Pricing

AI Platform offers scalable, flexible pricing options to fit your project and budget. AI Platform charges you for training your models and for getting predictions, but managing your machine learning resources in the cloud is free of charge.

Pricing overview

The following tables summarize the pricing for training and prediction in each region where AI Platform is available.

Training prices

The tables below provide the price per hour of various training configurations, as well as the number of training units2 used by each configuration. Training units measure the resource usage of your job; the price per hour of a machine configuration is the number of training units it uses multiplied by the region's cost of training.

You can choose a predefined scale tier or a custom configuration of selected machine types. If you choose a custom configuration, sum the costs of the virtual machines you use.

Accelerator-enabled legacy machine types include the cost of the accelerators in their pricing. If you use Compute Engine machine types and attach accelerators, the cost of the accelerators is separate. To calculate this cost, multiply the prices in the table of accelerators below by how many of each type of accelerator you use.

Americas

The cost of a training job in all available Americas regions is $0.49 per hour, per training unit.

Predefined scale tiers - price per hour (training units)
BASIC $0.1900 (0.3878)
STANDARD_1 $1.9880 (4.0571)
PREMIUM_1 $16.5536 (33.7829)
BASIC_GPU $0.8300 (1.6939)
BASIC_TPU $4.6900 (9.5714)
CUSTOM If you select CUSTOM as your scale tier, you have control over the number and type of virtual machines used for your training job. See the table of machine types.
Machine types - price per hour (training units)
n1-standard-4 $0.1900 (0.3878)
n1-standard-8 $0.3800 (0.7755)
n1-standard-16 $0.7600 (1.5510)
n1-standard-32 $1.5200 (3.1020)
n1-standard-64 $3.0400 (6.2041)
n1-standard-96 $4.5600 (9.3061)
n1-highmem-2 $0.1184 (0.2416)
n1-highmem-4 $0.2368 (0.4833)
n1-highmem-8 $0.4736 (0.9665)
n1-highmem-16 $0.9472 (1.9331)
n1-highmem-32 $1.8944 (3.8661)
n1-highmem-64 $3.7888 (7.7322)
n1-highmem-96 $5.6832 (11.5984)
n1-highcpu-16 $0.5672 (1.1576)
n1-highcpu-32 $1.1344 (2.3151)
n1-highcpu-64 $2.2688 (4.6302)
n1-highcpu-96 $3.4020 (6.9429)
cloud_tpu6 $4.5000 (9.1840) or N/A if accelerators are attached explicitly6
standard $0.1900 (0.3878)
large_model $0.4736 (0.9665)
complex_model_s $0.2836 (0.5788)
complex_model_m $0.5672 (1.1576)
complex_model_l $1.1344 (2.3151)
standard_gpu $0.8300 (1.6939)
complex_model_m_gpu $2.5600 (5.2245)
complex_model_l_gpu $3.3200 (6.7755)
standard_p100 $1.8400 (3.7551)
complex_model_m_p100 $6.6000 (13.4694)
standard_v100 $2.8600 (5.8367)
large_model_v100 $2.9536 (6.0278)
complex_model_m_v100 $10.6800 (21.7959)
complex_model_l_v100 $21.3600 (43.5918)
Accelerators - price per hour (training units)
NVIDIA_TESLA_K80 $0.4500 (0.9184)
NVIDIA_TESLA_P4 $0.6000 (1.2245)
NVIDIA_TESLA_P100 $1.4600 (2.9796)
NVIDIA_TESLA_T4 $0.9500 (1.9388)
NVIDIA_TESLA_V100 $2.4800 (5.0612)
Eight TPU_V2 cores6 $4.5000 (9.1840)
Eight TPU_V3 cores6 (Beta) $8.0000 (16.3264)

Europe

The cost of a training job in all available Europe regions is $0.54 per hour, per training unit.

