Direct answer: public serverless/per-second GPU prices in 2026 are roughly: RunPod on-demand A100 80GB 1.79/hour, H100 SXM 3.99/hour; RunPod Serverless A100 2.72, H100 5.39; Modal A100 80GB 0.000694/second (2.50/hour) plus CPU/memory, H100 0.001097/second (3.95/hour) plus CPU/memory; Replicate A100 80GB 0.0014/second (5.04/hour, including CPU/RAM), H100 0.001525/second (5.49/hour). At 100/300/730 GPU-hours/month on A100: RunPod on-demand 179/537/1,306.70; RunPod Serverless 272/816/1,985.60; Modal (GPU+4vCPU+16GiB) 272.06/816.18/1,986.04; Replicate 504/1,512/3,679.20. Bottom line: sustained/predictable workloads are cheapest on RunPod on-demand; among serverless, Modal matches RunPod Serverless on A100 and beats it on H100; Replicate is the most expensive per GPU-hour but sells the model marketplace/API experience. Calculate on actual active GPU-hours, not a default 730-hour month.
This guide is for AI engineering, platform and procurement teams running model inference, fine-tuning, batch jobs, image/video generation and agent backends. Data cut-off: 7 November 2026; prices come from RunPod Pricing, Modal Pricing and Replicate Pricing, in USD, excluding tax; all TCOs are editorial calculations, not vendor quotes. “On-demand/serverless, GPU model, CPU/memory, cold start/idle” differ, so official facts, editorial assumptions and non-comparable items are kept separate.
RunPod on-demand: A100 1.79, H100 SXM 3.99; Serverless: A100 2.72, H100 5.39. Modal: A100 0.000694/s, H100 0.001097/s, plus CPU/memory (4vCPU+16GiB about 0.22/hour); Starter 0/month with $30 credit, Team 250/month with $100 credit. Replicate: A100 0.0014/s (5.04/hour, including 10x CPU/144GB RAM), H100 0.001525/s (5.49/hour), T4 0.000225/s (0.81/hour). Sustained workloads favor on-demand; serverless bills actual active seconds.
Three meters: on-demand hours, per-second serverless, per-second with CPU/RAM
RunPod: Pods on-demand by the hour (A100 PCIe/SXM 80GB 1.79, H100 PCIe 2.89, H100 SXM 3.99, A40 0.59, L40S 1.09, H200 5.29, B200 7.99); Serverless per-second/per-hour (A100 2.72, H100 5.39, L40S 1.75, A6000/A40 1.22, 16GB class 0.58, H200 6.99, B200 10.79). Modal: per-second, GPU A100 80GB 0.000694/s (2.50/hour), H100 0.001097/s (3.95/hour), T4 0.000164/s (0.59/hour), L40S 0.000542/s (1.95/hour); CPU 0.0000131/core/s (about 0.047/core-hour), memory 0.00000222/GiB/s (about 0.008/GiB-hour); Starter 0/month + $30/month credit, Team 250/month + $100 credit. Replicate: per-second hardware, A100 80GB 0.0014/s (5.04/hour, including 10x CPU/144GB RAM), H100 0.001525/s (5.49/hour), T4 0.000225/s (0.81/hour), L40S 0.000975/s (3.51/hour).
| Platform/model | Billing | A100 80GB | H100 80GB | 16GB/T4 class | Note |
|---|---|---|---|---|---|
| RunPod on-demand (Pods) | Per hour | 1.79 | 3.99 (SXM) | — | Cheaper for sustained/always-on |
| RunPod Serverless | Per second/hour | 2.72 | 5.39 | 0.58 (16GB class) | Bills active seconds |
| Modal | Per second | 2.50 + CPU/memory | 3.95 + CPU/memory | 0.59 (T4) + CPU/memory | Starter 0 + $30 credit |
| Replicate | Per second | 5.04 (incl. CPU/RAM) | 5.49 (incl. CPU/RAM) | 0.81 (T4) | Model marketplace/API |
Sources: RunPod Pricing, Modal Pricing and Replicate Pricing (US/USD, accessed 7 November 2026). Modal’s GPU price excludes CPU/memory (“+CPU/memory” in the table); Replicate’s hardware price includes the listed CPU/RAM; RunPod Pods/Serverless follow the platform’s billing. Excludes tax, storage, egress, model/API fees, free credits and commitment discounts.
