CoreWeave Validates Nvidia Vera Rubin as SoftBank Bets $85B on France AI Infrastructure

CoreWeave became the first AI cloud provider to bring up and validate Nvidia's Vera Rubin NVL72, the company announced June 1. The milestone includes rigorous system-level validation for the entire rack-scale architecture, which features 72 Rubin GPUs and 36 Vera CPUs per rack connected via a 260 TB/s NVLink 6th-generation fabric.

Separately, SoftBank revealed plans for up to 5 GW of AI infrastructure in France, an $85 billion bet that leverages French utility EDF's conversion of former power plant sites into data center campuses. The move underscores how electricity availability has become a competitive advantage in the AI race.

At its GTC Taipei conference, Nvidia confirmed that Vera Rubin and the new Vera CPU remain on track. The company also launched DSX OS, an operating system designed to run AI factories. Early adopters include Anthropic, OpenAI, and SpaceXAI.

CoreWeave developed several purpose-built innovations to support Vera Rubin at production scale. Valvey, a programmable per-rack valve assembly, turns liquid cooling into a software-defined control surface. Racky aggregates power, cooling, and environmental sensors into a unified management appliance. The company also supports multi-rail, multi-plane networking with NVIDIA Quantum-X800 InfiniBand and Spectrum-X Ethernet.

Vera Rubin NVL72 delivers up to 10 times better inference per watt compared to NVIDIA Blackwell, according to Nvidia. It requires up to one-fourth fewer GPUs and costs one-tenth per million tokens. These gains matter as agentic AI models reach trillion parameters and context windows extend to millions of tokens.

DigitalOcean addressed the hidden cost of inference at scale. The company introduced prefix-aware routing that eliminates redundant recomputation of shared prompt prefixes, reducing compute spend. Its serverless inference platform handles GPU contention, traffic unpredictability, and multi-model orchestration across text, vision, and audio. By 2030, inference is expected to account for the majority of AI compute globally, making such optimizations critical.

The parallel developments show an industry racing to balance performance, cost, and power. CoreWeave's validation, SoftBank's infrastructure bet, Nvidia's new chip and OS, and DigitalOcean's inference optimizations all point to a maturing ecosystem where inference efficiency determines which companies can scale.