CoreWeave details GTC 2025 launches: CAIOS, RTX PRO 6000 and NVCF support

Product announcements at GTC

According to CoreWeave, the headline release is CoreWeave AI Object Storage (CAIOS), now generally available. The company describes it as a next-generation object storage solution optimized for high I/O and large datasets in AI and ML workflows, offering performance at scale, integration with existing workflows and enterprise-grade reliability.

CoreWeave also said it will introduce new compute instances featuring the upcoming NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. Compared to the NVIDIA L40S, the company says the RTX PRO 6000 can deliver up to 5.3x faster LLM inference and 3.5x faster text-to-image generation.

On the software side, CoreWeave announced support for the NVIDIA AI Enterprise software platform and NVIDIA Cloud Functions (NVCF), a core technology of DGX Cloud Serverless Inference, on the CoreWeave Cloud Platform. The company says it is one of the first cloud providers to offer NVCF on NVIDIA GB200 NVL72.

CoreWeave also released findings on Model FLOPS Utilization (MFU). Its latest study found an MFU of over 50% on NVIDIA H100 GPUs, which the company says drives up to 1.2X higher performance than alternative solutions.

Main stage talks and partner sessions

CoreWeave executives spoke on the NVIDIA main stage about AI infrastructure, efficiency and sustainability in data center design, and performance in GPU clusters. CTO Peter Salanki covered networking at massive scale, including how CoreWeave implemented one of the largest-scale NVIDIA Quantum-2 InfiniBand networks available.

A second session, with Salanki, John Mancuso and Jane Street's Adam Canady, described how Jane Street uses CoreWeave infrastructure to train and fine-tune quantitative trading AI models. A third, led by SVP of Engineering Chen Goldberg and CPO Chetan Kapoor, addressed model loading, checkpointing and fleet management for fault handling. The team also shared details on the NVIDIA GB200 NVL72 clusters CoreWeave rolled out last fall.

At its booth, CoreWeave hosted talks with several partners:

• IBM: infrastructure for training and fine-tuning models such as IBM's Granite 3.2, plus CoreWeave's new IBM storage offering.
• Cohere: a first look at the GB200 NVL72 in practice. According to the recap, early benchmarks showed dramatic throughput gains over NVIDIA HGX H100s, with more to come.
• Rescale: a fireside chat on the generative AI boom, the economics of GPU access and cloud HPC.

On-demand content and what it signals

CoreWeave also supported two on-demand sessions. One, by solutions architects Navarre Pratt and Jacob Feldman, looked at how MLOps teams can improve resilience through visibility and automation, citing NVIDIA BlueField-3 DPUs and the self-healing properties of NVIDIA Quantum-2 InfiniBand. The other, by Principal Product Manager Jeff Braunstein, shared benchmarking results for CoreWeave Distributed File Storage in a 32K NVIDIA H200 GPU cluster.

The company also hosted social events, including a networking night with NVIDIA and Dell in San Jose and a Solidigm podcast recording at its booth.

For customers, the announcements point to a broader CoreWeave offering across storage, compute and serverless inference tooling. CoreWeave frames the conference message as one of accelerating AI computing demand, driven by reasoning AI and agentic AI. The performance figures cited are the company's own claims, and the RTX PRO 6000 instances are described as coming rather than available.