A100 80GB: 100/300/730 GPU-hours
A100 80GB at 100/300/730 active GPU-hours/month: RunPod on-demand 179/537/1,306.70; RunPod Serverless 272/816/1,985.60; Modal (GPU+4vCPU+16GiB, about 2.7206/hour) 272.06/816.18/1,986.04; Replicate 504/1,512/3,679.20. RunPod on-demand is cheapest; Modal and RunPod Serverless are almost the same (under $0.5 apart); Replicate is about 2.8× on-demand. If you run a few hours/day or spiky traffic, calculate on active GPU-hours — per-second serverless and scale-to-zero fit; at 730 hours always-on, on-demand/reserved is cheaper.
| A100 80GB (hours/month) | RunPod on-demand | RunPod Serverless | Modal (incl. CPU/memory) | Replicate |
|---|---|---|---|---|
| 100 | 179.00 | 272.00 | 272.06 | 504.00 |
| 300 | 537.00 | 816.00 | 816.18 | 1,512.00 |
| 730 | 1,306.70 | 1,985.60 | 1,986.04 | 3,679.20 |
Editorial calculation, not a quote. RunPod on-demand 1.79/hour, Serverless 2.72/hour; Modal 0.000694/s GPU + 4vCPU/16GiB (2 physical cores × 0.0000131 + 16 × 0.00000222/s ≈ 0.00006172/s, total 0.00075572/s ≈ 2.7206/hour); Replicate 0.0014/s (5.04/hour). Excludes tax, storage, egress, model/API fees, free credits and commitment discounts.
H100 80GB: 100/300/730 GPU-hours
H100 80GB: RunPod on-demand (SXM) 399/1,197/2,912.70; RunPod Serverless 539/1,617/3,934.70; Modal (GPU+4vCPU+16GiB, about 4.1714/hour) 417.14/1,251.42/3,045.12; Replicate 549/1,647/4,007.70. RunPod on-demand remains cheapest; Modal is about 22% below RunPod Serverless at 100 hours (417 vs 539) and about 24% below Replicate; Replicate is close to RunPod Serverless but includes CPU/RAM and the model marketplace. The absolute H100 gap is larger than A100 — the wrong platform costs thousands per year.
| H100 80GB (hours/month) | RunPod on-demand | RunPod Serverless | Modal (incl. CPU/memory) | Replicate |
|---|---|---|---|---|
| 100 | 399.00 | 539.00 | 417.14 | 549.00 |
| 300 | 1,197.00 | 1,617.00 | 1,251.42 | 1,647.00 |
| 730 | 2,912.70 | 3,934.70 | 3,045.12 | 4,007.70 |
Editorial calculation, not a quote. RunPod on-demand H100 SXM 3.99/hour, Serverless 5.39/hour; Modal 0.001097/s GPU + 4vCPU/16GiB ≈ 0.00006172/s, total 0.00115872/s ≈ 4.1714/hour; Replicate 0.001525/s (5.49/hour). Excludes tax, storage, egress, model/API fees, free credits and commitment discounts.
16GB/T4 class: small models/image inference
Low-end 16GB: RunPod Serverless 16GB (A4000/A4500/RTX 4000/RTX 2000) 0.58/hour → 58/174/423.40; Modal T4 (+4vCPU+16GiB, about 0.8126/hour) 81.26/243.78/593.20; Replicate T4 0.81/hour → 81/243/591.30. RunPod is clearly cheapest in this class; Modal and Replicate are almost the same. Note the 16GB GPUs are not equivalent (T4 vs A4000/A4500/RTX 4000); calibrate with model memory/throughput, not just “16GB”.
| 16GB/T4 (hours/month) | RunPod 16GB class | Modal T4 | Replicate T4 |
|---|---|---|---|
| 100 | 58.00 | 81.26 | 81.00 |
| 300 | 174.00 | 243.78 | 243.00 |
| 730 | 423.40 | 593.20 | 591.30 |
Editorial calculation, not a quote. RunPod Serverless 16GB class 0.58/hour (A4000/A4500/RTX 4000/RTX 2000); Modal T4 0.000164/s + 4vCPU/16GiB ≈ 0.00006172/s, total 0.00022572/s ≈ 0.8126/hour; Replicate T4 0.000225/s (0.81/hour, including 4x CPU/16GB RAM). GPUs are not performance-equivalent; excludes tax, storage, egress, model/API fees and commitment discounts.
