Crusoe

💰 Pricing & Plans

🔧 API & Developer Experience

• •API Style: OpenAI-compatible REST API for Managed Inference (POST /v1/chat/completions + /v1/completions + /v1/embeddings). Existing OpenAI / Anthropic SDK clients can target Crusoe with a base URL change and API key swap — no code rewrite. Aug 12 Pydantic AI native integration confirmed (PydanticAI SDK lists Crusoe as a first-class model provider).
• •Model Catalog: 17+ open-weight models on Managed Inference (Aug 2026): DeepSeek V3 0324, V4 Pro, V4 Flash; GLM 5.1 / 5.2; GPT-OSS 20B / 120B; Gemma 4 31B-it; Kimi K2.6; Llama 3.1 8B Instruct, 3.3 70B Instruct; Nemotron 3 VoiceChat, Ultra 550B, 3.5 Lightning, 3 Nano, 3 Nano Omni 30B, 3 Super 120B; Qwen3 8B, 235B A22B Instruct 2507, Qwen3.5 2B/9B, Qwen3.6 35B A3B; Yutori n1.5.
• •MemoryAlloy™ Technology: Cluster-native memory fabric that persists across nodes, with a cluster-wide KV cache that is reused across requests. Crusoe's published data: up to 9.9x faster time-to-first-token and 5x higher throughput via speculative decoding + dynamic batching. Designed to make prefix-cache hits automatic for long-context RAG / agent workloads.
• •Serverless Inference: Managed Inference runs serverless — no capacity planning, no warm-up, no idle GPU costs. Each request is auto-routed to the cheapest tier that fits the model. Per-token billing only; no monthly minimums.
• •Serverless Fine-Tuning (Jul 2026): 4-tier pay-as-you-go fine-tuning, same parameter-count pricing as inference ($0.40 / $2.50 / $6.00 / $10.00 per 1M tokens). Customize open models on proprietary data without standing up GPU clusters. Still early ecosystem — fine-tune-result quality benchmarks are limited.
• •Self-Serve Deployments: Spin up dedicated endpoints for any open or fine-tuned model in minutes, no sales engagement. NVIDIA H100 80GB HGX $5.50/GPU-hour, NVIDIA H200 141GB HGX $6.00/GPU-hour. Best for AI labs that need benchmarked, reproducible latency and dedicated capacity.
• •Tailored Deployments + Provisioned Throughput: Sales-engagement tiers for highest-level optimization (custom benchmarked endpoint + bring-your-own-model) and reserved AI Model Units (AMUs) for guaranteed throughput. Longer commitment = lower unit price. No public list pricing.
• •Object Storage + Container Registry: Object Storage $0.06/GiB-month for datasets and model artifacts; Container Registry $0.10/GiB-month for images. Managed Kubernetes $0.10/cluster-hour ties it all together for AI app deployment across GPU + CPU resources.
• •Pricing Predictability: Per-token billing on Managed Inference (no hourly reservation) + per-minute billing on Self-Serve GPU rentals (no idle waste) = both deployment models are cost-transparent. The 4-tier pricing rule means estimating spend from a model name only requires knowing its parameter count, not chasing individual per-model prices.
• •Rate Limits & Error Handling: Standard HTTP 429 with Retry-After on rate-limit; detailed error codes for context-length / token-budget / model-not-found / quota-exceeded. Rate-limit thresholds scale with spend — high-volume customers get higher ceilings via Tailored Deployments.

🧠 MemoryAlloy & Open-Model Inference Capabilities

Crusoe's flagship product is Managed Inference for open-weight models, powered by a proprietary cluster-orchestration layer called MemoryAlloy™. Unlike hyperscaler LLM endpoints that put KV cache in single-instance GPU memory, MemoryAlloy persists a cluster-wide key-value cache and uses intelligent routing to eliminate redundant prefill computation when multiple requests share a system prompt. Combined with speculative decoding and dynamic batching, Crusoe reports up to 9.9x faster time-to-first-token and 5x higher throughput versus a naive single-instance deployment. The 17+ model catalog covers the full open-weight frontier — DeepSeek V4 Pro and V4 Flash for agent workloads, GPT-OSS 120B for budget 70B+ inference, Qwen3 235B A22B Instruct for the MoE sweet spot, Kimi K2.6 for Moonshot quality, GLM 5.2 for the >300B tier, plus full Llama 3.x, Gemma 4, and Nemotron 3 families. For teams running multi-agent or RAG systems on long shared contexts, MemoryAlloy is the single biggest technical differentiator — the same workload on Together AI or Fireworks AI pays full prefill cost on every request.

🌐 Regional Availability & Latency

Crusoe is a US-based AI infrastructure company (San Francisco headquarters) with data center operations concentrated in the United States (Colorado, Texas, and other US locations used by the Crusoe AI factory concept — facilities purpose-built to convert stranded energy into AI compute). The Managed Inference public API is served from these US regions; there is no mainland China direct endpoint and no announced APAC or EU regional expansion as of August 2026. China-based production deployments require either an overseas proxy or running Self-Serve Deployments on a different provider that has local presence. Latency-wise, Managed Inference is optimized for North American + European user bases; cross-Pacific requests from Asia to US Crusoe regions typically see 150-250ms first-byte latency versus ~50-80ms from a regional provider like Cloudflare Workers AI, Alibaba Bailian, or ByteDance Volcano. For teams in those regions, MemoryAlloy's cluster-cache benefit does not outweigh the network cost — pick a regional provider instead.