Predefined scale tiers - price per hour (training units)
BASIC $0.2200 (0.4074)
STANDARD_1 $2.3020 (4.2630)
PREMIUM_1 $19.1640 (35.4889)
BASIC_GPU $0.9300 (1.7222)
BASIC_TPU (Not available)
CUSTOM If you select CUSTOM as your scale tier, you have control over the number and type of virtual machines used for your training job. See the table of machine types.
Machine types - price per hour (training units)
n1-standard-4 $0.2200 (0.4074)
n1-standard-8 $0.4400 (0.8148)
n1-standard-16 $0.8800 (1.6296)
n1-standard-32 $1.7600 (3.2593)
n1-standard-64 $3.5200 (6.5185)
n1-standard-96 $5.2800 (9.7778)
n1-highmem-2 $0.1370 (0.2537)
n1-highmem-4 $0.2740 (0.5074)
n1-highmem-8 $0.5480 (1.0148)
n1-highmem-16 $1.0960 (2.0296)
n1-highmem-32 $2.1920 (4.0593)
n1-highmem-64 $4.3840 (8.1185)
n1-highmem-96 $6.5760 (12.1778)
n1-highcpu-16 $0.6568 (1.2163)
n1-highcpu-32 $1.3136 (2.4326)
n1-highcpu-64 $2.6272 (4.8652)
n1-highcpu-96 $3.9408 (7.2978)
cloud_tpu6 (Not available)
standard $0.2200 (0.4074)
large_model $0.5480 (1.0148)
complex_model_s $0.3284 (0.6081)
complex_model_m $0.6568 (1.2163)
complex_model_l $1.3136 (2.4326)
standard_gpu $0.9300 (1.7222)
complex_model_m_gpu $2.8400 (5.2593)
complex_model_l_gpu $3.7200 (6.8889)
standard_p100 $2.0400 (3.7778)
complex_model_m_p100 $7.2800 (13.4815)
standard_v100 $2.9684 (5.4970)
large_model_v100 $3.0708 (5.6867)
complex_model_m_v100 $11.0368 (20.4385)
complex_model_l_v100 $22.0736 (40.8770)
Accelerators - price per hour (training units)
NVIDIA_TESLA_K80 $0.4900 (0.9074)
NVIDIA_TESLA_P4 $0.6500 (1.2037)
NVIDIA_TESLA_P100 $1.6000 (2.9630)
NVIDIA_TESLA_T4 $1.0300 (1.9074)
NVIDIA_TESLA_V100 $2.5500 (4.7222)
Eight TPU_V2 cores6 (Not available)
Eight TPU_V3 cores6 (Beta) (Not available)

Asia Pacific

The cost of a training job in all available Asia Pacific regions is $0.54 per hour, per training unit.

Predefined scale tiers - price per hour (training units)
BASIC $0.2200 (0.4074)
STANDARD_1 $2.3020 (4.2630)
PREMIUM_1 $19.1640 (35.4889)
BASIC_GPU $0.9300 (1.7222)
BASIC_TPU (Not available)
CUSTOM If you select CUSTOM as your scale tier, you have control over the number and type of virtual machines used for your training job. See the table of machine types.
Machine types - price per hour (training units)
n1-standard-4 $0.2200 (0.4074)
n1-standard-8 $0.4400 (0.8148)
n1-standard-16 $0.8800 (1.6296)
n1-standard-32 $1.7600 (3.2593)
n1-standard-64 $3.5200 (6.5185)
n1-standard-96 $5.2800 (9.7778)
n1-highmem-2 $0.1370 (0.2537)
n1-highmem-4 $0.2740 (0.5074)
n1-highmem-8 $0.5480 (1.0148)
n1-highmem-16 $1.0960 (2.0296)
n1-highmem-32 $2.1920 (4.0593)
n1-highmem-64 $4.3840 (8.1185)
n1-highmem-96 $6.5760 (12.1778)
n1-highcpu-16 $0.6568 (1.2163)
n1-highcpu-32 $1.3136 (2.4326)
n1-highcpu-64 $2.6272 (4.8652)
n1-highcpu-96 $3.9408 (7.2978)
cloud_tpu6 (Not available)
standard $0.2200 (0.4074)
large_model $0.5480 (1.0148)
complex_model_s $0.3284 (0.6081)
complex_model_m $0.6568 (1.2163)
complex_model_l $1.3136 (2.4326)
standard_gpu $0.9300 (1.7222)
complex_model_m_gpu $2.8400 (5.2593)
complex_model_l_gpu $3.7200 (6.8889)
standard_p100 $2.0400 (3.7778)
complex_model_m_p100 $7.2800 (13.4815)
standard_v100 $2.9684 (5.4970)
large_model_v100 $3.0708 (5.6867)
complex_model_m_v100 $11.0368 (20.4385)
complex_model_l_v100 $22.0736 (40.8770)
Accelerators - price per hour (training units)
NVIDIA_TESLA_K80 $0.4900 (0.9074)
NVIDIA_TESLA_P4 $0.6500 (1.2037)
NVIDIA_TESLA_P100 $1.6000 (2.9630)
NVIDIA_TESLA_T4 $1.0300 (1.9074)
NVIDIA_TESLA_V100 $2.5500 (4.7222)
Eight TPU_V2 cores6 (Not available)
Eight TPU_V3 cores6 (Beta) (Not available)

Prediction prices

This table provides the prices of batch prediction and online prediction per node hour. A node hour represents the time a virtual machine spends running your prediction job or waiting in a ready state to handle prediction requests. Read more about calculating prediction costs.