Which option fits which scenario?
Sustained/always-on, lowest per-GPU-hour: RunPod on-demand (A100 1.79, H100 SXM 3.99); spiky/per-second/scale-to-zero in the Modal ecosystem: Modal (A100 about 2.72, H100 about 4.17 including CPU/memory); want a RunPod Serverless API/endpoint with per-second billing: RunPod Serverless; want a model marketplace/ready API without deploying: Replicate; low-end 16GB small models: RunPod 16GB class at 0.58; enterprise compliance/reserved/commitment: all three negotiate. Calculate on actual active GPU-hours, not a default 730; include cold start, queueing, storage/egress and model/API fees.
| Scenario | First choice | A100/H100 (USD/hour) | Check first |
|---|---|---|---|
| Sustained/always-on | RunPod on-demand | 1.79 / 3.99 | Region/inventory/reserved |
| Spiky/per-second/scale-to-zero | Modal | About 2.72 / 4.17 (incl. CPU/memory) | Cold start/concurrency/credits |
| Serverless API/endpoint | RunPod Serverless | 2.72 / 5.39 | Cold start/limits/storage |
| Model marketplace/ready API | Replicate | 5.04 / 5.49 | Model price/latency/quota |
| 16GB small models | RunPod 16GB class | 0.58 | Memory/compute/model |
| Enterprise compliance/reserved | All three contracts | Custom | SLA/compliance/commitment |
Sources: the three official pricing pages (accessed 7 November 2026). This is an editorial decision matrix, not vendor advice; throughput, cold start, queueing, availability, SLA, GPU equivalence and migration cost were not tested.
What should procurement and acceptance specify?
At least ten things: model/memory, GPU model/count, monthly active GPU-hours, peak concurrency, cold start/queueing, CPU/memory, storage/model cache, egress, SLA/compliance, contract/reserved. Run a 2–4 week POC on real inference and record GPU-hours, cold start, throughput, failure rate, storage/egress and the bill; do not compare only “per GPU-hour”.
- Model: parameter count/memory, precision, batch size, context length.
- GPU: A100/H100/L40S/T4, single/multi-GPU, region.
- Usage: monthly active GPU-hours, peaks, idle ratio.
- Cold start: first load, model cache, warm-up, queueing.
- CPU/memory: included in the GPU price or extra; over/under-provisioning.
- Storage: model/data/cache, volumes, cross-region.
- Egress: API responses/downloads, CDN, cross-region.
- SLA/compliance: uptime, residency, audit, SOC2.
- Contract: per-second/hour, reserved/commitment, renewal.
- POC: 2–4 weeks of real inference; GPU-hours, cold start, throughput, failure rate and bill.
FAQ
Which is cheapest?
Per GPU-hour: RunPod on-demand is cheapest (A100 1.79, H100 SXM 3.99); among serverless, Modal A100 about 2.72, H100 about 4.17, RunPod Serverless 2.72/5.39, Replicate 5.04/5.49. Low-end 16GB: RunPod 0.58 is lowest; Modal/Replicate T4 about 0.81.
Why is RunPod Serverless more expensive than on-demand?
On-demand Pods are wholesale rates for a long-held instance; serverless bills per second, autoscales and charges per request, so the platform spreads scheduling/cold-start/idle capacity into the unit price. Sustained always-on favors on-demand; spiky/low-frequency favors serverless.
Does Modal’s GPU price include CPU/memory?
No. Modal’s GPU price is GPU only; CPU is 0.0000131/core/s and memory 0.00000222/GiB/s. This article assumes 4vCPU (2 physical cores) + 16GiB, about 0.22/hour, for a total of about 2.72 (A100) or 4.17 (H100); adjust to your actual ratio.
