Google Cloud is taking another big swing at making AI workloads easier to run, and this time it’s with serious hardware muscle. Cloud Run, Google’s serverless compute platform, now supports NVIDIA’s RTX Pro 6000 Blackwell GPUs—a move that signals how quickly high-end inference is becoming mainstream for developers and enterprises alike.

Traditionally, running massive models meant wrangling clusters, reserving GPUs, and babysitting virtual machines. It was messy, expensive, and time-consuming. Cloud Run’s promise has always been “just code, no infrastructure headaches,” and now that extends to workloads that demand cutting-edge GPUs. With the RTX Pro 6000, developers can deploy models like Llama 3.1 70B or Gemma 3 27B without worrying about provisioning or scaling. The platform spins up GPU-backed instances in under five seconds, and when traffic dies down, it scales back to zero—so you’re not paying for idle hardware.

The RTX Pro 6000 Blackwell isn’t just a minor upgrade over previous GPUs like the NVIDIA L4. It brings 96GB of vGPU memory, 1.6 TB/s of bandwidth, and support for FP4 and FP6 precision. That’s a big deal for generative AI, where efficiency and speed matter. Think text-to-image generation, multimodal applications, or fine-tuning large language models—all of these benefit from the GPU’s fifth-generation Tensor Cores and high-efficiency compute. For businesses, this means real-time AI applications can be built without the traditional infrastructure drag.

Google is also leaning into flexibility. Cloud Run lets you attach these GPUs to services, jobs, or worker pools, depending on whether you’re doing inference, fine-tuning, or specialized workloads. You can even configure up to 44 vCPUs and 176GB of RAM alongside the GPU, giving developers plenty of headroom for complex tasks. And because it’s serverless, you don’t need reservations—capacity is managed automatically, with zonal redundancy available for production-grade reliability.

The integration with Google Cloud’s broader ecosystem is another angle worth noting. You can mount Cloud Storage buckets directly to load massive model weights, or secure traffic with Identity-Aware Proxy. It’s a reminder that Google isn’t just offering raw compute; it’s trying to make the entire AI pipeline—from storage to deployment—seamless.

Availability is still limited. The RTX Pro 6000 GPUs are in preview, with regions like us-central1 and europe-west4 getting first dibs, and partial rollout in Asia. But the direction is clear: Google wants Cloud Run to be the go-to for developers who want to experiment with large models without the operational baggage.

What’s striking here is how serverless is evolving. It started as a way to run lightweight apps without infrastructure overhead. Now, it’s being stretched to handle some of the heaviest workloads in AI. For developers, that’s liberating. For enterprises, it’s a chance to scale innovation without scaling costs. And for Google, it’s a way to position Cloud Run as not just a convenience tool, but a serious contender in the AI infrastructure race.

This move also reflects a broader trend: GPUs are no longer just for specialized ML teams. They’re becoming accessible to anyone who wants to build AI-powered applications, whether it’s a startup experimenting with generative art or a Fortune 500 company deploying multimodal assistants. By abstracting away the infrastructure, Google is betting that more developers will take the leap into large-scale AI—and that Cloud Run will be the platform they choose to do it.