NVIDIA, Google Cloud launch L4 generative AI platform

NVIDIA announced a co‑engineered generative AI platform with Google Cloud on June 25, 2026, saying Google Cloud will offer NVIDIA’s L4 Tensor Core GPU in private preview alongside new integration with Vertex AI.

The company described the work as an inference platform tuned for generative workloads, and said L4 capacity would be available in Google Cloud’s Compute Engine on a private‑preview basis.

Google Cloud’s materials confirm the vendor collaboration and highlight G2 virtual machines — branded VMs that pair L4 GPUs with Vertex AI tooling — as entering private preview as part of the announcement.

NVIDIA and Google Cloud said Vertex AI will receive optimized support so developers can build, tune and deploy modern generative models with lower overhead and faster time to production.

The launch is strikingly image‑friendly: NVIDIA supplied new product boards, VM‑branded art and data‑center photos for the joint release, underscoring the marketing push behind the technical announcement.

NVIDIA described the L4 as a “universal” GPU aimed at inference and media workloads, saying it offers large gains in AI video performance and energy efficiency compared with CPU‑only configurations. The company positioned L4 as a cost‑efficient option for production inference.

The tie‑up comes as hyperscalers broaden the silicon mix they offer to customers — adding bespoke ASICs, new rack‑scale systems and multiple GPU families — while still leaning on NVIDIA’s software and reference stacks. Google Cloud’s recent GTC notes and NVIDIA’s own roadmaps show both sides continuing to co‑engineer across hardware and software.

At the customer level, companies such as Descript and WOMBO were cited as early L4 users in the joint materials, examples intended to show how generative image and audio services can migrate to GPU‑accelerated inference in the cloud.

On the technical side, Google and NVIDIA reiterated features that make L4 attractive for inference: multi‑precision support including FP8 and INT8, efficient video encode/decode, and throughput gains for mid‑sized generative models. Google’s documentation and NVIDIA notes provide the performance context for workloads from text generation to image rendering.

The announcement also sits alongside broader Google Cloud product moves at GTC 2026 — including fractional GPU options and early roadmaps for next‑generation rack platforms — signaling that cloud providers want both finer‑grained cost options and access to flagship accelerators.

For NVIDIA the deal continues a pattern: the company’s inference platform and developer tooling keep appearing across hyperscaler stacks, even as some cloud operators explore or promote alternate ASICs and custom accelerators in parallel. That mix is shifting how large customers architect AI services, with multi‑vendor stacks becoming the norm.

Practically, L4 on Google Cloud is initially gated behind private preview access and sign‑up; enterprise customers and startups that want early access can apply through Google's usual preview channels. The announcement is both a technical update and a branding exercise — new VM names, images and photos make the partnership visible in marketing as well as in data centers.