CoreWeave is a specialized GPU cloud provider that raised over $12 billion in funding to build data centres exclusively for AI workloads, offering NVIDIA H100 and H200 instances at prices 30-50% below AWS and Azure equivalents. If you need dedicated GPU capacity without hyperscaler lock-in, CoreWeave has become the most credible alternative in the market.
What Is CoreWeave and How Did It Become a GPU Cloud Giant
CoreWeave started in 2017 as Atlantic Crypto, a cryptocurrency mining operation run by three founders: Michael Intrator, Brian Venturo, and Brannin McBee. When the crypto market crashed in 2018, they pivoted to renting out their NVIDIA GPUs for general compute workloads. By 2022, the AI training boom created explosive demand for GPU capacity that hyperscalers could not fill fast enough. CoreWeave repositioned itself as a pure-play GPU cloud, and the timing proved transformative.
In 2023, CoreWeave secured a $2.3 billion debt facility backed by its NVIDIA GPU inventory. The company then raised a $1.1 billion Series C round led by Coatue Management in late 2023, followed by an $8.6 billion funding round in early 2025 that valued the company at approximately $28 billion. CoreWeave completed its IPO on the Nasdaq in March 2025 at a $23 billion valuation, raising $1.5 billion. Total contracted revenue backlog exceeded $15 billion at the time of IPO, with major customers including Microsoft, Meta, and several leading AI labs.
CoreWeave GPU Pricing vs AWS and Azure
You will find CoreWeave consistently undercuts AWS and Azure GPU pricing by a significant margin. CoreWeave offers NVIDIA H100 SXM instances at approximately $2.06 per GPU-hour on reserved contracts, compared to $3.50-$4.00 per GPU-hour for equivalent AWS P5 or Azure ND H100 v5 instances on demand. H200 instances are available at roughly $3.19 per GPU-hour, and the newer GB200 NVL72 configurations launched in late 2025 at around $3.82 per GPU-hour.
The pricing advantage comes from CoreWeave’s architecture. Unlike hyperscalers that spread capital across hundreds of services, CoreWeave builds infrastructure optimized exclusively for GPU compute. Its data centres use direct liquid cooling across all GPU racks, achieving higher density per square foot and lower power consumption per TFLOP. This specialization translates to lower per-unit costs that CoreWeave passes through as competitive pricing.
How CoreWeave Networking Compares to Hyperscaler Clusters
CoreWeave deploys InfiniBand NDR networking at 400 Gb/s between nodes, matching what Azure offers on its ND H100 v5 series. Multi-node training jobs benefit from RDMA (Remote Direct Memory Access) support across all GPU instances. If you are evaluating Lambda Cloud vs AWS GPU options, CoreWeave sits between them, offering better networking than Lambda while maintaining pricing well below AWS.
Who Uses CoreWeave and Why
CoreWeave’s customer base spans three segments. First, AI research labs that need thousands of GPUs for large model training runs and cannot secure sufficient allocation from hyperscalers. Second, enterprises running inference workloads at scale, where CoreWeave’s lower per-hour pricing directly reduces cost per query. Third, companies like Microsoft itself, which signed a multi-billion dollar deal with CoreWeave in 2024 to supplement its own Azure GPU capacity during periods of peak demand.
The company operates 32 data centre locations across North America and Europe as of early 2026. Its strategy of leasing and building facilities near major power substations allows rapid deployment, with new data centres going operational within 9-12 months compared to the 2-3 year timelines typical for hyperscaler campuses. For teams evaluating the best cloud for AI training, CoreWeave deserves serious consideration alongside the traditional big three.
Risks and Limitations of CoreWeave
CoreWeave carries significant financial risk. The company reported $1.9 billion in revenue for 2024 against $2.5 billion in operating expenses and $8 billion in total debt. Customer concentration is also a concern, with Microsoft accounting for approximately 62% of contracted revenue. If Microsoft reduces its CoreWeave usage as its own GPU capacity scales up, the financial impact would be severe. You should also consider that CoreWeave lacks the breadth of services that hyperscalers provide, including managed databases, serverless compute, and global CDN networks.
Frequently Asked Questions
Is CoreWeave cheaper than AWS for AI training?
Yes. CoreWeave H100 instances cost approximately $2.06 per GPU-hour on reserved plans, compared to $3.50-$4.00 on AWS P5 on-demand instances. The gap narrows on AWS reserved pricing but CoreWeave typically remains 20-35% cheaper for equivalent GPU configurations.
Can CoreWeave handle large-scale training runs?
CoreWeave supports multi-thousand GPU training clusters with InfiniBand NDR networking at 400 Gb/s. Several AI labs have run training jobs exceeding 4,000 H100 GPUs on CoreWeave infrastructure, making it viable for frontier model development.
What are the biggest risks of using CoreWeave?
The primary risks are financial stability, customer concentration (62% revenue from Microsoft), and limited service breadth compared to AWS or Azure. CoreWeave also lacks the global data centre footprint of hyperscalers, which may matter for latency-sensitive inference deployments across multiple regions.