CoreWeave

What CoreWeave actually is

CoreWeave is a GPU cloud built specifically for AI training and inference workloads, sold primarily as whole NVIDIA nodes rather than as general-purpose virtual machines. It is not a VPS provider with a GPU SKU bolted on: the product is dedicated accelerated compute, managed Kubernetes and Slurm, InfiniBand-connected clusters, and the storage and networking that keep eight-figure GPU fleets fed. In our view it belongs in a different buying conversation than general GPU cloud options — this is infrastructure for teams whose bottleneck is GPU-hours, not web traffic.

The company has an unusual origin story that is worth knowing before you sign anything. It was formed in September 2017 in Delaware as The Atlantic Crypto Corporation, a cryptocurrency mining venture, renamed CoreWeave in December 2019 when it pivoted to accelerated cloud computing, and headquartered in Livingston, New Jersey. It filed its S-1 in March 2025 and listed on Nasdaq under CRWV later that month at $40.00 per share. NVIDIA is a shareholder and a close partner, which cuts both ways: strong supply relationships, and real dependence on one vendor's roadmap.

The scale claim on CoreWeave's own pages is 50+ active data centers across the United States and Europe, and the marketing line is that 9 of 10 leading AI labs build and run AI on the platform. Take the customer claim as marketing; the footprint claim is the one you can plan against.

Pricing: what you really pay

CoreWeave publishes a public pricing page, which we count in its favor — a large share of the GPU cloud market still hides numbers behind "contact sales." Current on-demand rates for the eight-GPU HGX nodes are $49.24 per hour for H100 (80 GB), $50.44 per hour for H200 (141 GB) and $68.80 per hour for B200 (180 GB). Spot capacity runs roughly half of on-demand: the H100 node drops to $19.71 per hour and the B200 node to $34.11 per hour, with preemption risk attached. Mid-range and inference-oriented lines are cheaper: an 8-GPU L40 node is $10.00 per hour on-demand ($6.27 spot), the RTX PRO 6000 Blackwell node is $20.00 per hour, and a single-GPU GH200 is $6.50 per hour. Several newer configurations — GB300 NVL72, HGX B300, the standard-memory RTX PRO 6000 — show spot prices but route on-demand through sales.

There is also a per-GPU "classic" line where vCPU, RAM and storage are billed on top: third-party price trackers list the H100 SXM at $4.76 per GPU-hour on that basis, plus components. Storage follows the same metered pattern — around $0.07 per GB-month for distributed file storage and $0.06 per GB-month hot object storage in the numbers we saw, with a $4.00 monthly charge per public IPv4 address and egress reported as free by at least one tracker. We treat the component figures as indicative because they are tracker-rendered; the node rates above are straight off the vendor's page.

Two structural points about the money. First, minimums are real: most capacity is sold as whole 8-GPU nodes, so your practical entry point is tens of dollars per hour, not a cheap sandbox. Single-GPU inference instances and spot pools soften this, but the product is built for sustained load. Second, the honest comparison is against spot resellers, where a preemptible H100 can cost less than half of CoreWeave's per-GPU rate. You pay CoreWeave's premium for dedicated hardware, no preemption on on-demand, NVLink within nodes, InfiniBand across them and a contractual SLA. Whether that premium is worth it depends entirely on whether a preemption would cost you a training run.

We have not benchmarked CoreWeave hardware ourselves, so we make no throughput claims here. Billing is hourly usage invoiced in arrears; multi-year commitments are reported to discount 15-30%, negotiated on order forms.

Performance & infrastructure

The hardware roster is the broadest tier of the NVIDIA catalog: L40 and L40S for inference and rendering, A100 and H100 for mainstream training, H200 for memory-bound inference, B200 and the GB200/GB300 NVL72 rack systems for frontier-scale work, plus the Grace Hopper GH200. Nodes carry serious local resources — the HGX H100 configuration pairs 8 GPUs with 128 vCPUs, 2,048 GB of system RAM and 61.44 TB of local NVMe — which matters because data pipelines starve GPUs faster than most teams expect.

