How Much Does Modal Cost for GPU Inference and Training? (2026 Rates)

iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more

Modal is pay-as-you-go. You pay for GPU time by the second, then add metered CPU, memory, volume storage and network egress. There is no single price for a training run: estimate GPU rate × number of GPUs × runtime, then include the supporting resources your workload requests. The official 2026 pricing page lists rates from $0.000164 per GPU-second for a T4 to $0.001972 for a B300. Starter has no platform fee and includes $30 of monthly compute credits; Team costs $250 per month and includes $100 of credits.

Modal’s pricing model in one minute

Modal runs inference, training and fine-tuning as serverless workloads. A GPU is metered for the time your container requests it, at the rate for the selected model. CPU cores, memory, persistent volumes and outbound network transfer are separate line items. Because billing is usage-based, a short inference request and a multi-hour training job are calculated from the same basic formula rather than from a fixed job fee.

#	Preview	Product	Price
1		ASRock Radeon AI PRO R9700 Creator 32GB Professional Graphics Card, 2920 MHz Boost Clock, GDDR6, AMD...	$1,699.99	Buy on Amazon
2		Kinupute Mini PC AI Server, AI Computing Workstation, AI MAX+ 395(126TOPS,16C/32T), Win-11 Pro,...	$4,648.00	Buy on Amazon
3		ASUS ESC8000A-E13 4U AI GPU Server Barebones with 3+1 3200W Titanimum CRPS Supporting Eight (8)...	$9,499.99	Buy on Amazon
4		ASUS Pro WS WRX90E-SAGE SE EEB Workstation Motherboard, AMD Ryzen™ Threadripper™ PRO 7000...	$1,169.05	Buy on Amazon

Modal states that customers “always pay for what you use and nothing more” on its pricing page. Its billing documentation similarly describes the service as serverless: you pay for compute used or requested, with no reservation requirement and no minimum usage-time increment.

GPU rates and hourly equivalents

Modal publishes per-second GPU prices. The hourly figures below are direct conversions at 3,600 seconds per hour; they are not separate billing tiers.

#1 Best Overall

ASRock Radeon AI PRO R9700 Creator 32GB Professional Graphics Card, 2920 MHz Boost Clock, GDDR6, AMD RDNA 4, AI-Accelerators, DisplayPort 2.1a, PCIe 5.0, Blower Cooler

• Professional AI & Creator Workstation: AMD Radeon AI PRO R9700 GPU with 32GB GDDR6 is engineered for AI development, professional content creation, and compute-intensive workloads.
• Massive 32GB Memory Capacity: 32GB of GDDR6 memory on a 256-bit bus provides ample bandwidth for large AI models, 8K video editing, and complex 3D rendering.
• Advanced RDNA 4 with AI Accelerators: 64 Compute Units with 3rd Gen Ray Tracing and dedicated 2nd Gen AI Accelerators for groundbreaking AI performance and visual computing.
• Professional Blower Cooling: Efficient single blower design exhausts heat directly out of the chassis, ideal for multi-GPU workstation and server configurations.
• Enterprise-Grade Thermal Solution: Vapor chamber heatsink with industrial Honeywell PTM7950 thermal interface material ensures reliable cooling under sustained professional loads.

$1,699.99

Buy on Amazon

For example, one H100 SXM5 running for two hours is approximately 7,200 seconds × $0.001097, or $7.8984 in GPU charges. Four H100s for 10 hours consume 144,000 GPU-seconds and cost approximately $157.968 before CPU, memory, storage or egress. Actual totals depend on the resources and runtime your application uses.

How to calculate a Modal bill

1. Calculate GPU time

Use this equation for each GPU type:

GPU charge = rate per second × runtime in seconds × number of GPUs

Convert minutes or hours to seconds before multiplying. If a job uses different GPU types in separate stages, calculate each stage separately and add the results. A distributed training run that uses eight H100s for six hours, for instance, uses 172,800 GPU-seconds (8 × 6 × 3,600) and incurs about $189.5616 in H100 GPU charges.

