Local vs Cloud for H3: The Real Cost Math (September 2026 prices)

Updated 2026-09-20 · Every price below is sourced and dated; promotional rates change — check the platform's current page.

Global site · runninghub.ai Promoted Sign up via our link — 1,000 RH Coins free Sign up · 1,000 RH Coins → Campaign terms on the platform page; we may earn a commission at no extra cost

China site · runninghub.cn Promoted Sign up via our link — 1,000 RH Coins free Sign up · 1,000 RH Coins → Campaign terms on the platform page; we may earn a commission at no extra cost

The formulas first

Before arguing about "cheaper", write both billing models as formulas. Every number below is just these formulas with assumptions plugged in, so you can swap in your own electricity price, card price, and clip duration and check the math yourself.

Local cost per clip (card already owned):

cost per clip = system power draw (kW) × electricity price (per kWh) × time per clip (hours)

Local cost per clip (buying a card):

cost per clip = electricity + card price ÷ amortization years ÷ clips per year

A card you already own is a sunk cost — count it as 0 and only pay for electricity. Only a newly purchased card carries depreciation, typically amortized over 2-3 years. This is the single most important distinction in the article: "already own" and "no card yet" are two completely different questions.

Cloud cost per clip, hourly billing:

cost per clip = (time per clip + share of queueing/upload/download) × hourly rate

Cloud cost per clip, per-second billing:

cost per clip = price per second × seconds of video × number of takes

Two structural differences matter. First, cloud time is billed continuously — booting, loading models, downloading weights, and queueing can all burn money, while locally those steps cost a little electricity at most. Second, on per-second billing a failed take costs the same as a good one: the more you roll the dice, the more each keepable clip really costs. Local retries are almost free.

Four routes at a glance (September 2026)

Shared assumptions for the estimates: local system draw 350W (5090-class: 450-500W), electricity ¥0.6/kWh, per-clip times from this site's measured anchors; hourly cloud cost = time per clip × rate, excluding upload/download and queueing; RunningHub figures are member promo rates that can end at any time. These are worked examples and assumptions, not quotes — the platform's current page and your own electricity bill are the source of truth. Site anchors: RTX 5090 at 1344×768, 73 frames, 8-step acceleration: ~80.3 s end to end per clip (source: /articles/5090-speedup logs, 2026-09); a 12GB-class card at 1280×736 / 5 s / 4 steps: ~7 min per clip (source: /articles/vram-8gb, 2026-09). To estimate times on your own card, use the calculator or the per-GPU pages in the GPU encyclopedia.

Worked example 1: an existing RTX 4070 12GB, 5 short clips a day

Assumptions (spelled out so you can re-run them):

• The GPU is already owned; depreciation counts as 0 (sunk cost — electricity only);
• Each clip is 1280×736 / 5 seconds / 4-step acceleration, measured at ~7 minutes per clip on a 12GB-class card (source: this site's /articles/vram-8gb table, 2026-09);
• System draw under load 350W, electricity ¥0.6/kWh (illustrative assumptions — substitute your own rate);
• 30 days × 5 clips = 150 clips/month.

Local electricity: 150 clips × 7 min = 1,050 min = 17.5 hours; 17.5 × 0.35kW × ¥0.6 ≈ ¥3.7/month, or about ¥0.025 per clip (~2.5 fen). (At the lighter 576×832 draft setting — ~80 seconds per clip — the monthly bill drops to about ¥0.7, roughly half a fen per clip.)

Renting a 4090 for the same output: on the same basis, 720p / 5 s / 4 steps takes roughly 2-4 minutes; call it 3, so 150 clips = 7.5 hours —

• AutoDL at ¥1.88/h: 7.5 h = ¥14.1/month;
• Compshare at ¥2.15/h: 7.5 h = ¥16.1/month;
• Add the time spent uploading/downloading weights and queueing (roughly 20% here) → about ¥17-19/month in practice.

Verdict: same output, ~¥3.7/month local versus ~¥17-19/month cloud — 4-5× cheaper if you already own the card. The trade-off: ~7 minutes per clip locally versus ~3 in the cloud. If you'll pay 5× to save 4 minutes, that's who cloud pricing is for. Note the premise: the card is already paid for. Amortize the original purchase over "5 clips a day" and the picture changes — which is exactly what example 2 answers.

Worked example 2: a student with no GPU — rent first or buy?

The logic is simple: the extra money you spend on a card has to be earned back through the hours it saves you versus renting. How much cheaper local is per hour sets the payback period.

break-even hours = card price ÷ (cloud hourly rate − local marginal hourly cost)

Assumptions:

• Local marginal hourly cost = 0.35kW × ¥0.6/kWh = ¥0.21/hour (same illustrative assumption as above);
• Cloud = the cheapest 4090 rate, AutoDL at ¥1.88/hour (homepage price list, 2026-09);
• Card price: an illustrative ¥10,000 (hardware/system spend — a formula demo, not a quote from anyone).