Americas

Prediction
Batch prediction $0.0791 per node hour
Online prediction
Machine types - price per node hour
mls1-c1-m2 (default)

$0.0401

mls1-c4-m2 (Beta)

$0.1349

n1-standard-2 (Beta)

approximately $0.0950

n1-standard-4 (Beta)

approximately $0.1900

n1-standard-8 (Beta)

approximately $0.3800

n1-standard-16 (Beta)

approximately $0.7600

n1-standard-32 (Beta)

approximately $1.5200

n1-highmem-2 (Beta)

approximately $0.1183

n1-highmem-4 (Beta)

approximately $0.2366

n1-highmem-8 (Beta)

approximately $0.4732

n1-highmem-16 (Beta)

approximately $0.9464

n1-highmem-32 (Beta)

approximately $1.8928

n1-highcpu-2 (Beta)

approximately $0.0708

n1-highcpu-4 (Beta)

approximately $0.1417

n1-highcpu-8 (Beta)

approximately $0.2834

n1-highcpu-16 (Beta)

approximately $0.5668

n1-highcpu-32 (Beta)

approximately $1.1336

Compute Engine (N1) machine types (beta) are only available in the us-central1 region, and their pricing is charged as two separate SKUs on your Google Cloud bill:

  • vCPU cost, measured in vCPU hours
  • RAM cost, measured in GB hours

The prices for Compute Engine (N1) machine types in the previous table approximate the total hourly cost for each prediction node of a model version using that machine type. For example, since an n1-highcpu-32 machine type includes 32 vCPUs and 28.8 GB of RAM, the hourly pricing per node is equal to 32 vCPU hours + 28.8 GB hours.

The prices in the previous table are provided to help you estimate online prediction costs. The following table shows the vCPU and RAM pricing for Compute Engine (N1) machine types, which more precisely reflect the SKUs that you will be charged for:

Compute Engine (N1) machine type SKUs
vCPU $0.031611 per vCPU hour
RAM $0.004237 per GB hour

You can optionally use GPU accelerators for online prediction with Compute Engine (N1) machine types. GPUs incur an additional charge, separate from those described in the previous table. The following table describes the pricing for each type of GPU:

Accelerators - price per hour
NVIDIA_TESLA_K80 $0.4500
NVIDIA_TESLA_P4 $0.6000
NVIDIA_TESLA_P100 $1.4600
NVIDIA_TESLA_T4 $0.9500
NVIDIA_TESLA_V100 $2.4800

Note that the pricing is per GPU, so if you use multiple GPUs per prediction node (or if your version scales to use multiple nodes), then costs scale accordingly.

Europe

Prediction
Batch prediction $0.0861 per node hour
Online prediction
Machine types - price per node hour
mls1-c1-m2 (default)

$0.0441

mls1-c4-m2 (Beta)

$0.1484

Asia Pacific

Prediction
Batch prediction $0.0861 per node hour
Online prediction
Machine types - price per node hour
mls1-c1-m2 (default)

$0.0515

mls1-c4-m2 (Beta)

$0.1733

Notes:

  1. All use is subject to the AI Platform quota policy.
  2. Note the difference between training unit used on this page, and Consumed ML units shown on your Job Details page. The duration is already factored into the Consumed ML units. See the details below.
  3. You are required to store your data and program files in Google Cloud Storage buckets during the AI Platform lifecycle. See more about Cloud Storage usage.
  4. For volume-based discounts, contact the Sales team.
  5. If you pay in a currency other than USD, the prices listed in your currency on Cloud Platform SKUs apply.
  6. The cloud_tpu machine type currently provides a TPU v2 device with 8 cores by default, even if you don't explicitly attach any accelerators to your configuration. In this case, you pay the price listed for cloud_tpu in the machine types table.

    If you do explicitly attach TPU v2 or TPU v3 accelerators to a cloud_tpu machine type in your configuration, you do not pay for the cloud_tpu machine type itself. Instead, you pay for the TPU_V2 or TPU_V3 accelerator attachments.

The pricing calculator

Use the pricing calculator to estimate your training and prediction costs.

More about training costs

You are charged for training your models in the cloud:

  • In one-minute increments.
  • At a price per hour as shown in the above table, where the price per hour is calculated from a base price and a number of training units, determined by the processing configuration you choose when you start your training job.
  • With a minimum of 10 minutes per training job.
  • From the moment when resources are provisioned for a job until the job finishes.