2. Add CPU and memory

Modal lists CPU at $0.0000131 per physical core-second and memory at $0.00000222 per GiB-second. These charges follow the resources allocated to the container, not just the amount of GPU memory. A workload that keeps a large CPU process or data loader alive can therefore cost more than a GPU-only estimate suggests.

3. Add storage and network transfer

Persistent volumes are listed at $0.09 per GiB-month. Outbound network egress is $0.04 per GiB, subject to plan-specific allowances. Include model downloads, checkpoint uploads, dataset transfers and generated outputs when estimating egress. Modal also lists up to 1 TiB per month of volume-storage allowance; check the allowance attached to your workspace and plan before assuming all storage is free.

Rank #2

Kinupute Mini PC AI Server, AI Computing Workstation, AI MAX+ 395(126TOPS,16C/32T), Win-11 Pro, Radeon 8060S GPU, 128G LPDDR5X-8400, 8T M.2 SSD, 10G+2.5G LAN, Quad Screen, 4xM.2 PCIe 4.0 Slots, WiFi 7

• 【AI Max+ 395 AI Workstation】16 cores, 32 threads, up to 5.1 GHz boost and 80 MB cache. Integrated Radeon 8060S graphics with 40 CUs, RDNA 3.5, delivers performance close to RTX 4060/4070 laptop GPUs. Triple-engine design(CPU+GPU+XDNA 2 NPU) with up to 126 TOPS total, including 50+ TOPS dedicated NPU for local AI inference and machine learning acceleration. Ideal for AI development, content creation, virtualization, data analysis, and demanding multitasking. Compact, high-performance workstation.
• 【256-bit LPDDR5X MAX 128GB】The LPDDR5X onboard memory reaches 8400 MT/s - 1.5x faster than DDR5 SODIMM. Unlock the full potential of your graphics with massive 128GB memory pooling. This system allows you to manually assign up to 128GB of the onboard RAM to serve as video memory (VRAM) directly within the BIOS setup, delivering unparalleled performance for 4K video editing, and AI model training without the need for a discrete graphics card.
• 【Lastest GPU 8060S & XDNA 2 NPU】Built on the RDNA 3.5 architecture, the AMD Radeon 8060S Graphics iGPU features 40 compute units (2,560 stream processors). It delivers performance on par with NVIDIA's mobile RTX 4070, efficient encoding/decoding for AVC, HEVC, VP9, and AV1 video codecs. And It can connect 4 screens via HDMI & DisplayPort & Full Featured USB4 x2 to efficiently handle your tasks and meet your specific needs. Supports 8K/4K resolution displays.
• 【Dual LAN (2.5GbE+10GbE)& WiFi 7】The computer has double LAN, one is 2.5GbE (I226), the other is 10GbE(AQC113). provides more applications, such as firewall, soft routing, multichannel aggregation. Built-in WiFi module, support WiFi 7 and Bluetooth5.4. Known as 802.11be, Wi-Fi 7 promises up to 46Gbps theoretical throughput, making it 4.8x faster than Wi-Fi 6. and computer has 4 built-in NVMe SSD slots, 1 SD card slot, allowing you to expand its storage capacity.
• 【Engineered to Endure】The computer measures 7.13 x 7.24 x 2.99 inches. AI mini pc is encased in a premium all-aluminium chassis. Dual turbo CPU fans deliver silent, ultra-efficient cooling, To enable the computer to maintain stable operation for a long time. We offer up to 2 years warranty and lifetime professional customer service. Please feel free to contact us if any issues happened. thanks

$4,648.00

Buy on Amazon

4. Subtract applicable credits or allowances

Plan credits reduce eligible compute charges, while free allowances can cover defined amounts of storage or egress. They do not turn every resource into an unlimited free service. Review the workspace’s usage and budget settings rather than treating the GPU table as an invoice total.

What inference costs on Modal

Custom inference containers

For a custom inference stack, choose a GPU and pay for the time that GPU is requested. The runtime can include model initialization, request processing and any period in which the container remains active. CPU and memory allocations are metered alongside the GPU. Concurrency, batching, container lifetime and model-loading behavior determine how much work each billed GPU-second accomplishes.

