Hugging Face and Google Cloud team up on TPUs and custom CDNs

Anyone who has deployed open-source AI models in the cloud knows the pain. You write three lines of code, spin up an instance, and then sit around twiddling your thumbs while a 50GB model downloads over the public web.

Now, Hugging Face and Google Cloud are teaming up to fix that bottleneck.

The two companies just announced a deeper strategic partnership aimed at bridging open-source models with cloud infrastructure. Over the past three years, Google Cloud customers have increased their Hugging Face usage by 10x, pulling tens of petabytes in model weights across billions of requests each month.

Dedicated CDN caching and native TPU support

To handle that deluge of traffic, the centerpiece of this deal is a new dedicated CDN Gateway. Built using Hugging Face's Xet storage transfer tech alongside Google Cloud's networking backbone, it caches models and datasets directly inside Google Cloud.

Whether you are spinning up models via Vertex AI Model Garden, GKE, Cloud Run, or plain Compute Engine VMs, the transfer times and time-to-first-token are slated to drop dramatically.

On the compute side, Hugging Face's managed Inference Endpoints will tap into Google Cloud instances, bringing both newer hardware options and promised price cuts.

Just as notably, Hugging Face is adding native library support for Google's custom AI silicon, now in its seventh-generation TPUs. The goal is to make compiling and running open-source models on TPUs as turnkey as it currently is on standard GPUs.

Google is also throwing its security weight behind the Hub. Technologies from VirusTotal, Google Threat Intelligence, and Mandiant will actively scan the millions of models, datasets, and Spaces hosted across Hugging Face to catch supply chain vulnerabilities.

Why it matters

For builders, this is a massive quality-of-life upgrade. Pulling massive weights across the open internet has always been a hidden tax on cloud agility. Internal caching directly lowers cold-start times and makes running open-source stacks in production much less painful.

At the same time, native TPU integration is Google's smartest play yet to chip away at Nvidia's hardware monopoly. Hugging Face is the front door for 10 million AI developers; if Google can make its cheaper, custom chips work seamlessly with popular open-source repos, developers will have a compelling financial reason to switch.

We think this partnership highlights a major shift in cloud strategy. The era of trying to lock developers into proprietary model walled gardens is giving way to infrastructure races, where the cloud that runs open models the fastest and cheapest takes the crown. The main risk looks to be vendor lock-in creeping in through optimized data transport, but right now, faster downloads and cheaper inference are hard to turn down.