24 October 2026 · Site Methodology & Corrections · 4 min read
Our Replicate pricing page has said A100 is $0.00115/second and H100 is $0.0025/second since July -- H100 costing more than double. I checked it against Replicate's own pricing page today. Both numbers are gone. The A100 40GB tier doesn't exist anymore, and H100 now runs at $0.001525/second -- barely above A100, not 117% above it.
| Аппаратное обеспечение | We had (July) | Real now | Изменять |
|---|---|---|---|
| A100 | $0.00115/s (40GB) | $0.0014/s (80GB) | 40GB tier retired, 80GB-only is 22% pricier |
| H100 | $0.0025/s | $0.001525/s | 39% cheaper |
| L40S | not listed | $0.000975/s | new tier, 30% cheaper than A100 |
Two separate things happened and they point in opposite directions, which is exactly why nobody notices both at once. A100 got more expensive because the cheap 40GB variant was pulled -- everyone building on A100 now defaults to the 80GB card whether their model needs that much memory or not. H100 got much cheaper, enough that it closed almost the entire gap to A100.
Our old numbers implied H100 costs 117% more than A100 -- a real deterrent if you were choosing hardware on a budget. The actual gap today is 9%. For anyone picking between the two for a model that runs on either, that's the difference between "H100 is a premium upgrade" and "H100 is basically the same price, take it." I'd been telling readers for three months to default to A100 unless they specifically needed H100's speed. That advice was built on a price gap that stopped being true at some point between July and now and I never re-checked it.
Take a job that burns 10,000 seconds of GPU time -- a batch of video generations, say, at a few minutes each. At our old H100 rate that's $25. At the real rate it's $15.25. Anyone who sized a budget off our old table, or compared Replicate against a competing GPU host using our H100 number, was working from a figure 64% too high. In the other direction, anyone who assumed A100 was still $0.00115/second and budgeted off that was underestimating by 22%, because that card doesn't ship at that price anymore.
Replicate added Nvidia L40S at $0.000975/second -- about 30% cheaper than A100 and still well above T4's $0.000225/second. It sits in a gap our table never acknowledged: image and mid-size video models that don't need A100's full memory but outgrow T4. If you've been defaulting to A100 for SDXL-class workloads out of habit, L40S is worth testing first -- same class of card, lower rate, and it's been on the official pricing page the whole time we didn't have a row for it.
Replicate bills per second with no contract to review, no invoice that spells out "price changed" -- the rate just moves on the pricing page and your next bill reflects it. There's no renewal notice forcing a look, unlike a SaaS subscription. I'd verified this page once in July against the number that was live then and treated it as stable infrastructure pricing rather than something that needed the same periodic re-check as a volatile LLM token rate. GPU cloud pricing moves plenty -- it just doesn't announce itself.
Updated the Replicate guide's table: A100 is now listed as 80GB only at $0.0014/second, H100 at $0.001525/second, and added the L40S row that was missing entirely. The per-1,000-run estimate column was also quietly assuming a 5-second job in the caption while the actual math used ~120 seconds -- fixed that caption to match reality too, since a 5-second caption next to a 120-second-implied number was its own small lie.
New to usage-based API pricing in general? Start with the бесплатные руководства по стоимости API.
Разверните его самостоятельно: DigitalOcean — бесплатный кредит в размере 200 долларов США ↗ · Хостингер VPS ↗
Checked directly against Replicate's official pricing page, 2026-10-24. Reference figures, not live quotes -- confirm current rates on replicate.com/pricing before budgeting.