NVIDIA T4 GPUs on Google Cloud will reach end of support (EOS) on August 1, 2027.
What you need to know
As of August 1, 2027, you cannot create, launch, or access any Google Cloud resources that run NVIDIA T4 GPUs. This deprecation affects resources for the following services:
- Compute Engine: Compute Engine instances
- Google Kubernetes Engine (GKE): nodes
- Gemini Enterprise Agent Platform: models, jobs, and endpoints
- Cloud Workstations: workstations
- Dataflow: pipeline jobs
- Managed Service for Apache Spark: clusters and serverless batches
- Deep Learning VM and Container-Optimized OS VM instances
Committed use discounts (CUDs)
Google Cloud no longer offers new 3-year committed use discounts (CUDs) for NVIDIA T4 instances. Your existing active 1-year and 3-year commitments remain valid until their scheduled expiration date, but you cannot purchase or renew 3-year commitments for T4 GPUs.
How this EOS impacts your existing resources
Until August 1, 2027, your resources that run NVIDIA T4 GPUs continue to function normally. However, because NVIDIA T4 GPUs are approaching their EOS, you must migrate your workloads to modern GPU alternatives before the decommission date. Google Cloud does not support EOS products or features.
As of the final decommission date on August 1, 2027, Google Cloud will shut down any remaining compute instances and resources running NVIDIA T4 GPUs, and you can no longer create, launch, or access them.
What you need to do
To transition your workloads to supported GPUs, choose a new GPU model, review its availability in your regions, and migrate your workloads before August 1, 2027.
Choose a GPU model
We recommend migrating your workloads to modern GPU alternatives, specifically the G4 or G2 machine series. To determine the best GPU model for your workload, compare the following options:
| Feature | G2 (NVIDIA L4) | G4 (NVIDIA RTX PRO 6000) |
|---|---|---|
| Best for | High-performance AI inference, generative AI, video processing and streaming, graphics, and light machine learning (ML) training. L4 GPUs provide up to four times the inference performance of T4 GPUs. | High-performance simulation, NVIDIA Omniverse workloads, 3D graphics rendering, local LLM fine-tuning and inference, virtual workstations (vWS), and fractional GPU (vGPU) sharing. |
| Architecture and memory | Ada Lovelace architecture with 24 GB GDDR6 GPU memory. | Blackwell architecture with 96 GB GPU memory per GPU. |
| Machine series | G2 accelerator-optimized | G4 accelerator-optimized |
For a comprehensive list and comparison of all GPU models, see GPU machine types.
Review available locations
To verify that your selected GPU model is available, review the location details for your service:
- Gemini Enterprise Agent Platform: review Gemini Enterprise Agent Platform accelerator locations.
- Cloud Workstations: review Cloud Workstations GPU availability by region and zone.
- For all other services: review the available GPU regions and zones.
Migrate your workloads
Based on the service that you use, select one of the following transition paths:
Compute Engine workloads
To transition your Compute Engine workloads to the recommended G2 (NVIDIA L4) or G4 (NVIDIA RTX PRO 6000) machine series, you must move to a new compute instance. You cannot modify an existing compute instance in-place from an N1 general-purpose machine type to an accelerator-optimized machine type.
To move to a new compute instance, complete the following steps:
- If your existing compute instance uses Local SSD disks that contain data you want to keep, move the contents of those disks to a Persistent Disk volume.
Create a new compute instance. For instructions, see Create a G2 or G4 instance.
Move Persistent Disk volumes by detaching them from the old compute instance and attaching them to the new compute instance. You can alternatively transfer files from one compute instance to the other.
Optional: Move the saved data from Persistent Disk volumes back to a Local SSD disk.
Reassign any static IP addresses associated with the old compute instance to the new compute instance.
Install GPU drivers on the new compute instance.
Install your applications on the new compute instance.
Delete the old compute instance.
Other Google Cloud workloads
If your workloads run on services other than Compute Engine such as GKE, Gemini Enterprise Agent Platform, Cloud Workstations, Dataflow, or Managed Service for Apache Spark, then do the following:
Update your configurations to reference a supported GPU model.
- For GKE, Gemini Enterprise Agent Platform, or Cloud Workstations, update your configuration templates.
- For Dataflow, update your pipeline specifications.
- For Managed Service for Apache Spark, update your cluster definitions.
For instructions about configuring GPUs for your service, see the product documentation for that service.
Restart or recreate your resources.
What's next
- Learn more about GPUs.
- Learn about GPU machines in the accelerator-optimized machine family.
- Review GPU pricing.