CoreWeave

💰 Pricing & Plans

🔧 API & Developer Experience

• •API Style: All three inference products expose OpenAI API-compatible endpoints. Existing OpenAI client libraries, agents, and tooling connect with minimal changes. Dedicated Inference is configured through the CoreWeave Inference management API (REST/JSON at api.coreweave.com, gRPC over HTTP/2, and a Terraform provider).
• •Deployment Options: Three ways to serve models: Serverless Inference (CoreWeave-managed catalog, auto-scaling, no infra to manage), Dedicated Inference (bring your own model weights on dedicated GPU infra with full control over gateways, scaling, and capacity reservations), and Inference on CKS (complete control over your serving stack, runtimes, and networking on CoreWeave Kubernetes Service).
• •NVIDIA GPU Access: H100, H200, B200, A100, L40S, L40, and GB200 NVL72 on demand with spot discounts of 40-60%. First cloud to deploy NVIDIA Vera Rubin NVL72 and GB300 NVL72 — early access to the newest NVIDIA platforms for those who need it.
• •Model Catalog (Serverless): CoreWeave-managed catalog of popular models — GLM 5.2, Kimi K2.6/K2.7, DeepSeek R1/V3, Llama 3.x, Qwen — plus the option to deploy any catalog model or bring your own weights via Dedicated. MLPerf records: CoreWeave trained DeepSeek V3 in ~2 minutes and leads inference speed/price-performance for Kimi K2.6.
• •Training + Inference Loop: Unlike pure inference relays, CoreWeave lets you train or fine-tune on the same GPU fleet (CKS, SUNK, Mission Control observability) and then serve via OpenAI-compatible endpoints — closing the training-to-inference gap for autonomous agent improvement.
• •Sandboxes / RL: CoreWeave Sandboxes (launched 2026) accelerate reinforcement-learning agent tool-use and model evaluation; a serverless RL capability was launched for building reliable RL reward systems. Good for teams that iterate agents, not just call static endpoints.
• •Storage & Data: AI Object Storage (2 GB/s per GPU throughput), distributed file storage, and dedicated VAST storage. The 0EM (Zero Egress Migration) program lets you migrate into CoreWeave with no egress fees and no lock-in.
• •Observability: Mission Control provides fleet lifecycle controller, node lifecycle controller, observability, security, and audit visibility — the operating standard for AI at scale. Tensorizer accelerates model loading for inference.
• •Rate Limits & Enterprise: Capacity through On-Demand, Spot, Capacity Plans, or reserved nodes. High-demand GPUs (e.g. GB200) may require waitlisting or reservations. Enterprise agreements with OpenAI, Anthropic, Meta, and hedge funds like Jane Street indicate strong capacity scale.

🧠 GPU-Native Training-to-Inference Capability

CoreWeave's defining capability is that it is a GPU-native cloud rather than a pure inference API relay. A team can rent an NVIDIA H100/H200/B200 instance on demand (or reserve capacity), train or fine-tune a model on CoreWeave Kubernetes Service (CKS) with Mission Control observability, then serve that same model through an OpenAI-compatible inference endpoint — Serverless (managed catalog) or Dedicated (bring-your-own weights). This closes the training-to-inference loop that pure relays like OpenRouter or Portkey cannot offer, because the serving happens on the same GPU fleet where the weights were built. CoreWeave leads MLPerf both on training (DeepSeek V3 in ~2 minutes) and inference price-performance (Kimi K2.6), and was first to deploy NVIDIA Vera Rubin NVL72 and GB300 NVL72. For API consumers, the OpenAI-compatible surface means you can prototype with a hosted catalog model and later swap to a Dedicated endpoint running your own fine-tune with zero client changes. The tradeoff: this power is aimed at engineers who can operate Kubernetes and reason about GPU capacity, not at a one-line drop-in replacement for a frontier API.

🌐 Regional Availability & Latency

CoreWeave is a US-headquartered AI cloud (Roseland, New Jersey) with data centers concentrated in the United States. In 2026 it expanded internationally: two UK data centers are operational, it announced expansion into Indonesia as its first Asia-Pacific node, and Sweden capacity via a Conapto partnership. There is no mainland China direct endpoint; the Inference management API and console are served from US-fronted regions. For China-based production deployments, expect to use a proxy or a regional provider with local presence. Latency wise, CoreWeave is optimized for North American and European user bases: cross-Pacific requests from Asia to US regions typically add 150-250ms of first-byte latency versus ~50-80ms from a regional provider like Alibaba Bailian, ByteDance Volcano Engine, or Cloudflare Workers AI. For teams serving APAC users, the Indonesia expansion (2026) is the first step toward reducing that distance, but mainland China latency remains a factor. Choose CoreWeave for training-to-inference workloads anchored in the US/EU; pick a regional provider for latency-sensitive Asia-facing traffic.