Nscale
Listed at https://www.nscale.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Nscale 采用充值余额（credit）按量计费，价格按模型分列，在 Nscale 控制台 AI Services → Models 与 /v1/models API 中按每 1M 输入 token 报价（文本模型）；图像模型输入按 0 计费。定价结构经 serverless-openapi.yaml ModelPricing schema 验证（input = 每 1M 输入 token 美元价）。 | per million tokens |
| Output | 输出按每 1M 输出 token 计费（文本模型）；图像模型按每百万像素计费（ModelPricing.output 定义为 per-million output pixels）。具体美元单价在控制台与 /v1/models 端点内按模型展示，注册充值后可查。 | per million tokens |
| Cache Read | Nscale 文档未单独公布"缓存输入"折扣档——计费以 model 为单位按 input/output 两档计（对照 OpenAI/Anthropic 的缓存输入差价，Nscale 暂无独立缓存价格线）。 | Discounted |
🤖 Supported Models (10)
✨ Pros
- ✓Serverless inference: managed endpoints ship in minutes with autoscaling — no clusters, GPUs, or infrastructure to operate
- ✓OpenAI-compatible: base_url=https://inference.api.nscale.com/v1, official OpenAI SDKs work as a drop-in
- ✓Model library: Qwen3, GPT OSS, Llama 4, Kimi K2.5, Mistral 8x22B, FLUX.1, SDXL open models
- ✓Fine-tuning + Prompt Workbench: browser prompt iteration, versioning, and cost visibility
- ✓EU data sovereignty: UK/Norway/Iceland data centers with in-country processing for GDPR and data-residency compliance
- ✓Full stack: on-demand GPU instances, Kubernetes (NKS), Slurm training, GB300 reserved capacity
⚠️ Cons
- ×Pricing not public: per-model rates require logging into the Console or calling /v1/models (needs a key) — no open pricing page, less transparent than OpenAI/Anthropic
- ×Newer brand: ecosystem (third-party SDKs, tutorials, community) is far smaller than OpenAI/Anthropic/Google
- ×No mainland China direct endpoint; positioned for Europe/US regions
- ×Requires prepaid credit before using the API, and signup is Google SSO only
- ×No public free tier (only possible promotional credits, not guaranteed)
🎯 Best For
European teams that want OpenAI-compatible managed serverless inference for open models (Qwen3, GPT OSS, Llama, Kimi, Mistral); enterprises and governments prioritizing EU data sovereignty, in-country processing, and GDPR compliance; developers who want a one-stop full-stack (serverless + on-demand GPUs + fine-tuning) and accept prepaid credit + Google SSO signup.
💰 Pricing & Plans
| Plan / Model Tier | Billing | Input | Output | Notes |
|---|---|---|---|---|
| Serverless text models | Per 1M tokens (credit) | Per-model rate | Per-model rate | Qwen3, GPT OSS, Llama 4, Kimi K2.5, Mistral 8x22B |
| Image generation | Per output pixels | 0 | Per 1M px | FLUX.1 [schnell], SDXL Base 1.0 via /v1/images/generations |
| Dedicated inference | Reserved endpoint | Per-model | Per-model | Strict customer isolation for compliance |
| Fine-tuning | Per job/usage | — | — | Customize foundation models; governed production flow |
| Compute instances | Per GPU-hour | — | — | On-demand GPU/CPU VMs, Kubernetes (NKS), Slurm |
🔧 API & Developer Experience
- •API Style: OpenAI-compatible REST API at https://inference.api.nscale.com/v1 — the official OpenAI Python/TypeScript SDK works as a drop-in by changing base_url.
- •Endpoints: Chat Completions (/v1/chat/completions), Completions (/v1/completions), Embeddings (/v1/embeddings), Image Generation (/v1/images/generations), plus /v1/models with per-model pricing + context length.
- •Auth: Bearer service token (JWT) created in Settings → Service Tokens; tokens are scoped, time-limited, and shown only once at creation.
- •Serverless vs Dedicated: Serverless autoscaling endpoints ship in minutes with integrated observability; dedicated endpoints add strict customer isolation for compliance-heavy teams.
- •Fine-tuning: Customize foundation models on your own data with a streamlined workflow, then move tuned models into production in a repeatable, governed flow.
- •Prompt Workbench: Browser-based prompt experimentation with versioning, preset saving, and cost visibility — reproducible runs without burning GPU hours.
- •Model Catalog: Open models only (Qwen3, GPT OSS, Llama 4, Kimi K2.5, Mistral, FLUX.1, SDXL) — no proprietary frontier models like GPT-5 or Claude; catalog managed per organization.
🧠 Sovereign Serverless Inference
Nscale's differentiator is a full-stack AI cloud built around European data sovereignty. Serverless inference runs on Nscale-managed GPUs across UK (Loughton), Norway (Glomfjord, Narvik), and US (Texas, West Virginia) data centers, with in-country processing to satisfy GDPR and data-residency requirements. The model catalog is open-weights only — Qwen3, GPT OSS, Llama 4, Kimi K2.5, Mistral 8x22B, FLUX.1, and SDXL — plus fine-tuning and a Prompt Workbench, so teams move from experimentation to production on one governed platform. Nscale's raising of a $900M revolving credit facility and its acquisition of Anyscale (announced late July 2026) signal aggressive expansion of this sovereign AI footprint.
🌐 Regional Availability & Latency
Nscale concentrates capacity in Europe and North America: own data centers at Loughton (UK), Glomfjord and Narvik (Norway), Texas and West Virginia (US), with partner-run sites in Sines (Portugal), Keflavik (Iceland), and Stavanger/Oslo (Norway). inference.api.nscale.com is a single global endpoint, so European and US clients see low single-digit tens of ms first-byte latency, while mainland China needs a proxy with typical 200-350ms transatlantic/transpacific latency. There is no Asia-Pacific or China-region endpoint. Teams prioritizing in-country data processing, GDPR alignment, or UK/Nordic sovereignty should consider Nscale; low-latency China workloads are better served by domestic providers.