Nscale is a full-stack AI infrastructure company headquartered in the UK with data centers in Loughton (UK), Glomfjord and Narvik (Norway), and Texas and West Virginia (US). It runs a serverless inference API at inference.api.nscale.com that is OpenAI-compatible, serving open-weight models only — Qwen3, OpenAI GPT OSS, Meta Llama 4, Moonshot Kimi K2.5, Mistral 8x22B, plus FLUX.1 and Stable Diffusion for image generation. Nscale emphasizes European data sovereignty and in-country processing for GDPR and data-residency compliance.

How does Nscale pricing work?

Nscale uses prepaid, credit-based pay-as-you-go billing. You top up credit via Stripe in the Nscale Console, then spend per model. Text models bill per 1M input and per 1M output tokens; image models bill per million output pixels. Exact per-model USD rates are shown in the Console under AI Services → Models and via the /v1/models API endpoint (which requires a Bearer service token). There is no permanent free tier, though the quickstart notes early users may qualify for promotional credits.

Is Nscale cheaper than OpenAI or Anthropic?

Because Nscale serves open-weight models (Qwen3, GPT OSS, Llama 4, Kimi K2.5, Mistral) rather than proprietary frontier models, token prices tend to sit below GPT-5 or Claude Opus tiers — but Nscale does not publish an open pricing page. You must log in or call /v1/models to see per-model rates, which makes direct head-to-head comparison harder than with OpenAI or Anthropic who publish full public pricing.

What models are available on Nscale serverless inference?

Nscale serves open models only. The featured catalog (verified on nscale.com/product/inference) includes Moonshot Kimi K2.5, Alibaba Qwen3 8B and Qwen3 4B (Thinking + Instruct), Meta Llama 4 Scout, Mistral 8x22B Instruct, OpenAI GPT OSS 20B and GPT OSS 120B, Black Forest Labs FLUX.1 [schnell], and Stability AI Stable Diffusion XL Base 1.0. Chat, vision, embeddings, and image generation are exposed through OpenAI-compatible endpoints.

Does Nscale work in China?

Nscale has no mainland China direct endpoint. inference.api.nscale.com is a single global endpoint with data centers concentrated in Europe and the US, so China production traffic needs a proxy with typical 200-350ms transatlantic/transpacific latency. Nscale positions itself for European data sovereignty and GDPR-aligned in-country processing rather than China-region low latency.

Did Nscale acquire Anyscale?

Yes — Nscale announced the acquisition of Anyscale in late July 2026 (verified via the Nscale newsroom and multiple industry outlets). Anyscale is already listed as a separate provider on APIRank. The acquisition, plus a $900M revolving credit facility, signals Nscale's aggressive expansion of its full-stack sovereign AI cloud. The two entries remain separate because they serve different surfaces: Anyscale for its Ray ecosystem and serverless model serving, Nscale as the underlying GPU infrastructure + inference platform.

Nscale Serverless Inference API Pricing 2026: Open Models & EU Sovereignty

Most "AI API" reviews in 2026 start from a proprietary frontier model and walk down to price. Nscale inverts that: it is a full-stack AI infrastructure company — data centers, GPUs, Kubernetes, Slurm training, fine-tuning, and a serverless inference API at https://inference.api.nscale.com/v1 — that serves open-weight models only. The catalog (verified on nscale.com/product/inference, captured 2026-08-27) is Qwen3 4B/8B, OpenAI GPT OSS 20B/120B, Meta Llama 4 Scout, Moonshot Kimi K2.5, Mistral 8x22B Instruct, plus FLUX.1 [schnell] and Stable Diffusion XL for image generation. The pitch is European data sovereignty: in-country processing at UK and Norway data centers, GDPR alignment, and OpenAI-compatible endpoints so a developer can point the official OpenAI SDK at Nscale by changing base_url. This review covers the verified API surface, the credit-based pricing model, and where Nscale fits for a team deciding between an open-model aggregator like Fireworks or Groq and a proprietary vendor.

🌍 Quick verdict: Nscale is a full-stack, EU-sovereign AI cloud whose serverless inference API is OpenAI-compatible and serves open models only (Qwen3, GPT OSS, Llama 4, Kimi K2.5, Mistral, FLUX.1, SDXL). Pricing is prepaid, credit-based, per-model — text per 1M input/output tokens, images per million output pixels — with per-model rates shown in the Console and via /v1/models. There is no permanent free tier and no public pricing page, which lowers transparency vs OpenAI/Anthropic. The differentiator is sovereign compute: UK, Norway, and US data centers with in-country processing for GDPR and data-residency workloads, plus a $900M credit facility and the late-July 2026 acquisition of Anyscale signaling aggressive expansion.

What Nscale actually is

Nscale is not a model vendor — it is an infrastructure company with its own data centers. The homepage describes it as "the engine of superintelligence," and the offering splits into four layers: AI Services (inference endpoints, fine-tuning, Prompt Workbench), Platform Services (managed Slurm, Kubernetes, instances), Infrastructure Services (compute, networking, storage), and Fleet Operations (observability, Radar API). Own data centers are at Loughton (UK), Glomfjord and Narvik (Norway), and Texas and West Virginia (US), with partner-run sites in Sines (Portugal), Keflavik (Iceland), and Stavanger and Oslo (Norway). CEO and founder Josh Payne leads the company, which the newsroom keeps positioned around sovereign AI, in-country data residency, and sub-30ms edge inference.

