- Cost per generated second is the only unit that compares renting, buying, reserving, bursting and APIs. A $6 card doing 60 seconds an hour beats a $2 card doing 5, by four times.
- Pipeline choice moves cost more than card choice. Distilled runs 8 sigmas, DFR is the production path, and gradient estimation cuts guided runs from 40 steps to 20 or 30.
- Reserved capacity is the option most comparisons skip: roughly 32% off at one year, 50% at three, with no purchase order. Spot runs about half of on-demand if your work is restartable.
GPU cost is the line item most video teams discover after the pilot, not during it.
Rent an H100 today and you will see anything from $1.49 to $6.98 per GPU-hour, per CloudZero's H100 analysis, rates verified against provider pages on 20 August 2026.
A 4.7x spread on one card tells you the hourly rate decides nothing. Cost per generated second of video does, and it is the only unit that compares renting, buying, reserving and per-second APIs on one axis.
This guide converts all of them into that unit and names the costs that never reach the spreadsheet.
The unit that matters
Cost per generated second is the unit. Every other figure is an input to it.
A $2 per hour GPU rendering 5 seconds of 1080p video per hour costs $0.40 per generated second. A $6 per hour GPU rendering 60 seconds per hour costs $0.10.
The expensive card is four times cheaper, and no hourly comparison shows you that.
Why dollars per GPU-hour misleads
Hourly pricing hides two multipliers: throughput, which depends on your pipeline, and utilization, which depends on how your team works.
A rented GPU charges the same whether it renders or idles, so a card busy 20% of the day costs five times more per output second than one busy all day.
What drives GPU hours
GPU hours per finished clip scale with resolution, frame count, sampling steps, and how many takes you throw away.
Step count is the one teams underweight, and on LTX-2.5 it is the largest single lever.
The open-source documentation states that DistilledPipeline is the fastest inference path, running 8 predefined sigmas. The guided two-stage pipelines run the full model instead, and the gradient-estimation denoising loop allows 20 to 30 steps instead of 40 while maintaining quality.
Choosing distilled over a 40-step guided run is not a 10% saving. It is a different order of magnitude.
There is a third option worth pricing separately. DFRPipeline is the documented production-quality path, running the same distilled transformer plus a detailing pass with optional temporal upscaling rounds. It costs more than distilled and less than a full guided run, and it is the one most teams should be budgeting for on final renders. The capabilities overview covers where each fits.
The frame-count tax
Frame count adds a wrinkle that quietly inflates budgets.
The LTX-2.5 Video VAE compresses frames temporally, and frame count must satisfy (F-1) % 8 == 0. Width and height must also be divisible by 32.
You cannot render 100 frames. At 25fps the valid counts are 9, 17, 25, 33, 41, 49, 57, 65, 73, 81, 89, 97, 105, 113, 121 and 129.
A 5-second brief is 125 frames, which is not valid. The counts either side are 121, at 4.84 seconds, and 129, at 5.16 seconds. If the brief is a floor rather than a target you round up to 129, and every clip carries about 3% more compute than a round-seconds budget predicted. Across a campaign that compounds.
You can also omit the frame count entirely and let the duration head pick a length from the prompt, which removes the arithmetic but makes per-clip cost less predictable.
Rent, buy or burst at a glance
Rent, buy, reserve or burst
Four options, and the right one is decided by utilization rather than volume.
Renting is priced per GPU-hour with no commitment. It suits real volume on a workload that is still moving, and it converts a six-figure guess into an hourly experiment you can end on Friday.
Buying removes the per-hour charge and adds power, cooling, ops and a purchase order. It wins only when the card stays busy most of the day for long enough to amortize.
Reserving is the option most comparisons skip. Committing to a term cuts the rate substantially without a capital purchase: the GetDeploying index puts a one-year reservation at about 32% below on-demand and a three-year at about 50% below, measured against each provider's own on-demand price for the same card.
Bursting on spot capacity trades reliability for price. Spot lists at roughly half of on-demand across the market, with H100 spot at about 53% of the on-demand rate.
What does renting cost?
H100 on-demand rates run from $1.49 on marketplace capacity such as Vast.ai, through $1.99 on RunPod and $3.99 on Lambda, to $4.25 on CoreWeave and as high as $6.98 on hyperscalers, per CloudZero's 20 August 2026 verification.
That hyperscaler gap is structural, not a negotiating failure. The GetDeploying index, last updated 14 September 2026, puts the median H100 at $5.00 per GPU-hour on dedicated GPU clouds against $7.89 on hyperscalers, a 58% premium for identical silicon.
Blackwell sits higher. The same index puts the B200 median at $7.88 per GPU-hour on dedicated clouds and $15.26 on hyperscalers.
Pick the card by VRAM before price. The LTX documentation targets NVIDIA GPUs with 80GB or more of VRAM for the standard configuration, on Linux with Python 3.12 or newer, CUDA 12.7 or newer, and PyTorch 2.7, with distilled variants supported on 32GB using FP8 quantization.