A useful estimate is:

1. Measure or conservatively estimate the time a container stays allocated for one request or batch.
2. Multiply that time by the selected GPU’s per-second rate.
3. Include the CPU cores and GiB of memory allocated during the same interval.
4. Account for cold-start downloads, persistent volume use and outbound data.
5. Divide by requests, images or tokens served to obtain a unit cost.

Modal’s Stable Diffusion example

Modal’s official example describes a Stable Diffusion task on an A10G taking about 1.5 seconds per image. It reports approximately $0.000458 in GPU charges, $0.000020 in CPU charges and $0.000013 in memory charges, for a total of $0.000491 per image, or $0.491 per 1,000 images. This is an illustrative configuration, not a guaranteed quote. Model-load time, container lifetime, concurrency and resource choices can change the result.

Shared Endpoints

Modal also offers Shared Endpoints with token-based pricing. Token usage is billed from the first request and is separate from the compute credits included with the Starter plan. Do not apply a GPU-second estimate to a Shared Endpoint invoice; use the endpoint’s published token price and then account for any other resources your architecture separately. Details are on Modal’s inference product page.

What training and fine-tuning cost

Training and fine-tuning use the same serverless billing model, so there is no universal “price per training job.” A credible estimate needs at least the GPU model, GPU count, elapsed time, CPU and memory allocations, storage footprint and data transfer.

Rank #3

ASUS ESC8000A-E13 4U AI GPU Server Barebones with 3+1 3200W Titanimum CRPS Supporting Eight (8) 2-Slot Server GPUs (e.g. Pro 6000, H200), Dual (2) EPYC 9005 CPUs & 24-Channels of DDR5 ECC RDIMM RAM

• [ Maximum AI Compute Power ] Dominate complex workloads with the ASUS ESC8000A-E13. This 4U rack server is a powerhouse engineered for mass-scale AI, machine learning, and deep training. Featuring support for dual AMD EPYC 9005/9004 processors and up to eight dual-slot GPUs, it delivers the raw computational muscle required to train LLMs and run complex simulations effortlessly. Accelerate your data science pipeline and transform raw data into actionable intelligence faster than ever.
• [ Advanced Thermal Efficiency ] High performance demands elite cooling. The ESC8000A-E13 features a cutting-edge aerodynamic design with independent CPU and GPU airflow tunnels. Equipped with redundant hot-swap fans and optimized for liquid cooling integrations, this 4U server ensures maximum uptime under heavy, sustained workloads. Keep your data center running cool, quiet, and highly efficient while preventing thermal throttling during mission-critical enterprise operations.
• [ Scale with Flexible Storage ] Future-proof your infrastructure with unmatched storage and expansion flexibility. This offers comprehensive front-panel drive bays supporting Gen5 NVMe, SAS, or SATA drives alongside multiple PCIe 5.0 slots. Designed as a high-density 4U server capable of housing eight dual-slot GPUs: NVD H200, RTX PRO 6000 Blackwell, RTX PRO 4500 Blackwell or AMD Instinct MI350P PCIe Card, each supporting up to 600 watts.
• [ Enterprise-Grade Reliability ] Minimize downtime and secure your ecosystem with server-grade redundancy. The ESC8000A-E13 is built for 24/7 continuous operation, boasting 2+2 redundant (3200W total) 80 PLUS Titanium power supplies and integrated ASUS ASMB11-iKVM for comprehensive out-of-band management. Ideal for cloud service providers, rendering farms, and large enterprise infrastructure, it combines robust physical hardware with smart remote monitoring to safeguard your digital assets.
• [Reliability Guaranteed] Shop with total peace of mind knowing that every new computer component we sell is backed by our EPC 3-year warranty. Whether you are investing in high-speed DDR5 RAM or a powerhouse GPU, we protect your build against defects and performance failures. We stand firmly behind the quality of our hardware, ensuring that your setup remains fast, stable, and secure for years to come.

