Incumbent Clouds GPU Allocation Tradeoff
Nscale
The real edge for AWS, Azure, and Google Cloud is that they can bundle GPUs into an existing enterprise stack, but that same breadth makes scarce Blackwell supply harder to aim at any one workload. A model team already buying storage, security, databases, and support from a hyperscaler can add GPU instances with less procurement friction. The tradeoff is that these clouds must spread limited top tier GPUs across many customers and products, while specialists can concentrate inventory on AI jobs first.
-
The incumbents are already rolling out Blackwell. Google introduced A4 VMs with NVIDIA B200 in preview in January 2025. AWS made P6e GB200 UltraServers generally available in July 2025, and later launched P6 B300 in November 2025. Azure also brought ND GB200 v6 online and built a large GB300 cluster for OpenAI workloads.
-
Integrated service offerings matter because the hyperscaler sale is not just a GPU rental. A customer can train on cloud GPUs, keep data in the same cloud, plug into identity and security tools, then deploy into managed inference and application services. That convenience is harder for an independent GPU cloud to match, even if the raw hardware offer is stronger.
-
Supply allocation is a real bottleneck, not a theoretical one. CoreWeave built much of its position on preferential NVIDIA allocation, and Microsoft committed a multi year deal for GPU capacity because Azure demand outstripped what it could serve directly. That shows how even the largest clouds can have customer demand that runs ahead of available top end GPU inventory.
Going forward, the winners among hyperscalers will be the ones that turn GPU access from a scarce component into a default feature of the broader cloud contract. As Blackwell and then Rubin capacity expands, the big clouds should get stronger with enterprises, while focused providers like Nscale, CoreWeave, Crusoe, and Lambda keep winning where concentrated inventory, speed, or power economics matter most.