Startup Profile: CoreWeave's NVIDIA-Powered Cloud Service

Word has it that NVIDIA (Nasdaq: NVDA) has favored a handful of startups with sales of its coveted H100 graphics processing units (GPUs), the chip giant’s top-performing components for generative artificial intelligence (AI) workloads. And leading the small pack is CoreWeave, an alternative to the big cloud hyperscalers when it comes to offering compute capabilities for AI.

As noted by The Information earlier this year, NVIDIA is targeting CoreWeave and similar companies for early sales of the H100, which is in short supply due to big demand. The motivation behind the special treatment is that CoreWeave, unlike AWS and Google Cloud, doesn’t intend to make its own processors to lessen its reliance on NVIDIA. Instead, CoreWeave, with NVIDIA’s help, offers compute services that directly compete with the hyperscalers'.

Indeed, in its press releases, CoreWeave claims to “build cloud solutions for compute-intensive use cases — VFX [visual effects] and rendering, machine learning and AI, batch processing and pixel streaming — that are up to 35 times faster and 80% less expensive than the large, generalized public clouds.”

Why SmartNICs are Key in the Broadcom-VMware Deal

Broadcom sees compute offload and networking via merchant silicon as it key growth areas. VMware's support of multiple DPUs just might fit that plan

Can AI Save Everyone's Bacon?

Both hyperscalers and telcos are saying there's a universal delay in closing orders. But AI could make a difference -- at least for the hyperscalers

Cloud Native Climbs at KubeCon

CNCF takes on AI -- and much more

UPDATE: CoreWeave's close relationship with NVIDIA was further revealed on August 3, when news broke that the startup had secured $2.3 billion in a debt facility for which NVIDIA H100 chips were used as collateral.

Notably, Microsoft appears to be alone among the top three cloud titans in choosing not to create its own GPUs but to rely on NVIDIA’s H100. (This despite the vendor’s purchase of Fungible, a data processing unit [DPU] innovator, earlier this year.) To ensure an adequate supply of compute power, Microsoft allegedly plans to rely on CoreWeave for added H100 access in a deal that could cost billions and run into multiple years, according to CNBC. Microsoft hasn’t publicly acknowledged the deal.

Graduating from Crypto to AI

CoreWeave was founded in 2017 in Roseland, New Jersey, by Michael Intrator (ex-Hudson Ridge, a natural gas hedge fund), who is now CEO; Brian Venturo (ex-Hudson Ridge), now CTO; and Brannin McBee (ex-trader for Active Power Investments, which specializes in natural gas, agriculture, and power), now CSO. The company started as a cryptocurrency mining firm, but pivoted to AI computing around 2019 in response to market demand.

In the shift-over, CoreWeave built on its relationship with NVIDIA. It now enjoys the status of offering not just H100 services but also ones based on many other NVIDIA chips, including the A100, the RTX, and A40 GPUs. (In 2021, CoreWeave claimed to offer North America’s largest installation of A40 GPUs.) CoreWeave also deploys NVIDIA’s InfiniBand computing platform and BlueField DPUs in its cloud.

CoreWeave has created a comprehensive cloud-native, Kubernetes-powered user interface for its cloud services, featuring applications, APIs, object storage, namespaces, Grafana monitoring, and other features.

NVIDIA Backs CoreWeave

CoreWeave has raised over $500 million ($576.5 million, according to Crunchbase). In May 2023, the company announced a Series B tranche of $421 million, including an extension of $200 million on initial funding of $221 million. The round was led by Magnetar Capital, with a contribution from NVIDIA and a “rounding out” amount from individual investors Nat Friedman and Daniel Gross.

In its initial funding announcement, CoreWeave stated that the new money would be used for expansion purposes, including building additional datacenters to provide its services. Presently, CoreWeave has datacenters in Weehawken, N.J.; Chicago, Ill.; and Las Vegas, Nev. The vendor says the datacenters are linked to one another by dark fiber supporting rates of 400 Gb/s, and each supports redundant, 200-Gb/s onramps to Tier 1 ISPs.

The company plans to launch two new datacenters in 2023, bringing its total to five. It recently announced plans to set up a facility in Plano, Texas.

CoreWeave isn’t the only small cloud alternative compute provider favored by NVIDIA. Lambda Labs also sports H100 compute power for rent. And Crusoe Energy has been named as another recipient of the NVIDIA components. Clearly, the rental of powerful compute capabilities for generative AI is a fast-growth segment that could account for many billions of dollars in revenue next year.