# The United States of Inference > There are so many inference providers. How do you tell them apart? A map that gives each of 50 AI inference providers the US state it is most like (not where it is based), with what makes each one different. Made by Subconscious (https://subconscious.dev). Each state is matched to one AI inference provider, with a one-line reason and the things that set that provider apart. The match is about character, not location: a provider gets the state it is most like (its ethos, what it is known for, what it does), not the state where it is headquartered. Link to a state with https://subconscious-systems.github.io/united-states-of-inference/#, for example https://subconscious-systems.github.io/united-states-of-inference/#texas. ## States ### Alabama: Moonshot AI Huntsville engineers designed the Saturn V that carried Americans to the moon, which makes Rocket City the obvious home for a lab named after the trip. Beijing lab behind Kimi and the open-weight Kimi K models. - Kimi K3 is an open-weight ~2.8T-param MoE, among the largest open models you can download - 1M-token context, aimed at long documents and agent work - Kimi models are a common pick for agentic coding on other hosts too Website: https://www.moonshot.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#alabama ### Alaska: Crusoe Alaska is oil rigs and off-grid cabins at the end of the road, and Crusoe got its start running GPUs on gas that oil wells used to burn off. Vertically integrated AI cloud, from energy to managed inference. - Owns the whole stack: power generation, data centers, GPUs and the inference API - Builds next to stranded and renewable energy, including the 1.2 GW Abilene Stargate campus - Managed Inference sits on capacity it controls end to end, not rented GPUs Website: https://www.crusoe.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#alaska ### Arizona: General Compute The desert once known for cactus now builds chip fabs, and General Compute runs inference on custom ASICs with no GPU in sight. Inference cloud running on ASICs instead of GPUs. - Serves models on SambaNova ASICs instead of GPUs - Pitches speed and cost that come from the chip, not from GPU tuning Website: https://www.generalcompute.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#arizona ### Arkansas: DeepInfra Walmart started in Arkansas on the promise of everyday low prices, and DeepInfra brings that energy to open models with a giant catalog and no minimums. Low-cost serverless API for open-source models. - Among the lowest per-token prices for popular open models - Huge catalog: text, embeddings, image and speech models behind one OpenAI-compatible API - Simple pay-as-you-go, no commitments Website: https://deepinfra.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#arkansas ### California: OpenAI Most of what feels normal today got tried in California first. ChatGPT is how most of the world met AI. Frontier lab behind the GPT models, ChatGPT and the OpenAI API. - Serves its own frontier models: GPT, reasoning models, image, audio and realtime voice - The OpenAI API format is the one most other providers copy - The only first-party home for the newest GPT models Website: https://openai.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#california ### Colorado: Baseten Every fourteener climb starts at base camp, and so does a lot of AI headed for production. Inference platform for serving open-source and custom models. - Built for production: dedicated deployments of your own or open models with autoscaling - Truss packaging and its own inference stack tuned per model - Runs across multiple clouds so capacity follows demand Website: https://www.baseten.co On the map: https://subconscious-systems.github.io/united-states-of-inference/#colorado ### Connecticut: Together AI Hartford got rich getting everyone to pool their risk. Together started out as a decentralized cloud pooling GPUs. AI native cloud for open-model inference, fine-tuning and GPUs. - Research-driven: home of FlashAttention, with Tri Dao as chief scientist - Serverless, dedicated, fine-tuning and raw GPU clusters in one place - A broad catalog of open models, often added on release day Website: https://www.together.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#connecticut ### Delaware: Modal Over a million companies call Delaware home without ever setting foot there. Modal runs your code on GPUs without you ever setting foot on a server. Serverless GPU cloud for running AI inference and sandboxes. - Run arbitrary Python on GPUs, not just a fixed model menu - Containers cold-start in seconds and scale to zero - Sandboxes for running agent-generated code are a big part of the business Website: https://modal.