Scale tiers for predefined configurations

You can control the type of processing cluster to use when training your model. The simplest way is to choose from one of the predefined configurations called scale tiers. Read more about scale tiers.

Machine types for custom configurations

If you select CUSTOM as your scale tier, you have control over the number and type of virtual machines to use for the cluster's master, worker and parameter servers. Read more about machine types.

The cost of training with a custom processing cluster is the sum of all the machines you specify. You are charged for the total time of the job, not for the active processing time of individual machines.

Examples: Calculate training cost using training units

Use training units to calculate the cost of your training job, with the following formula:

(training units * base price / 60) * job duration in minutes

Examples:

  • A data scientist in an Americas region runs a training job and selects the STANDARD_1 scale tier, which uses 4.0571 training units. Their job takes 15 minutes:

    (4.0571 training units * $0.49 per hour / 60) * 15 minutes
    

    For a total of $0.50 for the job.

  • A computer science professor in an Americas region runs a training job using the CUSTOM scale tier. They have a very large model, so they want to take advantage of the large model VMs for their parameter servers. They configure their processing cluster like this:

    • A complex_model_s machine for their master (0.5788 training units.)
    • 5 parameter servers on large_model VMs (5 @ 0.9665 = 4.8325 training units).
    • 8 workers on complex_model_s VMs (8 @ 0.5788 = 4.6304 training units).

    Their job runs for 2 hours and 26 minutes:

    (10.0417 training units * $0.49 per hour / 60) * 146 minutes
    

    For a total of $11.97 for the job.

Examples: Calculate training cost using price per hour

Instead of training units, you can use the price per hour shown in the above table. The formula is as follows:

(Price per hour / 60) * job duration in minutes

Examples:

  • A data scientist in an Americas region runs a training job and selects the STANDARD_1 scale tier. Their job takes 15 minutes:

    ($1.9880 per hour / 60) * 15 minutes
    

    For a total of $0.50 for the job.

  • A computer science professor in an Americas region runs a training job using the CUSTOM scale tier. They have a very large model, so they want to take advantage of the large model VMs for their parameter servers. They configure their processing cluster like this:

    • A complex_model_s machine for their master ($0.2836).
    • 5 parameter servers on large_model VMs (5 @ $0.4736 = $2.3680).
    • 8 workers on complex_model_s VMs (8 @ $0.2836 = $2.2688).

    Their job runs for 2 hours and 26 minutes:

    (($0.2836 + $2.368 + $2.2688) per hour / 60) * 146 minutes
    

    For a total of $11.97 for the job.

Examples: Calculate training cost using "Consumed ML units"

The Consumed ML units (Consumed Machine Learning units) shown on your Job details page are equivalent to training units with the duration of the job factored in. When using Consumed ML units in your calculations, use the following formula:

Consumed ML units * $0.49

Example:

  • A data scientist in an Americas region runs a training job. The field Consumed ML units on their Job details page shows 55.75. The calculation is as follows:

    55.75 consumed ML units * $0.49
    

    For a total of $27.32 for the job.

To find your Job details page, go to the Jobs list and click the link for a specific job.

More about prediction costs

AI Platform Prediction serves predictions from your model by running a number of virtual machines ("nodes"). By default, AI Platform automatically scales the number of nodes running at any time. For online prediction, the number of nodes scales to meet demand. Each node can respond to multiple prediction requests. For batch prediction, the number of nodes scales to reduce the total time it takes to run a job. You can customize how prediction nodes scale.

You are charged for the time that each node runs for your model, including:

  • When the node is processing a batch prediction job.
  • When the node is processing an online prediction request.
  • When the node is in a ready state for serving online predictions.

The cost of one node running for one hour is a node hour. The table of prediction prices describes the price of a node hour, which varies across regions and between online prediction and batch prediction.

You can consume node hours in fractional increments. For example, one node running for 30 minutes costs 0.5 node hours. However, several rules govern cost calculations:

Cost calculations for legacy (MLS1) machine types and batch prediction

  • The running time of a node is measured in one-minute increments, rounded up to the nearest minute. For example, if a node runs for 20.1 minutes, calculate its cost as if it ran for 21 minutes.
  • The running time for nodes that run for less than 10 minutes is rounded up to 10 minutes. For example, if a node runs for only 3 minutes, calculate its cost as if it ran for 10 minutes.

Cost calculations for Compute Engine (N1) machine types

  • The running time of a node is billed in 30-second increments. This means that every 30 seconds, your project is billed for 30 seconds worth of whatever vCPU, RAM, and GPU resources that your node is using at that moment.