For an AI API user, the entry point is serverless inference — fully managed endpoints that you call with a Bearer service token and OpenAI-compatible payloads. The /v1 surface mirrors OpenAI: /v1/chat/completions, /v1/completions, /v1/embeddings, /v1/images/generations, and /v1/models (which returns per-model pricing and context length). I verified the live endpoint returns HTTP 401 Unauthorized without a token — a clean signal the API exists and needs an API key, unlike a defunct or placeholder surface.

Nscale model catalog (verified)

Model	Vendor	Type
Qwen3 8B / Qwen3 4B Instruct / Qwen3 4B Thinking	Alibaba Cloud	Chat + reasoning
GPT OSS 20B / GPT OSS 120B	OpenAI	Chat (open weights)
Llama 4 Scout	Meta	Multimodal chat
Kimi K2.5	Moonshot AI	Chat
Mistral 8x22B Instruct	Mistral AI	Mixture-of-experts chat
FLUX.1 [schnell]	Black Forest Labs	Image generation
Stable Diffusion XL Base 1.0	Stability AI	Image generation

The catalog is deliberately open-model-only. There are no proprietary frontier models (no GPT-5, no Claude, no Gemini) beyond the open GPT OSS line. That makes Nscale a direct alternative for teams that want to run and pay for open weights on managed infrastructure — the same niche as Fireworks AI, Groq, and Together AI — but with the added twist that Nscale owns the GPUs and data centers underneath instead of renting them, giving it tighter control over sovereign placement thresholds.

Pricing model (verified structure)

Nscale uses prepaid credit-based billing: you sign up with Google SSO, top up credit via Stripe in the Console, then spend per model. The pricing structure is verified against the ModelPricing schema in the public serverless-openapi.yaml: text models bill input per 1M input tokens and output per 1M output tokens; image models bill input as 0 and output per million pixels. The per-model USD rates themselves are served from the Console and the authenticated /v1/models endpoint, not from an open pricing page.

⚠️ Transparency trade-off: unlike OpenAI, Anthropic, or OpenRouter, Nscale does not publish a public pricing table. Per-model token costs require a logged-in Console session or an authenticated /v1/models call. This is the main friction for a cost-comparison review — the structure is verified, the exact numbers are not public.

API & developer experience

OpenAI-compatible endpoints: base_url=https://inference.api.nscale.com/v1; the official OpenAI Python and TypeScript SDKs work by changing base_url and passing the service token.
Four modalities on one surface: chat completions, text completions, embeddings, and image generation are all under /v1, matching OpenAI's shape so existing tooling mostly ports over.
Service-token auth: scoped, time-limited JWTs created in Settings → Service Tokens, shown once at creation — a more modern auth model than plain API keys.
Serverless vs dedicated: autoscaling serverless endpoints with integrated observability, or dedicated endpoints with strict customer isolation for compliance-heavy teams.
Fine-tuning + Prompt Workbench: customize foundation models on your own data and iterate prompts in a versioned browser workbench with cost visibility.
Full-stack optionality: the same account can spin up GPU instances, Kubernetes (NKS), Slurm training clusters, reservations (GB300 NVL72), and object storage — inference is one layer of a much bigger platform.

Sovereign inference & recent news

The strategic differentiator is sovereignty. Nscale positions in-country processing as the answer to GDPR and data-residency requirements: a UK or Norwegian enterprise can run inference on GPUs physically inside its own jurisdiction, on dedicated infrastructure it does not have to operate. Two recent signals confirm the expansion trajectory. First, Nscale closed a $900 million revolving credit facility (newsroom, 2026). Second, Nscale announced the acquisition of Anyscale in late July 2026 — notable because Anyscale is already a listed APIRank provider for its Ray ecosystem and serverless serving. The acquisition plus the credit facility indicates Nscale intends to become a top-tier EU-sovereign alternative to US hyperscale clouds.

Where Nscale fits

Nscale is the strongest pick when:

You need EU data sovereignty: open models running on UK/Norway GPUs with in-country processing is hard to match from US-only vendors.
You want OpenAI-compatible open-model serving: Qwen3, GPT OSS, Llama 4, Kimi, and Mistral behind one /v1 drop-in, comparable to Fireworks or Together but with owned infrastructure.
You want a single governed platform: serverless + on-demand GPUs + fine-tuning + workbench under one account and one credit balance, instead of stitching vendors.

Nscale is a weaker pick when:

You need public, verifiable pricing: the lack of an open pricing page makes budgeting and benchmarking harder than OpenAI/Anthropic/OpenRouter.
You need proprietary frontier models: no GPT-5, no Claude — open weights only.
You serve mainland China: no China-region endpoint; expect 200-350ms proxied latency.
You want zero-friction signup: Google SSO only, prepaid credit before first call, no permanent free tier.

Bottom line

Nscale is a credible, EU-sovereign full-stack AI cloud whose serverless inference API is genuinely OpenAI-compatible and serves a solid open-model catalog (Qwen3, GPT OSS, Llama 4, Kimi K2.5, Mistral, FLUX.1, SDXL — all verified on nscale.com/product/inference, captured 2026-08-27). Pricing is prepaid and credit-based, structured per 1M input/output tokens for text and per million output pixels for images (verified against the public serverless-openapi.yaml), with per-model rates in the Console and authenticated /v1/models. The trade-offs are real but coherent: no public pricing page, open models only, no China endpoint, Google-SSO signup, and no free tier. The authoritative sources are the AI services page, the quickstart, the docs index, and the newsroom for the Anyscale and $900M updates. For a team that values sovereign open-model inference and is happy to manage credit in the Console, Nscale is a genuinely differentiated pick.

If you are routing Nscale alongside OpenAI, Anthropic, or Groq behind a single OpenAI-compatible key with cross-region failover, FreeModel is the simplest handoff: one dashboard, one billing relationship, and routing controls without glue code.