Plug in: 10,000 ÷ (1.88 − 0.21) ≈ 5,988 hours ≈ 6,000 hours. You'd have to run ~6,000 cumulative hours before buying matches the cash cost of renting 4090s forever. Spread that over months —

Buy column = ¥10,000 ÷ 36 months + electricity (hours × ¥0.21). Rent column = hours × ¥1.88. Estimates under the stated assumptions, not quotes.

So "how many hours a month before buying pays off" has no fixed answer — it's a threshold set jointly by card price and cloud rate. Under the assumptions above it lands around 165 hours a month (~5.5 h/day) sustained for three years. For a student, concretely:

• Rent for 2-3 months first, log your real hours, then re-run this formula with your own card price. Most people find they use 20-60 hours a month — far below the threshold.
• Selling a used card recovers part of the price, which lowers the break-even point — but hardware depreciates fast, so don't count on a generous resale value.
• Conversely, at a dozen hours a month, rent forever: a cloud bill under ¥40/month beats any hardware that sits idle and takes up desk space.

Hidden cloud costs (put these in the comparison too)

• Upload/download time is billable time: the full H3 stack is 40GB+ (pruned INT8 main model 19.5GB + text encoder 14.6GB + VAEs; source: this site's /articles/vram-8gb). At 100Mbps that's roughly 53 minutes — an hour of billed time if the meter is running. Public model libraries save the download, but your own LoRAs and workflows still travel.
• Promo rates move: RunningHub's ¥0.05/s and ¥0.21/s are member promos (2026-08 and 2026-09 windows). They end. Budget on promo pricing and your numbers quietly stop working — check the platform's current page.
• Failed takes are billed like good ones: on per-second billing a wasted generation costs exactly as much as a keeper. At a 1-in-5 keeper rate, multiply the cost of a usable clip by five — locally, retries cost only electricity. This is the most underrated local advantage.
• Data privacy: your footage lands on someone else's machine. Client contracts, unreleased products, and people's faces deserve a hard think before uploading.
• Long jobs get interrupted: preemptible instances, session timeouts, and maintenance windows can take a half-finished run. Split long clips into segments, checkpoint, and budget for re-runs.

Decision checklist

Next steps

• Estimate time before money: check the calculator for your card and settings, or the per-GPU pages in the GPU encyclopedia;
• Downloads and quantized files live on the model files page (the 40GB+ of weights and disk space are part of the cost too);
• Acceleration choice (4/6/8 steps) drives time, and time drives cost — compare options in the acceleration leaderboard write-up before optimizing pennies;
• A cloud deployment guide (which instance, how to use public model libraries, when to stop the meter) is in the works and will appear in the articles index;
• Disclosure: some outbound links on this site may carry referral parameters; they don't change the formulas or conclusions here. Prices are as sourced below — see the disclosure page.

Sources

• Compshare (by UCloud): RTX 4090 ¥2.15/hour, RTX 5090 ¥3.32/hour — source: compshare.cn homepage price list, checked 2026-09
• AutoDL: RTX 4090 ¥1.88/hour, RTX 5090 ¥2.78/hour, RTX 3090 ¥1.32/hour — source: autodl.com homepage price list, checked 2026-09
• RunningHub (cloud ComfyUI) member promo rates: Wan3.0 480P from ¥0.05/s (2026-08-25 launch promo); Seedance 2.0 720P ¥0.3/s (2026-06 sale); Seedance 2.5 480P from ¥0.21/s (2026-09 promo) — source: RunningHub promo pages, checked 2026-09; promo rates change — the platform's current page is authoritative
• Comfy Cloud (official ComfyUI cloud): Standard $20/mo (4,200 credits ≈ 380 five-second videos), Creator $35/mo, Pro $100/mo — source: comfy.org/cloud/pricing, checked 2026-09
• Site anchors: RTX 5090 @1344×768 / 73 frames / 8 steps ≈ 80.3 s/clip (/articles/5090-speedup, 2026-09); 12GB-class @1280×736 / 5 s / 4 steps ≈ 7 min/clip (/articles/vram-8gb, 2026-09); RTX 4070 12GB @576×832 / 5 s / 4 steps ≈ 80 s/clip (/articles/hardware-guide, 2026-09)
• Local electricity and depreciation are formulas + illustrative assumptions (350W system, 450-500W for 5090-class, ¥0.6/kWh, ¥10,000 example card price) — not quotes from any platform or region; substitute your own rates