More about automatic scaling of prediction nodes

Online prediction Batch prediction
The priority of the scaling is to reduce the latency of individual requests. The service keeps your model in a ready state for a few idle minutes after servicing a request. The priority of the scaling is to reduce the total elapsed time of the job.
Scaling affects your total charges each month: the more numerous and frequent your requests, the more nodes will be used. Scaling should have little effect on the price of your job, though there is some overhead involved in bringing up a new node.

You can choose to let the service scale in response to traffic (automatic scaling) or you can specify a number of nodes to run constantly to avoid latency (manual scaling).

  • If you choose automatic scaling, the number of nodes scales automatically, and can scale down to zero for no-traffic durations.
  • If you choose manual scaling, you specify a number of nodes to keep running all the time. You are charged for all of the time that these nodes are running, starting at the time of deployment and persisting until you delete the model version.
You can affect scaling by setting a maximum number of nodes to use for a batch prediction job, and by setting the number of nodes to keep running for a model when you deploy it.

Minimum 10 minute charge

Recall that if a node runs for less than 10 minutes, you are charged as if it ran for 10 minutes. For example, suppose you use automatic scaling. During a period with no traffic, zero nodes are in use. If you receive a single online prediction request, then one node scales up to handle the request. After it handles the request, the node continues running for few minutes in a ready state. Then it stops running. Even if the node ran for less than 10 minutes, you are charged for 10 node minutes (0.17 node hour) for this node's work.

Alternatively, if a single node scales up and handles many online prediction requests within a 10 minute-period before shutting down, you are also charged for 10 node minutes.

You can use manual scaling to control exactly how many nodes run for a certain amount of time. However, if a node runs for less than 10 minutes you are still charged as if it ran for 10 minutes.

Learn more about node allocation and scaling.

Example of a prediction calculation

A real-estate company in an Americas region runs a weekly prediction of housing values in areas it serves. In one month, it run predictions for four weeks in batches of 3920, 4277, 3849, and 3961. Jobs are limited to one node and each instance takes an average of 0.72 seconds of processing.

First calculate the length of time that each job ran:

3920 instances * (0.72 seconds / 1 instance) * (1 minute / 60 seconds) = 47.04 minutes
4277 instances * (0.72 seconds / 1 instance) * (1 minute / 60 seconds) = 51.324 minutes
3849 instances * (0.72 seconds / 1 instance) * (1 minute / 60 seconds) = 46.188 minutes
3961 instances * (0.72 seconds / 1 instance) * (1 minute / 60 seconds) = 47.532 minutes

Each job ran for more than ten minutes, so it is charged for each minute of processing:

($0.0791 / 1 node hour) * (1 hour / 60 minutes) * 48 minutes * 1 node = $0.06328
($0.0791 / 1 node hour) * (1 hour / 60 minutes) * 52 minutes * 1 node = $0.06855
($0.0791 / 1 node hour) * (1 hour / 60 minutes) * 47 minutes * 1 node = $0.06196
($0.0791 / 1 node hour) * (1 hour / 60 minutes) * 48 minutes * 1 node = $0.06328

The total charge for the month is $0.26.

This example assumed jobs ran on a single node and took a consistent amount of time per input instance. In real usage, make sure to account for multiple nodes and use the actual amount of time each node spends running for your calculations.

Note about AI Platform Prediction charges for AI Explanations

AI Explanations comes at no extra charge to AI Platform Prediction prices. However, explanations take longer to process than normal predictions, so heavy usage of AI Explanations along with auto-scaling could result in more nodes being started, which would increase AI Platform Prediction charges.

Required use of Cloud Storage

In addition to the costs described in this document, you are required to store data and program files in Cloud Storage buckets during the AI Platform lifecycle. This storage is subject to the Cloud Storage pricing policy.

Required use of Cloud Storage includes:

  • Staging your training application package.

  • Storing your training input data.

  • Staging your model files when you are ready to deploy a model version.

  • Storing your input data for batch prediction.

  • Storing the output of your batch prediction jobs. AI Platform does not require long-term storage of these items. You can remove the files as soon as the operation is complete.

  • Storing the output of your training jobs. AI Platform does not require long-term storage of these items. You can remove the files as soon as the operation is complete.

Free operations for managing your resources

The resource management operations provided by AI Platform are available free of charge. The AI Platform quota policy does limit some of these operations.

Resource Free operations
models create, get, list, delete
versions create, get, list, delete, setDefault
jobs get, list, cancel
operations get, list, cancel, delete

What's next

Was this page helpful? Let us know how we did:

Send feedback about...

AI Platform
Need help? Visit our support page.