Does an infinite supply of GPUs actually solve the AI bottleneck?

The technical specifications center on the G4 VM family and the broader AI Hypercomputer architecture. The G4 VMs utilize the NVIDIA RTX 6000 Blackwell Server Edition GPU, featuring Google’s custom peer-to-peer communication protocols. Google also introduced fractional G4 VMs, leveraging NVIDIA virtual GPU (vGPU) technology to partition a single physical Blackwell GPU. This allows developers to allocate specific VRAM and compute for tasks like real-time 3D rendering.

On the software side, the integration of GKE Inference Gateway with Dynamo creates a modular control plane. For extreme workloads, Google detailed the A4X Max, powered by the NVIDIA GB300 NVL72, featuring third-generation liquid cooling and the Jupiter network fabric to eliminate "tail latency."

Fractional GPU-aaS

Google Cloud's preview of fractional G4 VMs introduces a more efficient entry point for AI and graphics workloads by utilizing NVIDIA virtual GPU technology. These VMs partition a single NVIDIA RTX PRO 6000 Blackwell Server Edition GPU into flexible slices, allowing customers to choose from 1/2, 1/4, or 1/8 GPU increments based on their specific needs. This granular sizing supports a wide range of tasks, from intensive LLM inference and robotics simulations to lightweight remote desktops and entry-level streaming. By providing smaller configurations, Google Cloud enables users to pay only for the fractional resources they consume, significantly reducing operational overhead. The infrastructure is deeply integrated with Google Kubernetes Engine, which uses advanced container binpacking to maximize hardware utilization and price-performance. Furthermore, the Dynamic Workload Scheduler automates the process of finding available GPU slices, ensuring higher obtainability for diverse workloads. As part of a co-engineered stack, these VMs work seamlessly with NVIDIA NeMo on Vertex AI and NVIDIA Dynamo to support complex reasoning and Mixture-of-Experts models. Ultimately, this offering provides an open and high-performance platform that helps enterprises right-size their infrastructure and maximize their ROI in the evolving AI landscape.

Competitive Landscape for Fractional GPUs

We recently wrote about CoreWeave’s announcement around fractional GPU-aaS, and while both Google Cloud and CoreWeave leverage the NVIDIA RTX PRO 6000 Blackwell Server Edition, their strategies diverge significantly in how they deliver this high-performance hardware to users. Google Cloud’s new fractional G4 VMs prioritize extreme granularity, offering slices as small as 1/8 of a GPU to provide a highly accessible entry point for lightweight AI and graphics tasks. In contrast, CoreWeave maintains a performance-first approach, typically favoring full-GPU or 8-GPU node configurations optimized for large-scale "AI Factory" workloads.

Google integrates these fractional resources deeply into its managed ecosystem, utilizing GKE and Vertex AI to automate "container binpacking" and resource allocation. CoreWeave differentiates itself through a specialized, Kubernetes-native stack that emphasizes "goodput" and reliability for massive clusters rather than fine-grained slicing. While Google’s pay-per-slice model is designed to maximize ROI for enterprises with diverse, agentic AI needs, CoreWeave’s flexible capacity plans are tailored for labs requiring elite-tier performance without thermal throttling. This creates a clear market split: Google serves the "right-sized" general-purpose cloud market, whereas CoreWeave remains the choice for specialized, high-density infrastructure. Ultimately, Google’s strength lies in its "AI Hypercomputer" platform’s ability to blend TPUs and GPUs, while CoreWeave’s advantage is its mature, liquid-cooled Blackwell deployment that has been in general availability since mid-2025. Together, these two providers offer a spectrum of choice ranging from precise, fractional efficiency to raw, uncompromised power. Let the competitive games begin!

Tension Remains

Google Cloud’s strategy involves a delicate balancing act, aggressively co-engineering new infrastructure with NVIDIA while simultaneously expanding its own custom-built Tensor Processing Units (TPUs) to capture more of the AI chip market. The announcement blog post emphasizes "giving customers even more options," a phrasing that underscores the competitive reality where Google's TPUs and NVIDIA's Blackwell GPUs vie for the same high-performance AI workloads.

While Google serves as an "Elite sponsor" for NVIDIA’s latest hardware, like the Vera Rubin systems, it is also positioning its own Axion and TPU silicon as high-efficiency alternatives to reduce its long-term reliance on third-party vendors. This "open ecosystem" approach allows Google to benefit from NVIDIA's industry-standard CUDA software stack while quietly developing the AI Hypercomputer as a platform that can eventually swap out GPUs for in-house silicon. Ultimately, the tension lies in Google’s dual role: it is one of NVIDIA’s most important distribution partners in the cloud, yet it remains a formidable rival in the race to define the next generation of AI-optimized hardware.