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#delaware ### Florida: Meta Nobody can predict Florida Man or Meta: Llama one year, Muse the next, a retired API in between. Your grandma probably thinks Muse is cute. Meta Superintelligence Labs serving Muse models; maker of Llama. - Made the Llama family, which kicked off the open-weight model wave - Now serves Muse models from Meta Superintelligence Labs via the Meta Model API - Muse Spark Open is an 816B-param open-weight MoE (42B active) Website: https://www.meta.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#florida ### Georgia: OpenRouter Atlanta’s Hartsfield-Jackson is the busiest airport in the country, and OpenRouter is where the AI industry changes planes. Unified API routing requests across hundreds of LLMs. - One API and one bill for hundreds of models from dozens of providers - Routes each request to a provider by price, speed or uptime, with automatic fallback - Its public usage rankings are a de facto leaderboard for what developers actually use Website: https://openrouter.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#georgia ### Hawaii: Dolphin Hawaii’s spinner dolphins swim wherever they want. Dolphin’s models say whatever they want, running on home GPUs scattered across a network. Uncensored Dolphin models served on a network of distributed consumer GPUs. - Serves the uncensored Dolphin models - Runs on a decentralized network of idle consumer GPUs instead of data centers Website: https://dphn.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#hawaii ### Idaho: Makora Boise is home to Micron, which makes the memory stacked inside today’s GPUs. Makora’s agents rewrite the code running on those GPUs until it’s up to 5x faster. Inference endpoints tuned by AI agents, down to the GPU kernels. - Agents optimize every layer: routing, speculative decoding, vLLM/SGLang tuning and custom CUDA kernels - Claims the highest tokens per second per user, up to 5x faster than other providers - Runs on NVIDIA, AMD, Google TPU, AWS Trainium and more, with no hardware lock-in Website: https://www.makora.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#idaho ### Illinois: Vercel AI Gateway Every trip across America seems to connect through O’Hare. Every model call can connect through Vercel. Vercel's single API gateway to hundreds of AI models. - A gateway, not a host: one key for hundreds of models across providers - No markup on tokens, even with your own keys - Built into the Vercel AI SDK, with automatic failover between providers Website: https://vercel.com/ai-gateway On the map: https://subconscious-systems.github.io/united-states-of-inference/#illinois ### Indiana: Groq Built for tight, predictable latency, Groq belongs at the Indy 500, where the winning car holds its line for 500 miles. Inference neocloud built on LPU and NVIDIA hardware. - Runs models on its own LPU chips, built for low-latency token generation - Known for very high tokens per second on open models - Licensed its tech to Nvidia in a ~$20B deal and runs 13 data centers Website: https://groq.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#indiana ### Iowa: Wafer Iowa farmers brag about bushels per acre, and Wafer brags about tokens per second on the same weights everyone else runs. Fast inference for open models. - Offers a flat-rate plan (Wafer Pass) as well as per-token pricing - Fast open-model inference on NVIDIA B300s Website: https://www.wafer.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#iowa ### Kansas: Inception A Kansas tornado turns a whole farm into noise in seconds. Mercury runs it in reverse: noise in, a whole answer out. Diffusion LLMs (Mercury) generating tokens in parallel. - Diffusion LLMs: Mercury writes many tokens in parallel instead of one at a time - That makes it several times faster than same-size autoregressive models - Mercury was the first commercial diffusion LLM Website: https://www.inceptionlabs.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#kansas ### Kentucky: CoreWeave Fort Knox guards the nation’s gold reserve, and CoreWeave guards the GPU reserve the biggest AI labs rent from. AI cloud built on NVIDIA GPU clusters, with open-model serving through W&B Inference. - The GPU cloud the big labs rent: huge NVIDIA clusters built for AI - Serves open models through W&B Inference, from the Weights & Biases acquisition - Training, tracking and inference in one place Website: https://www.coreweave.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#kentucky ### Louisiana: Mistral Louisiana still speaks a little French, and so does Europe’s frontier lab. Paris lab building open-weight models and sovereign AI infrastructure. - Europe's leading lab, with open weights under Apache 2.0 - Mistral Large 3 is a 675B-param open MoE - Sells sovereign deployments for governments and companies that need data to stay local Website: https://mistral.