CoreWeave Gets $100M From Magnetar and Plans NVIDIA H100 Access in Q1 2023
CoreWeave says Magnetar has backed the company with a $100 million investment, which will go into scaling its NVIDIA GPU-accelerated cloud infrastructure. In a year-end post, the company also said clients can reserve NVIDIA H100 capacity now, with availability at scale in Q1 2023.
- Magnetar invested $100 million in CoreWeave, money the company says will go directly into scaling infrastructure.
- NVIDIA H100 GPUs are due on CoreWeave Cloud in Q1 2023, starting at $2.23/hr, with reservations open now.
- CoreWeave cites up to 9x faster training and 30x inference speedup on LLMs for the H100.
- 2022 additions include object storage from $0.03 / GB per month, Goose AI and the CoreWeave Accelerator Program.
- A new New York City region data center (LGA1) and a headquarters in Roseland, New Jersey were also announced.
Funding aimed at infrastructure capacity
According to CoreWeave, the $100 million investment from Magnetar was announced in December 2022 and is going directly into scaling its infrastructure. The company says this will let it deliver a large capacity of NVIDIA GPU-accelerated compute to AI and machine learning companies, VFX studios, and other enterprise and start-up businesses.
CoreWeave frames the money as a response to rising demand. The company argues that as AI and machine learning evolve, demand for infrastructure rises with them, and that start-ups and enterprises will need reliable, flexible, and highly available compute resources at affordable prices. It points to recent events as evidence of the pace: ChatGPT, OpenAI's large language model, gained 1 million new users in the first week of its launch, and AI-generated self-portraits from Lensa AI, which used the Stable Diffusion model, filled social feeds.
CoreWeave also cites GPT-NeoX-20B, which it says became the largest publicly available language model earlier in the year. EleutherAI trained it on CoreWeave's NVIDIA A100 Tensor Core GPU training cluster. Since then, the company says, demand for its serverless Kubernetes infrastructure has grown among businesses and researchers building products powered by LLMs.
NVIDIA H100 coming to CoreWeave Cloud
In Q1 2023, CoreWeave says clients will have access to NVIDIA H100 Tensor Core GPUs on CoreWeave Cloud. The company says this makes it one of the first providers to offer cloud instances of NVIDIA H100-based supercomputers.
Clients can reserve H100 capacity today, according to CoreWeave. It will be available at scale in Q1 2023 and starting at $2.23/hr.
The post repeats performance claims for the chip. The H100 has fourth-generation Tensor Cores and the Transformer Engine with FP8 precision, and offers up to 9x faster training and a 30x inference speedup on large language models. For high-performance computing applications, the post says it delivers up to 7x higher performance.
Product additions, customers and expansion in 2022
CoreWeave listed several additions from the year:
- Object storage for high volumes of unstructured data, starting at $0.03 / GB per month, with no access or transfer fees.
- Integrations with Determined.AI for large-scale AI training, Zeet for managing Kubernetes infrastructure, and EleutherAI to make open-source LLMs more accessible.
- AI examples for fine-tuning and serving LLMs and generative AI models such as BLOOM and Stable Diffusion.
- Goose AI, a fully managed API to serve pre-trained, open-source models.
The company also launched the CoreWeave Accelerator Program, which aims to give startups NVIDIA GPU-accelerated compute on demand and at scale, along with visibility and partnership opportunities.
Two customers were highlighted. Spire Animation Studios presented at NVIDIA GTC 2022 on a cloud-native workflow built with Unreal Engine for an animated feature film. Rajesh Sharma, VP of Engineering at Spire, said the team can iterate interactively instead of "waiting 2 or 3 hours for a render to come back in a traditional setting." Yasu Seno and Bit192, Inc. trained a 20-billion parameter Japanese model from scratch using a slightly modified GPT-NeoX codebase, and released AI Novelist, a writing assistant used by professional writers, editors and casual users across Japan.
On the physical side, CoreWeave opened a data center in the New York City region (LGA1), which it says offers sub 1ms latency to Manhattan. Its new headquarters is in Roseland, New Jersey, with offices in Brooklyn, New York, and Philadelphia, Pennsylvania. The team has grown to 70+ employees across the U.S.
What it means for users
For AI teams, VFX studios and start-ups that rely on rented GPU capacity, the post signals that CoreWeave plans to add H100 instances early in 2023 and has fresh capital to expand. Reserving capacity is possible now, with a stated starting price of $2.23/hr. Performance figures and the "one of the first providers" claim come from CoreWeave itself, and the post gives no further detail on how much capacity will be available.
