🇺🇸 United States · CoreWeave — GPU Cloud for AI Workloads
Status: 🟩 COMPLETE 🟦 LIVING Section: 15 — Broader Tech Bonus
| Vendor | CoreWeave |
| Country/origin | 🇺🇸 United States (Roseland, New Jersey) |
| Recommended for AUS? | ✅ Yes — US-based; enterprise-grade; SOC 2 |
| Privacy summary | Enterprise data handling; SOC 2 Type II; standard cloud contracts; data in US/EU data centres |
| Free tier | No — credit-based trial only; enterprise-focused |
| Paid tiers | Per-GPU-hour pricing; enterprise contracts |
| First released | Founded 2017 (originally cryptocurrency mining); AI cloud pivot 2019; IPO 2025 |
| Last reviewed | June 2026 |
| Official site | https://coreweave.com |
What it is
CoreWeave is a specialist cloud computing provider focused entirely on GPU compute for AI workloads. Unlike general cloud providers (AWS, Azure, GCP) that offer GPU compute alongside dozens of other services, CoreWeave’s entire business is providing powerful GPU infrastructure — especially for AI model training and inference.
CoreWeave was originally a cryptocurrency mining company (using GPUs to mine Ethereum), but pivoted entirely to AI cloud services when AI demand exploded. They accumulated a massive inventory of NVIDIA GPUs and built a purpose-built cloud platform around them.
What makes CoreWeave different from AWS/Azure/GCP:
- GPU-first: Entire platform optimised for GPU workloads, not afterthought
- Better GPU availability: During periods of extreme NVIDIA GPU shortage (2022–2024), CoreWeave often had availability when AWS/Azure had waitlists
- Competitive pricing: Generally 30–50% cheaper than AWS/Azure for equivalent GPU instances
- Network: Specialised low-latency interconnects between GPUs (InfiniBand) — critical for large model training
- NVIDIA partnership: CoreWeave has a strategic partnership with NVIDIA; they often get new GPU hardware early
Key use cases:
- AI model training: Training large language models, image models, or custom AI models
- AI inference at scale: Hosting AI models for production API traffic
- HPC (High-Performance Computing): Scientific computing, simulation, rendering
Who uses CoreWeave
CoreWeave’s customers are AI companies and enterprises, not individual developers. Notable customers include:
- Microsoft / OpenAI: CoreWeave provides significant GPU capacity for OpenAI’s training workloads
- Stability AI: Used CoreWeave for Stable Diffusion training
- Mid-size AI startups: Teams that need more GPU flexibility or better pricing than AWS
How to access
CoreWeave is not a self-serve product in the way AWS is — it’s primarily sold through enterprise relationships and contracts. However:
- Go to https://coreweave.com → Contact Sales or Get Started
- For developers: CoreWeave Cloud (cloud.coreweave.com) provides Kubernetes-based access to GPU compute
- A credit trial is available for evaluation
This is primarily aimed at AI companies and engineering teams, not individual users.
What it costs
CoreWeave pricing is by GPU-hour, similar to other cloud providers. Approximate mid-2026 pricing for NVIDIA H100 GPUs:
- H100 SXM5 (80GB): ~2.40 USD/hour per GPU
- Compare: AWS p4d (A100) instances: ~$3.50+ USD/hour per GPU equivalent
For AI training runs that use hundreds of GPUs for weeks, CoreWeave’s cost savings can be substantial — hundreds of thousands of dollars.
How it compares to alternatives
| Provider | Specialisation | Best for |
|---|---|---|
| CoreWeave | GPU specialist; NVIDIA-focused | Large AI training; enterprise inference |
| Lambda Labs | Developer-friendly GPU cloud | Mid-size training; easier access |
| RunPod | Consumer-friendly GPU rental | Individual developers; smaller experiments |
| AWS | General cloud + GPU | Full AWS ecosystem integration |
| Azure | General cloud + GPU | Microsoft/OpenAI ecosystem |
| GCP | General cloud + GPU | Google TPU + GPU; Vertex AI |
| Modal | Serverless GPU | Developer-friendly; pay-per-second |
CoreWeave competes most directly with Lambda Labs for pure GPU cloud, and with AWS/Azure for enterprise GPU capacity.
Gotchas
- Not for individual developers / small projects. CoreWeave is designed for teams, not individual experimenters. The minimum commitment and complexity favour enterprise use.
- Australian latency: No Australian data centres as of mid-2026. US and EU locations available. Australian AI companies using CoreWeave should factor in cross-Pacific latency for interactive applications (less relevant for batch training).
- Kubernetes knowledge required. CoreWeave’s primary interface is Kubernetes — the container orchestration system used to manage large numbers of servers. If you don’t have a DevOps/infrastructure team, this is complex.
- GPU availability varies. Despite their inventory, specific GPU types can have waitlists during high-demand periods.
- Not a hyperscaler. CoreWeave provides GPU compute, not the full range of cloud services (databases, storage, networking) that AWS/Azure offer. You’ll likely use CoreWeave for GPU + other cloud providers for everything else.
Recent changes (LIVING)
- IPO (March 2025): CoreWeave listed on NASDAQ — one of the first pure-play AI infrastructure IPOs.
- Microsoft contract expansion (2024): Multi-billion dollar multi-year agreement with Microsoft for GPU capacity.
- EU data centres (2024): Expanded to European locations for EU-resident AI companies.
- CoreWeave Inference (2024): Managed inference service — deploy models without managing Kubernetes directly.
See also
- lambda-labs — more accessible GPU cloud alternative
- runpod — individual-developer GPU rental
- modal — serverless GPU compute
- aws-bedrock — managed AI on AWS (easier but more expensive)
- nvidia-ai — NVIDIA’s own AI cloud (CoreWeave is a major NVIDIA partner)
Sources
- CoreWeave official documentation: coreweave.com
- CoreWeave NASDAQ IPO filing (2025)
- Microsoft × CoreWeave partnership announcements (2023–2024)
- Sequoia Capital and NVIDIA investment in CoreWeave (2023)
- The Information, Bloomberg coverage of CoreWeave’s growth (2023–2025)