The software layer is Kubernetes-first. CoreWeave's managed Kubernetes (CKS) and managed Slurm (SUNK) control planes are included at no extra charge, InfiniBand fabric runs at up to 3,200 Gbps across multi-node clusters, and the stack includes schedulers, health checking and observability aimed at long-running distributed jobs. Compliance posture is enterprise-grade — SOC 2 and ISO certifications are referenced across the ecosystem — and the platform supports reserved capacity and dedicated regions for customers who need physical or contractual isolation.

Reliability is contractual rather than anecdotal: the published commitment is 99.9% measured across regions and 99% within a single region, with service credits (a percentage of the affected month's bill) as the sole remedy. That single-region number is the one to plan around if your workload is pinned to one geography — a 99% single-region commitment is a full hour of downtime per month on paper.

Where it falls short

Cost is the headline. At $68.80 per hour for a B200 node, a full month of uninterrupted compute is a five-figure line item before storage and networking, and H100 nodes are not far behind. Teams that cannot keep hardware busy — experimental workloads, spiky research jobs, staging environments — will burn money quickly. Spot helps, but then you are back to preemption semantics and careful checkpointing.

Second, the buying experience still leans enterprise. Some configurations are sales-gated, commitment discounts live on order forms, and the classic per-GPU line bills CPU, RAM and storage separately, which is precisely the kind of component pricing that surprises people at invoice time. The published node prices are clear; the total bill is clearer only after you have modeled storage, IPs and networking.

Third, there is concentration risk. CoreWeave's supply, technology partnerships and a large share of its identity are tied to NVIDIA, and its balance sheet carries the heavy debt financing typical of the GPU build-out. For a strategic supplier, that is a diligence item, not a footnote. We would also note that the company is a comparatively young public entity with a checkered pre-history in crypto mining; the pivot has been thorough, but buyers of multi-year capacity should assess the counterparty as carefully as the GPUs.

Finally, this is not a general cloud. There is no cheap shared tier, no meaningful story for ordinary web hosting, and little reason to run non-GPU workloads here. If your need is a small always-on server, you are shopping in the wrong aisle — see our benchmark comparisons for that class of machine.

Who should pick this

Our pick for AI labs, ML platform teams and research groups running sustained training or high-throughput inference where a lost node costs more than the premium over spot pricing. If you need guaranteed access to H100, H200 or Blackwell capacity, NVLink within nodes and InfiniBand across them, a 99.9% cross-region commitment and an accountable counterparty, CoreWeave is one of the few providers that sells exactly that, with public prices to compare.

We would not pick it for hobby projects, intermittent GPU experiments, or cost-sensitive batch work where preemption is acceptable — spot-first resellers win those cases decisively. And we would not pick it as a general-purpose cloud for teams that also need cheap VMs and object storage at commodity prices; the fit is too poor.

For teams in between: price out the classic per-GPU line before committing to whole nodes, model your storage and IP costs explicitly, and negotiate on a term if you have steady demand. The list prices are honest; the discounting is where the enterprise game is played.

Key takeaways

• GPU-specialist cloud, founded 2017 (as Atlantic Crypto), Livingston NJ, Nasdaq-listed since March 2025; 50+ data centers across the US and Europe.
• Published on-demand node pricing: HGX H100 $49.24/hr, H200 $50.44/hr, B200 $68.80/hr, with spot at roughly half and single-GPU GH200 at $6.50/hr.
• Whole 8-GPU nodes are the default unit — practical entry is tens of dollars per hour, so this is sustained-load infrastructure, not a sandbox.
• Managed Kubernetes and Slurm control planes included, InfiniBand at up to 3,200 Gbps, enterprise compliance (SOC 2, ISO).
• SLA: 99.9% across regions, 99% single region, with service credits as the sole remedy.
• Pay the premium when uptime of the GPU matters; use spot resellers when it does not.

Sources

• https://coreweave.com/pricing
• https://coreweave.com/about-us
• https://www.sec.gov/Archives/edgar/data/1769628/000176962825000062/R9.htm
• https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798ds1.htm
• https://www.coreweave.com/news/coreweave-announces-pricing-of-initial-public-offering
• https://gpufind.cloud/providers/coreweave/review
• https://gpubank.io/compare/gpubank-vs-coreweave
• https://deploybase.ai/articles/coreweave-gpu-pricing
• https://docs.coreweave.com/products/storage/object-storage