That sorts the market for you. An 80GB H100 or H200 runs the standard configuration, a 32GB card runs distilled with quantization. Renting a 24GB card to save money buys an out-of-memory error.
When is renting right?
Renting is right when your volume is real but your workload is still moving.
You keep the full local pipeline surface, every CLI flag and all twelve documented pipelines, without a hardware purchase order, plus the option to change your mind next quarter.
Should you buy the hardware?
Buying makes sense at one point, and the break-even is a utilization threshold rather than a volume one.
Measure your own throughput first. The LTX documentation publishes hardware requirements rather than throughput benchmarks, so any borrowed number will be wrong for your pipeline.
For a sense of the spread, LTX's own published figure is a 10-second 720p clip in 6.8 seconds, self-hosted on two NVIDIA GB200s at steady state. That is a best case at a specific resolution on specific hardware, which is exactly why it is an illustration rather than an input to your model. Run a representative clip at your target resolution, frame count and pipeline, and record seconds of output per GPU-hour.
Then run four numbers.
Rented equals hourly rate divided by output seconds per hour.
Owned equals purchase price divided by expected hours of life, plus power and ops per hour, divided by output seconds per hour.
Reserved equals the committed rate divided by output seconds per hour, which on a one-year term is roughly two thirds of the on-demand figure.
API equals the published per-second price at your resolution.
Multiply each by monthly output seconds. The winner is usually obvious, and usually not the one the team assumed.
The pattern is consistent. Owning wins only when the card stays busy most of the day long enough to amortize it. Otherwise renting, reserving and bursting win, because they convert a fixed cost into a variable one.
Where a consumer card fits
A 32GB consumer card such as the RTX 5090 matches the documented tier for quantized distilled configurations.
It will not run the standard 80GB config and it is wrong for a high-throughput render farm. It is right for a small team that generates continuously, keeps everything on premises, and wants no per-generation fee.
LTX Desktop is built around this case: local on consumer hardware, no cloud dependency, no per-generation fees.
When does bursting pay off?
Bursting pays off when demand is spiky and work is restartable.
Spot capacity sells at a discount because the provider can reclaim it, and across the market spot lists at roughly half of on-demand. Campaign renders, overnight batch passes and dataset preprocessing tolerate interruption. A client review session does not.
The hidden cost of bursting is the restart. A self-hosted deployment loads the full split checkpoint pack before it generates anything, and a preempted instance pays that load again. Batch enough work into each window that the load amortizes.
The costs the sticker hides
Three costs sit outside every hourly rate, and together they move the answer.
Failed generations. On a rented or owned GPU, an out-of-memory crash costs a full billing increment and produces nothing. The LTX API works differently: prepaid billing reserves the estimated cost before generation, commits on success and releases on failure, so failed jobs do not charge the account. If your pipeline has a real failure rate, that asymmetry is a line item.
Iteration, not output. Teams budget for the clip they ship and forget the nine they discarded. At LTX's published $0.13 per second for ltx-2-5-fast at 1080p, a 10-second clip is about $1.30 and ten takes are $13. Forecast takes, not deliverables.
Storage. The LTX-2.5 split checkpoint pack is tens of gigabytes before you store a single frame, and EXR output for colour pipelines is larger again than an MP4 master.
How to choose your model
Answer three questions in order, and stop at the first clear answer.
Monthly output under a few thousand seconds? Use a per-second API. ltx-2-5-fast runs $0.09 per second at 720p and $0.13 at 1080p, so 3,000 seconds of 1080p costs around $390 a month. No rented GPU idling two thirds of the day beats that. ltx-2-5-pro runs $0.12 and $0.17 at the same tiers.
Utilization steady but workload still changing? Rent, or reserve if you can commit to a year. Match VRAM to configuration first: 80GB or more for the standard config, 32GB for distilled with quantization.
Card busy for 18 months or more, footage that must stay on your hardware? Buy, and model power, cooling and ops into cost per output second before you sign.
Summary
Cost per generated second, not cost per GPU-hour, is the unit that compares renting, buying, reserving, bursting and per-second APIs.
H100 rental spans $1.49 to $6.98 per GPU-hour as verified in August 2026, with a median around $5.00 on dedicated GPU clouds against $7.89 on hyperscalers, and the B200 median at $7.88 against $15.26. Reservations cut roughly 32% at one year and 50% at three, and spot lists at about half of on-demand.
LTX-2.5 targets 80GB or more of VRAM with CUDA 12.7 or newer, and distilled variants run on 32GB using quantization, so VRAM sorts the shortlist before price does.
DistilledPipeline runs 8 predefined sigmas, DFRPipeline is the production path, and gradient estimation cuts guided runs from 40 steps to 20 or 30. Pipeline choice is the largest lever on compute cost.
Frame counts must satisfy (F-1) % 8 == 0, so a 5-second brief rounds up to 129 frames and round-second budgets understate real compute.
Want to measure it yourself? Pull the weights from HuggingFace, run the pipelines from the repository, or start in the LTX Playground. If a per-second rate is the right shape for now, current LTX API pricing is published by tier and resolution.