$9,499.99

Buy on Amazon

A practical training estimate

1. Define the hardware. Record the GPU type and number of GPUs per worker. Match the type to model-memory and throughput requirements rather than choosing by hourly price alone.
2. Measure a representative step. Run enough batches to include data loading, communication and checkpoint behavior. Extrapolate from measured throughput to the target number of steps or epochs.
3. Include startup and teardown. Container initialization, image pulls, dataset staging and checkpoint uploads can extend billed runtime.
4. Price every stage. Multiply each stage’s GPU-seconds by its GPU rate, then add CPU, memory, volume and egress charges.
5. Add a contingency. Failed runs, hyperparameter trials and validation jobs consume additional time; keep them as separate scenarios instead of hiding them in an optimistic single number.

As an illustration, a 12-hour run on two H200 SXM GPUs uses 86,400 GPU-seconds (2 × 12 × 3,600). At $0.001261 per second, its GPU portion is approximately $108.9504. CPU, memory, storage and egress are extra. If a later fine-tuning stage uses one A100 80 GB for three hours, that stage adds approximately $26.9824 (10,800 × $0.000694) before supporting resources.

Why training estimates vary

• Scaling efficiency: doubling GPUs can reduce elapsed time without halving the GPU bill; communication overhead may leave some GPUs underused.
• Checkpoint frequency: frequent uploads increase storage operations, runtime and potentially egress.
• Input pipeline: insufficient CPU, memory or storage throughput can keep an expensive GPU idle while those supporting resources continue billing.
• Retries and experiments: preempted or failed jobs still consume the resources used before failure, so budget for multiple runs when tuning.

Plans, free credits and platform fees

The figures above are Modal’s 2026 plan terms. Compute beyond included credits is charged at the usage rates. The Team platform fee is due even if compute use is low, so compare it with the value of unlimited seats, higher limits and the $100 credit. Enterprise pricing is negotiated; Modal’s AWS Marketplace listing directs enterprise buyers to contact Modal rather than publishing a fixed rate.

Other metered resources

These rates are listed on Modal’s pricing page. The page also describes free allowances, including up to 1 TiB per month of volume storage and plan-specific network-egress allowances. Confirm the current allowance for your workspace because an allowance does not necessarily apply identically to every plan or resource.

Billing mechanics and spend controls

Modal bills workspaces monthly. Its billing documentation says incremental charges are applied when usage crosses the specified thresholds. A payment method is required to use GPUs. The same documentation says reservations are not required and there are no minimum usage-time increments, so a workload is not forced into a one-minute or one-hour block merely because it ran briefly.

Budgets

Use workspace-level and environment-level budgets to limit exposure. Separate development, staging and production environments so an experiment cannot silently consume the production budget. Set an alert or cap before launching a multi-GPU job, then inspect usage after the first small run and update the estimate with measured runtime.

Rank #4

Sale

ASUS Pro WS WRX90E-SAGE SE EEB Workstation Motherboard, AMD Ryzen™ Threadripper™ PRO 7000 WX-Series, ECC R-DIMM DDR5, 32 Power-Stage,7xPCIe 5.0x16, PCIe 5.0 M.2, 10Gb & 2.5Gb LAN, Multi-GPU Support

• AMD socket sTR5 supports up to 96-core CPUs: Ready for AMD Ryzen Threadripper PRO 7000 WX-Series Processors.
• Ultrafast connectivity:Seven PCIe 5.0 x16 slots, dual 10 Gb LAN ports, four M.2 slots, two rear USB4 40Gbps Type-C and SlimSAS NVMe support.
• CPU and memory overclocking: Support for up to 2TB ECC R-DIMM DDR5 memory modules (1DPC)
• Robust power and thermal design: 32 power stages with two 8-pin power connectors for the CPU, massive VRM cooling, chipset and M.2 heatsinks with active fans, and M.2 thermal pad.
• PCIe Q-release Slim: Remove the graphics card by directly pulling it up, instead of pressing a PCIe latch.

$1,169.05

Buy on Amazon

What a bill can include

• GPU seconds for every allocated GPU, including multiple workers.
• CPU core-seconds and GiB-seconds of memory.
• Volume GiB-months for data that remains stored.
• Network egress beyond the applicable allowance.
• Any platform fee for the selected plan, offset by that plan’s included compute credits where eligible.

