CoreWeave, a cloud computing startup that specialises in providing GPU resources, has seemingly broken into the ranks of hyperscalers by raising a staggering $7.5 billion in debt financing from investors, including Blackstone and Magnetar, in one of the largest-ever private debt financings. 

In addition to this, the company recently announced a $1.1 billion Series C funding round led by Coatue, with participation from Magnetar, Altimeter Capital, Fidelity Management & Research Company, and Lykos Global Management.

With this, CoreWeave has raised $12 billion in equity and debt investments over as many months. It plans to use the funds to scale its data centre capacity to 300 megawatts, doubling the number of centres from 14 to 28 by year-end. 

Coreweave co-founder and CSO Brannin McBee recently mentioned that the company has already secured power and data centre space to more than double that amount in 2025.

This, coupled with the fact that they grew from three to 14 data centres and quadrupled their employees over the past year, speaks highly of the pace of their growth.

McBee also mentioned that another debt deal could happen this year, further fueling the company’s growth.

McBee believes the funding rounds have propelled CoreWeave into the ranks of hyperscalers, alongside industry giants like Amazon, Google, Microsoft, and Oracle. 

However, he emphasised that CoreWeave stands out as an “AI hyperscaler”, differentiating itself through its specialised focus on AI infrastructure.

US-based Coreweave, which started out as a GPU provider to crypto-miners in 2017, witnessed its revenue skyrocket after the pivot.

While the company lowered its projected revenue for the year from $630 million to just over $500 million, it still registered a revenue of $440 million in 2023, an increase of more than 17-fold or 1,660% from $25 million in 2022.

Looking ahead, they are even more optimistic, with projections of $2.3 billion in revenue for 2024 based on having over $7 billion in signed cloud contracts through 2026.

This rapid growth has attracted the attention of major investors, including NVIDIA, which has taken an equity stake in the company.

Competing with Cloud Giants

Despite the impressive revenue, the company still has a long way to go to catch up with the established cloud giants. 

In Q1 of 2024, AWS and Google Cloud generated $25 billion and $9.6 billion respectively. On the other hand, Microsoft’s Intelligent Cloud group, which includes Azure, reported a revenue of $26.7 billion.

However, CoreWeave’s focus on AI workloads, alongside its close partnership with NVIDIA and access to its latest AI chips, have given it a competitive edge in this fast-growing market segment.

NVIDIA is using its market dominance in GPUs to build up CoreWeave as a counterbalance to the cloud computing giants. The giant which supplies a majority of CoreWeave’s chips, views this trend favourably, possibly for reasons related to leverage. 

It is reported that NVIDIA has granted preferential access to its GPUs to certain alternative cloud providers.

NVIDIA allocated a significant portion of its limited H100 supply to CoreWeave and other smaller cloud providers, while larger players like AWS and Microsoft Azure struggled to secure adequate supplies. 

This allowed CoreWeave to attract major customers, including Microsoft itself, which signed a deal to rent GPU servers from CoreWeave to support its own Azure cloud customers. 

This would also ensure sufficient computational power for OpenAI to train its generative AI models.

CoreWeave boasts an impressive GPU infrastructure, delivering access to over 45,000 GPUs and offering the industry’s broadest range of NVIDIA GPUs with 13 SKUs available on demand. 

The company has deployed the largest fleet of NVIDIA A40 GPUs in North America and provides a variety of NVIDIA GPU types purpose-built for different use cases. 

This includes the Quadro RTX series and NVIDIA Ampere architecture GPUs like the A100 for PCIe (80GB), a40, H100 HGX (80GB), L40, L40S, GH200, B200, empowering businesses to harness the latest advancements in GPU technology for their computing needs.

CoreWeave also expects to have the availability of NVIDIA’s new GB200 chip in early 2025, addressing the immense demand for this cutting-edge technology.

The exact number of GPUs CoreWeave has in comparison to hyperscalers like Amazon, Google, and Microsoft might be lower. However, CoreWeave emphasises a key difference in its approach. It offers a much broader range of NVIDIA GPU types, contrasted with the “one size fits all” approach of large, generalised clouds. Furthermore, CoreWeave claims to deliver a performance-adjusted cost of up to 80% less expensive than its competitors.

For instance, Coreweave charges $2.39 per hour to rent an NVIDIA A100 40GB GPU, which is commonly used for model training and inference. This translates to a monthly cost of $1,200. 

In contrast, the same GPU on Azure costs $3.40 per hour, or $2,482 per month, while on Google Cloud, it costs $3.67 per hour or $2,682 per month.

CoreWeave’s Edge

Despite being much smaller, the company, in its investor pitch has drawn direct comparisons to major cloud providers, particularly AWS. 

The company’s documents have previously projected a gross margin of 85%, derived from the difference between operational costs and rental revenue. 

CoreWeave’s 2023 revenue is also higher than that of some competitors in the specialised GPU cloud market, such as Lambda Labs ($250 million) and Crusoe Energy ($100 million). 

However, it is still significantly lower than major cloud providers like AWS, Microsoft Azure, and Google Cloud.

But CoreWeave has a certain competitive edge. The company operates on a rental model, where customers pay for the computing power they use per hour and also offers additional services like data storage, networking, and CPU compute.

The cost of a high-end H100 PCIe card for CoreWeave is roughly $30,000, which is rented out at an average of $4.25 per hour. Assuming an 80% utilisation rate, it would generate about $29,473 in annual revenue, roughly breaking even. 

However, cheaper GPUs like the A40, purchased in bulk before the generative AI boom, could generate much higher margins. An A40 with a sticker price of $4,500 three years ago, rented out at $1.278 per hour, could generate $8,877 in annual revenue at 80% utilisation.

Together, a GPU reseller that rents GPUs from various sources and bundles them with AI training software, hit a $10M annual revenue run rate at the end of 2023, with 90% coming from their Forge product launched in June 2023.

Looking forward, CoreWeave’s advantage depends on the long-term state of the GPU industry and their ability to build a differentiated AI compute platform. 

The current GPU shortage, driven by limitations at TSMC, is expected to last until March 2026, with a new $2.9B packaging facility operational in 2027.

CoreWeave’s infrastructure, designed from the ground up for GPU compute at scale, has shown promising results. Its HGX H100 instances delivered benchmark results 29x faster than the next-fastest competitor. 

As AI workloads increase in size and complexity, CoreWeave’s specialisation in serving customers with the biggest compute needs could protect their moat even as GPU shortages ease.

Overcoming Bottlenecks

CoreWeave faces challenges, with a significant portion of its revenue coming from a small number of large customers like Microsoft and Inflection, while competitor Lambda Labs has a more extensive customer base. 

The company also reduced its 2023 revenue projection, possibly indicating difficulties in securing sufficient chips or data centre space. 

Bernstein analyst Stacy Rasgon has also noted that despite NVIDIA’s prioritisation of CoreWeave, the high demand for AI chips from various parties could lead to growing pains for cloud providers, who are struggling to keep pace with the AI boom. 

Data centre operators face challenges in establishing facilities to accommodate NVIDIA’s power-hungry chips and meet the surging demand. 

Despite these challenges, CoreWeave has been proactive in securing over 300 megawatts of data centre capacity and designing purpose-built facilities to handle AI workloads. 

The company’s strong partnerships with NVIDIA and major customers, combined with its early-mover advantage, position it well for continued growth in the AI infrastructure market.

The post How This Crypto-Miner Turned AI Hyperscaler appeared first on AIM.