Why is Replicate the most expensive?
Replicate’s hardware price includes the listed CPU/RAM (A100 10x CPU/144GB RAM, H100 13x CPU/144GB RAM) plus the model marketplace/ready API/managed experience. For bare GPUs, RunPod/Modal are cheaper; if you do not want to deploy models and want a ready API, the premium may be worth it.
How do we choose 100/300/730 hours?
730 hours is always-on; 100/300 is 3–10 active hours/day. Calculate on actual active GPU-hours, not a default always-on month. Always-on favors RunPod on-demand/reserved; spiky favors Modal/RunPod Serverless; Replicate fits model/API calls rather than long-held GPUs.
How do cold start/idle count?
Serverless bills active seconds and does not charge GPU when idle; but cold start adds latency and may trigger more instances/retries. On-demand Pods bill hourly even when idle; low utilization favors serverless, high utilization favors on-demand.
What hidden costs should we expect?
Storage/model cache, egress, CPU/memory over-provisioning, cold start/retries, model/API fees, support/SLA, compliance/audit, migration/dual-run and tax. Replicate/Modal model/API fees may be separate from hardware; RunPod storage/network are separate too.
How do we choose and accept?
Sustained/always-on: RunPod on-demand; spiky/scale-to-zero: Modal; serverless endpoint: RunPod Serverless; model marketplace: Replicate; 16GB small models: RunPod 0.58. A 2–4 week POC on real inference records GPU-hours, cold start, throughput, failure rate and the bill.
Research method and limitations
This article separates “official facts” from “editorial modelling”: RunPod’s on-demand A100 1.79/H100 SXM 3.99, Serverless A100 2.72/H100 5.39 and 16GB class 0.58 come from RunPod Pricing; Modal’s A100 0.000694/s, H100 0.001097/s, T4 0.000164/s, CPU 0.0000131/core/s, memory 0.00000222/GiB/s and Starter/Team credits come from Modal Pricing; Replicate’s A100 0.0014/s, H100 0.001525/s and T4 0.000225/s (including CPU/RAM) come from Replicate Pricing. The TCOs are arithmetic at 100/300/730 GPU-hours and a 4vCPU/16GiB assumption, not vendor quotes.
Eight limitations: (1) only US public list prices are used, excluding tax, commitments/reserved/enterprise discounts, free credits and regional differences; (2) only GPU/CPU/memory are counted, excluding storage, egress, model/API fees, support and migration; (3) Modal CPU/memory is estimated at 4vCPU/16GiB; the actual ratio changes the total; (4) RunPod’s 16GB class is A4000/A4500/RTX 4000/RTX 2000, not exactly T4; (5) throughput, cold start, queueing, availability, SLA and GPU equivalence were not tested; (6) AWS/GCP/Azure GPUs, Lambda/Fargate and other serverless GPU vendors are excluded; (7) migration, model conversion, cache/warm-up and dual-run are excluded; (8) prices and terms can change by contract, region and exchange rate. Treat this as a per-GPU-hour conversion template and replace assumptions with a 2–4 week real-inference POC.
Sources and basis
- RunPod, Pricing: runpod.io/pricing. On-demand A100 1.79, H100 SXM 3.99, A40 0.59, L40S 1.09; Serverless A100 2.72, H100 5.39, L40S 1.75, 16GB class 0.58, H200 6.99, B200 10.79. Accessed 7 November 2026.
- Modal, Pricing: modal.com/pricing. A100 80GB 0.000694/s, H100 0.001097/s, T4 0.000164/s, L40S 0.000542/s; CPU 0.0000131/core/s, memory 0.00000222/GiB/s; Starter 0/month + $30 credit, Team 250/month + $100 credit. Accessed 7 November 2026.
- Replicate, Pricing: replicate.com/pricing. A100 80GB 0.0014/s (5.04/hour, including 10x CPU/144GB RAM), H100 0.001525/s (5.49/hour), T4 0.000225/s (0.81/hour), L40S 0.000975/s (3.51/hour). Accessed 7 November 2026.
Data cut-off 7 November 2026; prices are public list prices excluding tax; all models are editorial calculations. This is not procurement advice.