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#louisiana ### Maine: Thinking Machines Mainers fix their own boats, plows and roofs. Thinking Machines built Tinker so you can fix up your own model without touching a GPU. Mira Murati’s research lab; Tinker API for fine-tuning and serving open models. - Fine-tune open models with Tinker, then run inference on them in the same place - You write the training loop, Tinker handles the distributed GPUs - Research lab from former OpenAI CTO Mira Murati Website: https://thinkingmachines.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#maine ### Maryland: Cloudflare Workers AI The NSA lives in Maryland. So does the company that sits in front of a big chunk of the internet. Serverless GPU inference on Cloudflare's global edge network. - Inference runs on GPUs spread across 180+ cities, close to users - Called from Workers, so the model sits next to your app logic - Pairs with Cloudflare's AI Gateway and Vectorize for caching and retrieval Website: https://www.cloudflare.com/developer-platform/products/workers-ai/ On the map: https://subconscious-systems.github.io/united-states-of-inference/#maryland ### Massachusetts: Subconscious Home of the oldest modern marathon, the Boston Marathon. Subconscious builds Marathon models that keep agents going for hours. Inference for long-running agents with 200k+ token sessions. - Excellent for coding and cybersecurity tasks - Marathon versions of GLM-5.3 and DeepSeek V4.1 tuned for long agent runs Website: https://www.subconscious.dev On the map: https://subconscious-systems.github.io/united-states-of-inference/#massachusetts ### Michigan: Morph Michigan gave the world the moving assembly line, where each specialist does one job at blistering speed, and Morph applies that idea to code edits. Fast Apply code-edit models and inference for coding agents. - Specialized models for coding agents, not general chat - Fast Apply merges LLM code edits at 4,500+ tok/s - Also offers WarpGrep code search and Compact context tools Website: https://morphllm.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#michigan ### Minnesota: Cohere Minnesota is where Target, 3M and General Mills keep their headquarters, a polite drive from Canada. Cohere is the Toronto lab that builds AI for companies exactly like them. Enterprise and sovereign AI models (Command) and the North platform. - Enterprise-first: models built for retrieval, tool use and business data - Deploys privately in your cloud or on-prem, including sovereign setups - Its Embed and Rerank models are widely used in search pipelines Website: https://cohere.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#minnesota ### Mississippi: StepFun The Ole Miss Rebels love picking fights with bigger SEC programs. StepFun is a startup picking fights with China’s tech giants. Shanghai lab building the Step family of foundation models. - Step models from a Shanghai lab, strong on multimodal and reasoning - Served through the StepFun Open Platform Website: https://stepfun.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#mississippi ### Missouri: Infron St. Louis has the Gateway to the West. Infron is a gateway to 400+ models. Unified API gateway to hundreds of AI models. - Provider routing picks which host serves each request - Passes through provider pricing with no markup, and supports your own keys Website: https://infron.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#missouri ### Montana: Armada Montana has ranches a hundred miles from the nearest data center. Armada ships the data center to the ranch. Containerized edge data centers for AI where data is generated. - Inference at the edge: containerized data centers shipped to where data is created - Built for remote sites like energy fields, defense and mining with little or no connectivity Website: https://www.armada.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#montana ### Nebraska: Parasail Warren Buffett got rich in Omaha buying what everyone else overlooked, and Parasail turns overlooked GPU capacity into cheap batch inference. Pay-per-token inference cloud for open models. - Owns no GPUs: rents across providers and passes on the savings - Pay-per-token open models plus dedicated deployments - Founded by Groq and Mythic veterans Website: https://www.parasail.io On the map: https://subconscious-systems.github.io/united-states-of-inference/#nebraska ### Nevada: Venice What happens in Vegas stays in Vegas, and Venice runs uncensored models without keeping your conversations. Private, uncensored AI platform on open-source models. - Privacy-first: no stored conversations - Uncensored open models for text, images and video - Stake the VVV token to get API access instead of paying per token Website: https://venice.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#nevada ### New Hampshire: Refuseless New Hampshire stamps Live Free or Die on every license plate, and Refuseless builds models that follow the motto. API for uncensored, abliterated open models. - Abliterated open models with refusals removed - Drop-in OpenAI-compatible API, supported in LiteLLM Website: https://refuseless.