Google Cloud TPUs land on Hugging Face

If you wanted to host an open-source AI model in the cloud, your default move was almost always renting an expensive Nvidia GPU. But Hugging Face is now giving builders another path.

Google Cloud TPU v5e chips are officially available directly inside Hugging Face Inference Endpoints and Spaces, letting developers run open models on Google's custom AI silicon in just a few clicks.

Three tiers starting at $1.38

The rollout brings Google Cloud TPU v5e instances to the platform. By choosing Google Cloud Platform and selecting the us-west1 region, developers can deploy models across three initial hardware configurations.

The entry-level v5litepod-1 features 1 core and 16 GB of memory for $1.375 per hour, designed to run models with up to 2 billion parameters without memory hiccups.

For larger models, the v5litepod-4 offers 4 cores and 64 GB of memory at $5.50 per hour. The largest tier, v5litepod-8, delivers 8 cores and 128 GB of memory for $11.00 per hour, which Hugging Face notes provides the lowest inference latency.

The same hardware tiers are also accessible within Hugging Face Spaces. Creators can head into the settings menu of their interactive demos and switch their backend to TPUs right away.

Bridging the software gap

Moving models from GPUs to TPUs used to mean grappling with complex software stacks and compilation headaches. To eliminate that friction, the two companies released an open-source library called Optimum TPU.

Inference Endpoints pairs Optimum TPU with Hugging Face's Text Generation Inference system to handle large language model serving automatically.

Right out of the gate, users can deploy Gemma, Llama, and Mistral model architectures with a few clicks, without writing custom low-level integration code.

Why it matters

We think this is one of the smartest ecosystem plays Google has made in recent memory. For years, open-source AI development has been virtually married to Nvidia CUDA. By dropping TPUs into Hugging Face as a simple drop-down option, Google is finally giving the broader developer community a frictionless reason to try its custom chips.

For builders, it represents a genuine alternative when GPUs are either too expensive or simply out of stock. Running an interactive demo or a 2B parameter model for under $1.40 an hour makes prototyping accessible without blowing through cloud credits.

The trade-off, at least for now, is architectural flexibility. While mainstream models like Llama and Gemma are ready to go, exotic architectures will still face hurdles compared to the well-worn GPU path. Still, breaking Nvidia's functional monopoly on turn-key AI hosting is a huge win for developers looking to trim their cloud bills.