A budget for running video generation on your own GPUs
Video Model Cost Calculator
Enter your workload to estimate the total cost of ownership and break-even point of running open-weight video generation models on your own GPUs, compared against official and hosted API per-second pricing.
Electricity for hardware you already own
This total includes only the electricity for the whole machine during the hours entered. Equipment purchase, depreciation, labor and cloud rental are excluded. It replaces the ownership estimate; do not add the two electricity totals.
Choose the currency on your electricity bill. Changing currency does not convert the rate; enter the new rate yourself.
Enter measured average power for the entire system, including CPU, RAM and PSU losses. GPU TDP is not a whole-machine measurement. No extra PUE multiplier is applied.
Enter elapsed running hours. Divide measured seconds by 3,600. Monthly time is limited to 744 hours; include idle hours only if the same average power covers them.
Use the variable energy rate on your bill. Fixed connection fees are separate.
Enter finite, non-negative power, hours and price. Maximum: 10,000,000 W, 1,000,000 task hours or 744 monthly hours, and 10,000 per kWh. Zero is allowed.
Time is entered directly here. Other workload calculators derive it only where matching performance data exists. Output duration, model memory fit and a GPU name do not establish generation time. For your own setup, measure the complete task, including offload and retries, before estimating its electricity.
API workload budget
Estimate metered usage in USD from the stored tariff. Include every attempt and retry. Taxes, credits, free allowances and fixed fees are separate. This is independent of the self-hosting capacity estimate.
Count every attempt, including retries. Output and reference-material charges apply to each attempt; the first 5 reference images are free per attempt. For reference video, enter 0 for none or at least 2 seconds.
Estimated usage cost: $400.00
5,000 output seconds × $0.08 · base · 768p
Accepted clips: 1000 · Attempts per accepted clip: 1 · Total attempts: 1000
Retry attempts: 0 · Of which, retry cost (included in total): $0.00
Reference materials: $0.00
Cost per accepted clip: $0.40
Tariff source and verification date ·
Reference fee source and verification date ·
Provider notes and billing conditions
Billed per second of output. Input materials: audio free; first 5 images free, then $0.04/image; reference video billed by input duration at the output rate.
Video models
Browse Video modelsGet the cost of one second of video, then decide how to run it
Turn the cost of self-hosted video generation into a number your team can check, compare, and use to discuss building versus calling an API.
Open weights are free to download, but running them is not. GPUs, power, and day-to-day operations all land on the bill, while API paths grow linearly with every output second.
Use this estimate to judge whether self-hosting belongs in your video pipeline plan — then take the budget to your team.
The total is only the start. Results list the upfront investment, monthly cost, per-output-second cost, and break-even months, so you can explain where the budget goes.
What determines the cost of one second of video?
This estimate looks at four things: resolution, workload, hardware profile, and the operational work needed to keep a deployment running.
Resolution tier
The open checkpoints self-host only at native 768p; 480p and 2K/4K outputs are API-only, and per-second pricing differs by tier.
Expected workload
Clip length × clips per month sets the GPU-hours required. The higher the volume, the thinner the fixed hardware investment spreads.
Hardware profile
Throughput samples come from 4× H100 80GB (BF16, SGLang). A different tier changes generation speed, GPU-hours, and utilization.
Operations and power
Deploying, monitoring, upgrading, and handling failures all take people; electricity and PUE add to the monthly bill based on your region.
You can adjust the exchange rate, electricity price, salary range, FTE ratio, PUE, system cost multiplier, and depreciation period.
Self-host estimates use one 4× H100 80GB deployment at 768p. Use the API path to compare other resolution tiers.
Method & assumptions
How the estimate works — the assumptions behind every number on this page.
- Workload is measured in seconds of output video: clip duration × clips per month. API prices are published per output second by resolution tier.
- Self-host throughput is an estimate: public samples report roughly 13–74 seconds to generate a 5-second 768p clip on 4× H100 80GB (BF16, SGLang). We use the conservative end-to-end figure and assume generation time scales approximately linearly with output duration at fixed resolution and step count. Treat all throughput-derived numbers as directional estimates.
- Local purchase cost = one-time hardware (× system multiplier) depreciated over the chosen period + electricity (GPU-hours × power × PUE × rate) + operations labor. Cloud rental bills only the GPU-hours actually spent generating.
- Hosted API cost = monthly output seconds × the per-second price of the selected resolution tier. Reference input materials (images, reference video) may be billed separately by the provider.
- Break-even months ≈ one-time local investment ÷ (cheapest API monthly cost − local monthly operating cost), shown only when the difference is positive. Directional only.
- Check the model license for permitted territories, revenue thresholds and the scope of open checkpoints before choosing a deployment path.
What this estimate does not include
Allow for the production costs and quality factors below when comparing video generation budgets.
- This is a planning estimate, not a vendor quote.
- Generation quality, prompt adherence, and aesthetics are outside the price estimate — the post-training gap of fal H3 Max cannot be measured in price alone.
- Content moderation, safety, and compliance costs are not included.
- Video storage, delivery bandwidth, and asset management may add to the monthly bill.
- The self-hosting estimate does not include compute burned on failed generations and retries. The API workload budget counts the attempts you enter.
- Throughput is based on sparse public samples; real speed varies with steps, framework, and load.
How to estimate the cost of self-hosted video generation
Set your workload, pick a hardware profile and assumptions, and get a shareable number you can take to a budget discussion.
- 01
Set the workload
Enter clip length, clips per month, and target resolution to fix your monthly output seconds.
- 02
Pick hardware and assumptions
Choose owned or rented GPU profiles, then adjust electricity, labor, and other cost assumptions for your region.
- 03
Compare the four paths
See monthly and per-output-second costs for local purchase, cloud rental, the official API, and fal H3 Max — then share the URL with your team.
Should you self-host video models?
The calculator shows you the self-hosting side of the bill first; the final call also depends on your volume, budget, region, and whether the team can run the infrastructure.
Self-hosting may fit when:
- Your monthly volume is stable and large enough to amortize the hardware.
- You need full control over data and the generation pipeline.
- Native 768p output meets your quality requirements.
- The team can deploy and maintain GPU inference services.
- Your region is within the MiniMax H3 license territory (which excludes the US, EU, UK, and South Korea).
A hosted API may fit when:
- You want to ship fast without operating GPU infrastructure.
- Volume is low, uncertain, or highly variable.
- You need 2K/4K output or API-only components like Context-IR.
- You need fal H3 Max's post-trained quality, and its weights are not released.
- Your region falls outside the H3 license's applicable territory.
Both approaches can be right. The calculator places the self-hosting estimate next to the published prices of the official MiniMax API and fal H3 Max, and shows break-even months against the cheapest API path.
Supported video models
Open-weight video generation models you can self-host, with their license terms and open checkpoint scope.
Data sources
Every price and throughput figure on this page traces back to a public source.
Hosted API list prices, manually verified: Last verified 2026-09-03.
- https://platform.minimax.io/docs/guides/pricing-paygo
- https://fal.ai/models/minimax/h3/text-to-video
- https://fal.ai/models/minimax/h3-max/text-to-video
The result’s "Hardware & cloud rental price sources" card lists the provider, region, update date and source link where available for each GPU purchase or rental price.
Prices are list prices for budget estimation only; promotions, discounts, and regional terms may differ. Confirm with the provider.
FAQ
Common questions about self-hosting video models and how this estimate works.