Ways to reduce inference and training cost

Choose hardware by total work, not hourly rate

A cheaper GPU can cost more overall if it takes substantially longer to finish a workload. Benchmark a representative batch on two candidate GPUs and compare total GPU-seconds, not just the per-hour number. Conversely, a premium GPU may be wasteful when the model is small or requests are sparse.

Keep expensive GPUs busy

Batch compatible inference requests, use dynamic batching where latency permits and size CPU and memory so the input pipeline does not starve the GPU. For training, monitor utilization and communication overhead; adding GPUs without useful scaling increases the bill.

Control container lifetime and data movement

Load models once per container when traffic justifies it, but avoid keeping idle capacity alive indefinitely. Cache datasets and model weights in a volume when repeated downloads would add runtime or egress. Delete obsolete checkpoints and cap output transfers. Measure cold-start and warm-container behavior because both affect the billed seconds in a serverless system.

Use budgets during experimentation

Run a short pilot, record actual GPU, CPU and memory usage, and extrapolate before launching a full sweep. Apply environment budgets to notebooks and test jobs. Treat retries, failed runs and validation as explicit rows in the estimate rather than assuming the first run is the only run.

How Modal compares on the dimensions that affect cost

If you are evaluating another GPU platform, compare like for like. The relevant questions are:

• Is the same GPU model available, and is the quoted price per GPU-second, per minute, per hour or per request?
• Does billing include cold-start, idle-container or reserved capacity time?
• Are CPU, memory, storage and egress metered separately, and what allowances apply?
• What concurrency, region and quota limits affect how many jobs can run at once?
• Is inference priced by GPU runtime or by tokens, and are training and inference on the same billing model?
• Are free credits recurring monthly, one-time, promotional or restricted to particular resources?

Modal’s distinguishing points for this calculation are per-second serverless metering, no reservation requirement, no minimum usage-time increment, plan credits and a separate token-priced Shared Endpoint option. Availability of a particular GPU or region can change over time, so verify it in the workspace before committing to a schedule.

Bottom line for budgeting

Start with the GPU-second equation, then add CPU, memory, storage, egress and the plan fee. For a quick upper bound, multiply the published hourly equivalent by GPU count and expected hours; for a production estimate, measure a representative run that includes startup, data loading and checkpoints. Starter can cover light experimentation with its $30 monthly compute credit, while Team’s $250 fee and $100 credit make sense only when its seats and higher limits justify the fixed cost. Training has no fixed sticker price: the model, hardware, scaling efficiency and runtime determine the invoice.

Quick Recap

Bestseller No. 1

ASRock Radeon AI PRO R9700 Creator 32GB Professional Graphics Card, 2920 MHz Boost Clock, GDDR6, AMD RDNA 4, AI-Accelerators, DisplayPort 2.1a, PCIe 5.0, Blower Cooler

$1,699.99

Bestseller No. 2

Kinupute Mini PC AI Server, AI Computing Workstation, AI MAX+ 395(126TOPS,16C/32T), Win-11 Pro, Radeon 8060S GPU, 128G LPDDR5X-8400, 8T M.2 SSD, 10G+2.5G LAN, Quad Screen, 4xM.2 PCIe 4.0 Slots, WiFi 7

$4,648.00

Bestseller No. 3

ASUS ESC8000A-E13 4U AI GPU Server Barebones with 3+1 3200W Titanimum CRPS Supporting Eight (8) 2-Slot Server GPUs (e.g. Pro 6000, H200), Dual (2) EPYC 9005 CPUs & 24-Channels of DDR5 ECC RDIMM RAM

$9,499.99

SaleBestseller No. 4

ASUS Pro WS WRX90E-SAGE SE EEB Workstation Motherboard, AMD Ryzen™ Threadripper™ PRO 7000 WX-Series, ECC R-DIMM DDR5, 32 Power-Stage,7xPCIe 5.0x16, PCIe 5.0 M.2, 10Gb & 2.5Gb LAN, Multi-GPU Support

CPU and memory overclocking: Support for up to 2TB ECC R-DIMM DDR5 memory modules (1DPC)

$1,169.05

Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.