In short: After August 7, 2026, a defined set of NVIDIA CUDA products stops receiving new updates on NVIDIA’s standalone nvidia channel. New releases continue on Anaconda’s main (defaults) channel and on conda-forge. The nvidia channel stays online, existing packages remain available, and your current environments keep working. If you already get these products from main or conda-forge, nothing changes. If you get them from the nvidia channel, switching is a one-line config change for most setups.

What’s changing

If you install CUDA or NVIDIA’s GPU libraries through conda, you’re getting them from one of three channels: nvidia, main, or conda-forge. These products stop receiving new updates on the nvidia channel after August 7, 2026:

cuda, cublasmp, cudnn, cudss, cuest, cufftmp, cuquantum, cusolvermp, cusparselt, cutensor, nccl, nvpl, nvshmem, cuda-python, numbast

This is not a full channel retirement. The nvidia channel stays online, every existing package remains downloadable, and current environments keep working. Products outside this list are unaffected. The only change is where new updates for these products are published: going forward, on main and conda-forge.

Why this is happening

The nvidia channel was created when it was the only way to get CUDA into conda. Since then, both main (through Anaconda’s CUDA redistribution partnership with NVIDIA) and conda-forge (through NVIDIA’s work with the community) have matured into complete distribution paths for CUDA, with new versions reaching main within weeks of an upstream release.

Maintaining a third update path for the same products meant doing the same work three times over. Concentrating updates on the two channels most conda users already rely on gives a clear, supported path for CUDA and removes the complexity of managing a third channel.

 

Which channel to migrate to

Both main and conda-forge are healthy homes for CUDA, and either keeps your updates coming. For most people migrating off the nvidia channel, and especially anyone running production, regulated, or security-sensitive workloads, main is the recommended path. It’s the curated, supported home for these packages, and if you use Anaconda Distribution or Miniconda it’s already in your configuration, so migrating is simply removing the nvidia channel. If you prefer a community-governed solution without a commercial vendor dependency, conda-forge is also a supported path. Choose conda-forge if you are comfortable supporting your own environments; you’ll need to add it explicitly.

The core trade-off is assurance and support versus raw speed:

 mainconda-forge
Maintained byAnaconda, end to endOpen-source community
Best forProduction and team environments; anywhere you need commercial support, environment stability, security curation, or regulated/sensitive workloadsDevelopers comfortable supporting their own environments; teams that want a fully community-governed path
SecurityCurated and rebuilt in access-controlled infrastructure; vulnerability remediation; signed metadata; a curated namespace, so not exposed to typosquatting or dependency-confusionCommunity-built; maintainers note it isn’t recommended for environments with sensitive information, since they can’t verify every package is free of tampering
Release speedNew CUDA within weeks of upstreamFast; new versions appear here first from NVIDIA
SupportFormal commercial supportCommunity

Before you switch: all affected products will be available on main and conda-forge by August 7, 2026. If you migrate before then, confirm the package and version you need is present on your target channel first. If you’re running an older release, the nvidia channel continues to serve existing packages; migrate when you want a newer version.

The guide below covers migrating to main. For conda-forge, the steps are nearly identical: substitute conda-forge for main throughout, and add the channel explicitly.

How to migrate to main

For most people, migration means one thing: keep installing the same packages, but get new updates from main instead of the nvidia channel.

# Remove the nvidia channel
conda config --remove channels nvidia

# Install/update as usual — no -c nvidia flag
conda install <package>

If you use Anaconda Distribution or Miniconda, main is likely already in your configuration (it’s part of the defaults channel set), so removing nvidia is likely all you need. The steps below confirm that and clean up environment files, scripts, and containers that still name the nvidia channel.

Step 1: Confirm your current channel configuration

See which channels conda is configured to use:

conda config --show channels

If you see nvidia listed, that’s the one to migrate. If you don’t, your config is already done.

To see which installed packages came from the nvidia channel, check the rightmost Channel column:

conda list

Step 2: Point conda at main

Find every .condarc affecting your setup:

conda config --show-sources

Remove the nvidia channel:

conda config --remove channels nvidia

Verify main is in your channel list:

conda config --show channels

You should see either main or defaults (the defaults set includes main). If neither is present, add it:

conda config --add channels defaults

Step 3: Update environment files, scripts, and containers

Anywhere you named the nvidia channel, remove that reference. Package names stay the same.

environment.yml — drop nvidia from the channels: list:

# Before
channels:
  - nvidia
  - defaults

# After
channels:
  - defaults

In install commands, CI/CD, and Dockerfiles, drop the -c nvidia flag:

# Before
conda install -c nvidia <package>

# After
conda install <package>

Step 4: Update your environments and verify

Update the affected packages so they resolve from main:

conda update <package>

Or, if you manage environments from a file, recreating from the updated environment.yml is the cleanest way to clear mixed-channel state:

conda env remove -n <env-name>
conda env create -f environment.yml

Then confirm nothing still points at nvidia:

# Should return nothing
conda config --show channels | grep nvidia

# The Channel column should show main, not nvidia
conda list <package>

Getting help

Anaconda enterprise customers: reach out to your account team or support with questions, including help moving to main.