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#new-hampshire ### New Jersey: ElevenLabs Edison recorded the first human voice in Menlo Park. ElevenLabs can now fake yours. Voice AI: text-to-speech, speech-to-text and voice agents. - Voice, not text: lifelike text-to-speech, speech-to-text and voice agents - 17,500+ voices in 90+ languages, plus voice cloning - A top choice for realtime voice agents Website: https://elevenlabs.io On the map: https://subconscious-systems.github.io/united-states-of-inference/#new-jersey ### New Mexico: fal Santa Fe is one of the biggest art markets in the country and Albuquerque keeps film crews busy, a fitting home for fal and its thousand-plus generative media models. Inference platform for generative media: image, video and audio models. - Generative media only: images, video, audio and 3D - 1,000+ media models, often available the day they launch - Its own inference engine tuned for diffusion models Website: https://fal.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#new-mexico ### New York: Anthropic New York is the business capital of the world, plus millions of regular people upstate. Anthropic dominates enterprise AI and still serves millions of individuals with Claude. AI safety lab building the Claude model family. - Serves the Claude models: Opus, Sonnet and Haiku - Known for coding and agentic work; Claude Code runs on it - Also available through Amazon Bedrock and Google Vertex AI Website: https://www.anthropic.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#new-york ### North Carolina: SambaNova NASCAR country loves heavy machines moving at absurd speed, and SambaNova pushes giant open models fast on its own dataflow chips. Fast inference on custom RDU chips. - Runs models on its own RDU dataflow chips, not GPUs - Fits very large models like full DeepSeek on few machines - Fast open-model inference through SambaNova Cloud Website: https://sambanova.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#north-carolina ### North Dakota: Boundless The prairie runs to the horizon in every direction, and Boundless tunes reserved capacity so you never hit the edge. Reserved, tuned inference capacity for open models. - Reserved, tuned capacity for open models rather than shared serverless - Runs on a ~4,000-GPU distributed network repurposed from zero-knowledge proving Website: https://boundless.network On the map: https://subconscious-systems.github.io/united-states-of-inference/#north-dakota ### Ohio: RunInfra The Wright brothers tested 200 wing designs in their Dayton bike shop before they ever left the ground. RunInfra tests your serving setup before it ships. Agent that benchmarks and deploys open-model inference stacks. - An agent picks the stack: it benchmarks engines and configs, then deploys the winner - OpenAI- and Anthropic-compatible APIs with simple $1 credits Website: https://runinfra.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#ohio ### Oklahoma: Particle.AI Oklahoma City invented the shopping cart so you could haul more for less effort. Particle hauls a million tokens for a quarter. Low-cost inference for open-weight models with 1M-token context. - Particle-1, post-trained from DeepSeek V4.1 Flash, with 1M-token context - Cheap: $0.25/M input and $1.00/M output Website: https://particle.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#oklahoma ### Oregon: Hugging Face Portland’s motto is Keep Portland Weird, and Hugging Face is the open-source town square where AI’s weirdest projects live. Open-source model hub with routed inference providers. - Home of open models: the Hub hosts nearly every open-weight release - Inference Providers routes one API to many partner hosts - Pick a model on the Hub and run it there in a click Website: https://huggingface.co On the map: https://subconscious-systems.github.io/united-states-of-inference/#oregon ### Pennsylvania: Fireworks AI Philadelphia lit up the first Fourth of July with fireworks in 1777. They’ve been the main event ever since. Inference and fine-tuning platform for open models. - Built by the team behind PyTorch at Meta - Its own serving engine for fast open-model inference - Fine-tune, then serve the result on the same platform Website: https://fireworks.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#pennsylvania ### Rhode Island: Sail Research Newport is America’s sailing capital, and Sail lets you pick a completion window and pay up to 85% less for waiting on the wind. Open-model API with completion windows that trade latency for lower prices. - Completion windows: wait longer, pay less - Built for batch and latency-tolerant work like evals and data pipelines Website: https://www.sailresearch.