Google Cloud launches A3 instances with Nvidia H100 GPUs

You might expect Google to bet everything on its in-house TPUs, but the ongoing generative AI surge has a way of bending even the strictest hardware loyalties.

At Google Cloud Next, NVIDIA CEO Jensen Huang walked onto the keynote stage alongside Google Cloud CEO Thomas Kurian to confirm that Google Cloud's A3 instances, powered by NVIDIA H100 Tensor Core GPUs, are now generally available.

NVIDIA was also named Google Cloud's Generative AI Partner of the Year, marking a deep alignment across silicon and software.

DeepMind embraces the competition's chips

Perhaps the most telling revelation from the keynote is that Google is not just renting these chips out to startups. Google DeepMind and other internal research teams are actively using NVIDIA H100 and A100 GPUs for frontier model training and inference.

To make that easier, the companies brought NVIDIA GPU acceleration to PaxML, Google's JAX-based framework for massive language models. PaxML is available immediately via the NVIDIA NGC container registry.

Beyond raw instances, NVIDIA AI Enterprise has landed on the Google Cloud Marketplace, with acceleration libraries directly embedded into the Vertex AI development suite. Google also pointed out that its Dataproc data processing service now leverages the RAPIDS Accelerator for Apache Spark.

More than a thousand generative AI startups are already tapping this combined stack. Marketing platform Writer used the NVIDIA NeMo framework on Google Cloud to scale its models from hundreds of millions of parameters up to 40 billion, serving enterprise clients like Uber and Deloitte. Video generation platform Runway is also leaning on the infrastructure to synthesize original footage.

Why it matters

This partnership shows that in the current generative AI landgrab, no cloud hyperscaler can afford to be precious about its proprietary silicon.

For developers and enterprise builders, general availability of A3 instances means you no longer have to leave the Google ecosystem just to tap premier H100 clusters. Being able to access NVIDIA's software stack alongside Vertex AI lowers the friction for training multi-billion-parameter models.

We think Google played this smartly by keeping PaxML open to NVIDIA hardware. Trying to force enterprise teams onto TPUs when the entire industry builds on CUDA would simply hand workloads over to Microsoft Azure or AWS. Embracing Jensen Huang's chips is the price of keeping AI startups inside Google Cloud.

The clear risk, however, is margin pressure and dependency. Building your flagship AI offerings around NVIDIA silicon leaves Google at the mercy of GPU allocation bottlenecks and pricing, even as it pours billions into developing its own chips.