate CogVideoX licence. This is a plain-words summary, not legal advice — each licence is linked from its model’s repository, and the three community licences deserve a full read before you sell anything made with them.
How we know
Verified
- Wan 2.2 T2V-A14B — 27B total / ~14B active MoE; Apache 2.0; 480p and 720p, 5-second videos; the single-GPU command “can run on a GPU with at least 80GB VRAM” with offloading; efficiency table (single GPU — A100: 480p 785.7s at 41.3GB peak, 720p 2,735.7s at 59.8GB peak; H100: 480p 326.9s, 720p 1,041.5s, same peaks) — from Wan-AI’s model card for Wan-AI/Wan2.2-T2V-A14B on Hugging Face, read Nov 4, 2026; the efficiency table is published on the card as an image and was read directly
- Wan 2.2 TI2V-5B — 5B dense; Apache 2.0; 720p at 24fps, text- and image-to-video; “at least 24GB VRAM (e.g, RTX 4090)”; RTX 4090 measured at 534.7s and 22.9GB peak for a 5-second 720p clip — from Wan-AI’s card for Wan-AI/Wan2.2-TI2V-5B, read Nov 4, 2026; the Wan 2.1 note (the 1.3B makes a 5-second 480p video on an RTX 4090 in about 4 minutes) is from Wan-AI/Wan2.1-T2V-14B’s card, read the same day
- Wan 2.2 is the newest downloadable Wan — the Wan-AI organisation’s model listing on Hugging Face, checked Nov 4, 2026, lists no Wan 2.5+ weights
- HunyuanVideo-1.5 — 8.3B; Tencent Hunyuan Community licence, repository flagged EU-disallowed in the Hugging Face metadata; minimum GPU memory 14GB with model offloading; 480p/720p with super-resolution to 1080p; default 121 frames at 24fps; the step-distilled 480p image-to-video build generating “within 75 seconds” on a single RTX 4090 (the card’s Dec 5, 2025 news entry) — from tencent’s card for tencent/HunyuanVideo-1.5, read Nov 4, 2026
- LTX-2 — 19B; LTX-2 Community License Agreement; bf16, FP8 and NVFP4 checkpoints plus a distilled checkpoint; two-stage pipeline; the card’s own example uses sequential CPU offload; no VRAM floor and no generation time published on the card — from Lightricks’ card for Lightricks/LTX-2, read Nov 4, 2026; checkpoint file sizes (bf16 43,285,058,242 bytes, about 43.3GB; FP8 27,078,716,018, about 27.1GB; NVFP4 19,988,416,674, about 20.0GB) are from that repository’s file listing via the Hugging Face API, read the same day
- CogVideoX-1.5-5B — 5B; released Nov 8, 2024; 1360×768; 5 or 10 seconds at 16fps; BF16 recommended; published inference speeds of ~1,000 seconds per 5-second video on a single A100 and ~550 seconds on a single H100 at 50 steps with the maker’s memory optimisations; memory from 76GB for the full BF16 route down to “from 10GB” for the optimised diffusers route; code Apache 2.0, weights under the CogVideoX licence — from the maker’s GitHub README (github.com/zai-org/CogVideo), read Nov 4, 2026; the model’s Hugging Face README endpoint returned an error, so the maker’s GitHub is the source of record here
- Mochi 1 — 10B; Apache 2.0; 480p preview at 848×480; the official route requires ~60GB VRAM on a single GPU and Genmo recommends at least one H100; a diffusers example at highest quality requires 42GB with CPU offload and VAE tiling; no per-clip generation time published — from genmo’s card for genmo/mochi-1-preview, read Nov 4, 2026
- GPU prices — our board check of Nov 4, 2026 (banked snapshot 2026-11-04): RunPod Community RTX 4090 $0.34/hr and H100 SXM $2.69/hr; Thunder Compute A100 80GB $1.09/hr; quoted in passing, RunPod Community L40S $0.79/hr, RTX 5090 $0.69/hr and RTX PRO 6000 $1.69/hr; TensorDock’s H100 SXM $2.25/hr is named only to flag its 8-GPU minimum and is never used as a single-card price. The arithmetic: clip cost = seconds ÷ 3,600 × $/hr (75s → $0.0071, printed $0.007; 534.7s → $0.0505, printed $0.05; 785.7s at $1.09 → $0.24; 326.9s at $2.69 → $0.24; ~1,000s at $1.09 → $0.30; ~550s at $2.69 → $0.41; 1,041.5s at $2.69 → $0.78; 2,735.7s at $1.09 → $0.83)
- Spheron’s competing video-generation guide was read via search on Nov 4, 2026; its 65–80GB bracket for Wan 2.2 at 720p and its rates are Spheron’s own stated figures, labelled as such where mentioned, and none of them is used as a verified price anywhere in this guide
Still unverified
- Any independent benchmark of these models’ generation times — every timing in this guide is the maker’s own, labelled maker-reported wherever it is used
- LTX-2’s real-world VRAM need and speed — Lightricks publishes neither; our 80GB pick is sizing from the published checkpoint files plus the maker’s own offload example, not a maker specification
- Mochi 1’s seconds-per-clip on any card — Genmo publishes none, so the model is priced per hour only
- HunyuanVideo-1.5’s base-model generation time — Tencent publishes the step-distilled figure only; the base model is costed by the hour on the same card
- The CogVideoX licence’s exact commercial thresholds — the licence file is linked from the maker’s repositories; read it before commercial use
Every timing on this page is maker-reported and labelled; every price is from the board check of Nov 4, 2026. Nothing else on this page is estimated.
Related guides
GPU Selection · 2026-10-30 · 14 min
Most “best GPU for inference” pages price their verdicts from a rate card the seller itself publishes — prices stated on the seller’s own pages, with no check date a reader can test. This page works the other way round. Every price below was read on the provider’s own published page and checked on Oct 30, 2026, and every cost per token is built from a throughput assumption printed next to it, so the maths can be rerun with your own numbers. The result is not the ranking the spec sheets suggest: on a mid-size model, the cheapest tokens in our checks come off a $0.34-an-hour gaming card, and the fastest card on the board finishes mid-table.
GPU Selection · 2026-10-30 · 8 min
On paper, AMD’s MI325X beats NVIDIA’s H200 where inference hurts most: 256GB of memory against 141GB, and 6 TB/s of bandwidth against 4.8. But there is a catch nobody puts in the headline — you can barely rent an MI325X at all: no verified price list we checked carries one. This page does the comparison honestly: the H200 prices you can actually pay today, the AMD prices that actually exist (MI300X and MI355X, verified on Oracle’s own list), and the one comparison the spec sheet can’t settle for you.
GPU Pricing · 2026-11-01 · 10 min
Search for an H100 rental price and you will meet several different numbers, every one of them presented as current. In late October 2026 we counted five different on-demand prices for the same H100 on a single provider’s own pages — on the same day. This guide puts the stated figures next to the verified ones, names the five mechanisms that pile prices up around one GPU, and ends with the three-step check that separates a price from a claim.
Go deeper