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#rhode-island ### South Carolina: Relace South Carolina’s beaches exist for relaxing, and Relace keeps inference cheap enough that you can. Open-model inference and apply models for coding agents. - Models made for coding agents: apply, embeddings and reranking - Turns an LLM's edit into a merged file quickly and reliably Website: https://relace.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#south-carolina ### South Dakota: Luminal Mount Rushmore carved four presidents straight into granite, and Luminal compiles models straight into native GPU code. ML compiler that turns models into optimized native GPU and ASIC code. - A compiler, not just a host: turns models into optimized native GPU and ASIC code - Open source, written in Rust (Apache 2.0) Website: https://luminal.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#south-dakota ### Tennessee: Suno Nashville has a songwriter on every barstool. Suno has one in every browser tab. AI music generation from text prompts. - Music, not text: full songs with vocals from a prompt - The best-known AI music generator Website: https://suno.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#tennessee ### Texas: xAI Everything is bigger in Texas, opinions included, and xAI is the lab that does things its own way and dares you to object. AI lab behind the Grok models, now part of SpaceX. - Serves the Grok models, with live data from X - Trains on Colossus in Memphis, one of the largest GPU clusters - Part of SpaceX since February 2026 Website: https://x.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#texas ### Utah: Cerebras Land speed records get set on the Bonneville Salt Flats, and Cerebras posts the fastest published output speed of any public host. Wafer-scale chips powering very fast inference. - Wafer-scale chips: one processor the size of a whole wafer - Among the fastest tokens per second for open models - Keeps model weights in on-chip memory, avoiding the GPU memory bottleneck Website: https://www.cerebras.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#utah ### Vermont: GMI Cloud Vermonters grow their own food and fix their own trucks. GMI Cloud owns the GPUs it runs on. GPU cloud offering serverless inference and dedicated clusters. - Owns GPU infrastructure in the US and Asia-Pacific, a rare mix - Same provider for serverless inference and reserved clusters Website: https://www.gmicloud.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#vermont ### Virginia: Amazon Bedrock Virginia is for lovers. It’s also for defense contractors and US-East-1. Amazon Bedrock embodies this great state. AWS managed service for foundation models and agents. - Models from many labs inside your AWS account and its security rules - Claude, Llama, Mistral and Amazon Nova behind one AWS API - Adds Guardrails, Knowledge Bases and AgentCore for enterprise apps Website: https://aws.amazon.com/bedrock/ On the map: https://subconscious-systems.github.io/united-states-of-inference/#virginia ### Washington: Google Seattle grows tech giants the way it grows moss. Google wrote the Transformer paper that every model on this map is built on. Google's Gemini API and Vertex AI model serving. - Runs on its own TPUs, not NVIDIA GPUs - Serves Gemini and the open Gemma models - Long context, plus native image, audio and video input Website: https://ai.google.dev On the map: https://subconscious-systems.github.io/united-states-of-inference/#washington ### West Virginia: TypeSafe AI Jevons noticed that making coal cheaper to burn only made the world burn more of it, and West Virginia dug plenty. TypeSafe named its first model Jev, making the same bet on decisions. Decision models for software that report calibrated confidence. - Decision models that report calibrated confidence, not just answers - Built for software that needs to know when to trust a model's call Website: https://typesafe.ai On the map: https://subconscious-systems.github.io/united-states-of-inference/#west-virginia ### Wisconsin: Nebius Wisconsin was built by European immigrants who showed up and got to work, and Nebius is the Amsterdam neocloud doing the same in American AI. Full-stack AI cloud with GPU clusters and Token Factory inference. - Designs its own servers and runs its own data centers in Europe and the US - Listed on Nasdaq (NBIS) after spinning out of Yandex - Serve open models or your own fine-tunes on dedicated capacity Website: https://nebius.com On the map: https://subconscious-systems.github.io/united-states-of-inference/#wisconsin ### Wyoming: Inference.net With the smallest population in the country, Wyoming has room to spare, and Inference.net runs batch jobs on spare GPU capacity. Inference infrastructure and custom model training for AI-native teams. - Cheap batch inference on spare GPU capacity - Trains custom models for AI-native teams, then serves them - Started as Kuzco, a distributed GPU network Website: https://inference.net On the map: https://subconscious-systems.github.io/united-states-of-inference/#wyoming