Token APIs and inference endpoints, as the operator publishes them. China and the EU stay on the list even when the city is sales-gated. This is not a bare-metal catalog, and it is not a market cap. 21 providers · 33 location rows · 14 named cities plotted · 0 state/region-only · 19 undisclosed. China: 7 operators. Europe: 3 operators. Sibling of the Silicon Tape and the Compute Net Worth Index™.
Country conversion signals and sourced silicon prints. Labeled terms only — no invented deltas. Read it on the public brief, or take the RSS.
RSS: /brief.xml · /silicon.xml · /inference.xml · /neoclouds.xml · /hyperscalers.xml · /wire.xml
Named cities only — Wikipedia centroids where the official page names a city. State-only and undisclosed rows are not plotted as fake points.
| Name ▼ | Kind | HQ | Regions | Named cities | Models / silicon | Source |
|---|---|---|---|---|---|---|
| Groq / GroqCloudgroq | inference | Mountain View · United States | 11 | 11 named | Groq LPU, NVIDIA Groq 3 LPX announced, Llama and other open models via GroqCloud — catalog rotates | source |
| Cerebrascerebras | inference | Sunnyvale · United States | 1 | undisclosed | Cerebras WSE, Llama, Qwen and others via Cerebras Inference — catalog rotates | source |
| Together AItogether | inference | San Francisco · United States | 1 | undisclosed | NVIDIA GPUs (dedicated-endpoint hardware flags), Large open-model catalog — see models page | source |
| Fireworks AIfireworks | inference | Redwood City · United States | 1 | undisclosed | NVIDIA GPUs, Open + custom; serverless / on-demand / reserved | source |
| OpenAIopenai | inference | San Francisco · United States | 1 | undisclosed | Undisclosed mix; partner neoclouds + Azure, GPT / o-series / embeddings — see models page | source |
| Anthropicanthropic | inference | San Francisco · United States | 1 | undisclosed | Undisclosed; AWS / GCP / CoreWeave / Fluidstack capacity, Claude Opus / Sonnet / Haiku | source |
| Google AI Studio / Vertex AIgoogle-ai-studio-vertex | inference | Mountain View · United States | 1 | undisclosed | TPU, NVIDIA GPUs on GCP, Gemini, Imagen, Veo, Model Garden | source |
| Amazon Bedrockamazon-bedrock | inference | Seattle · United States | 1 | undisclosed | Trainium/Inferentia + NVIDIA on AWS, Nova, Claude, Llama, Mistral, Cohere — marketplace | source |
| Azure OpenAI Serviceazure-openai | inference | Redmond · United States | 1 | undisclosed | Azure GPU/CPU mix, OpenAI models on Azure | source |
| Mistral AImistral | inference | Paris · France | 1 | undisclosed | Undisclosed; la Plateforme + partner clouds, Mistral Large/Medium/Small, Codestral, Pixtral, open-weight | source |
| Coherecohere | inference | Toronto · Canada | 1 | undisclosed | Undisclosed, Command, Embed, Rerank | source |
| SambaNovasambanova | inference | Palo Alto · United States | 1 | undisclosed | SambaNova SN40L / RDU, Llama and others on SambaCloud | source |
| DeepSeekdeepseek | inference | Hangzhou · China | 1 | undisclosed | Undisclosed China training cluster, DeepSeek-V3/V4/R1 families — names rotate | source |
| Moonshot / Kimimoonshot-kimi | inference | Beijing · China | 1 | undisclosed | Undisclosed, Kimi family | source |
| Zhipu / Z.ai (GLM)zhipu | inference | Beijing · China | 1 | undisclosed | Undisclosed, GLM family | source |
| Alibaba Qwen / DashScopealibaba-dashscope | inference | Hangzhou · China | 1 | undisclosed | Alibaba Cloud GPU/NPU, Qwen family | source |
| Baidu Qianfanbaidu-qianfan | inference | Beijing · China | 1 | undisclosed | Kunlun / Cloud GPU — English SKU page sales-gated, ERNIE / Wenxin + third-party | source |
| ByteDance / Volcengine Arkbytedance-volcengine-ark | inference | Beijing · China | 1 | undisclosed | Volcengine GPU fleet, Doubao / Seed + open models | source |
| Tencent Hunyuantencent-hunyuan | inference | Shenzhen · China | 1 | undisclosed | Tencent Cloud GPU, Hunyuan family | source |
| Scaleway Generative APIsscaleway-generative | inference | Paris · France | 3 | Paris, Amsterdam, Warsaw | Scaleway GPU instances in PAR/AMS/WAW, Managed open models — catalog on product page | source |
| Crusoe Managed / Serverless Inferencecrusoe-inference | inference | San Francisco · United States | 1 | undisclosed | H100, H200, B200, GB200 · +3 | source |
Click a row for the sourced location list. A missing city is undisclosed, never a guessed metro. Grain of the source: CoreWeave US rows stay state-only.
The question. Who sells tokens / APIs? Do not put here: Bare-metal clusters — except dual-list notes already in the catalog.
No invented cities. If a provider is real but the city list is sales-gated, the record exists with undisclosed. lat/lon only when a named city exists (Wikipedia centroid). State-only rows are not upgraded to a metro.
No market caps. No fake fleet counts. Marketing GW figures appear only if they already sit in the JSON notes.
China and the EU appear even when the city is undisclosed. Classification is already in the JSON — Scaleway and OVH are regional hyperscalers; DigitalOcean / Vultr stay neoclouds because the product we care about is GPU Droplets / Cloud GPU.
Machine-readable: inference.json (CC BY 4.0, attribution to compute.world). Cite as: Hamal, P. (2026). The Inference Index. compute.world.. Corrections: get in touch.
Who sells tokens / APIs? Do not put here: Bare-metal clusters — except dual-list notes already in the catalog.
No. A pin is a named city with a sourced lat/lon. Texas stays Texas. A sales-gated list stays undisclosed. China and the EU still appear as rows.
No. compute.world does not print a CoinMarketCap number on a cloud logo. These are sourced catalogs beside the Silicon Tape and the Compute Net Worth Index™.