Frontier AI companies

This page is an overview of the companies that lead the AI industry and the products they sell. It describes who does what, not which model is best this month. Model names and version numbers change every few months; the companies and their roles change much more slowly.

What “frontier” means

A frontier model is a general-purpose model at or near the limit of what current AI can do. The term became common in 2023, when Anthropic, Google, Microsoft, and OpenAI founded the Frontier Model Forum, and governments met at the first AI Safety Summit at Bletchley Park in the United Kingdom.

Training a frontier model needs tens of thousands of specialized chips, large amounts of data, and investments in the billions of dollars. That is why only a few companies compete at the top. Industry reports in 2026 describe the very front as a race between three labs: OpenAI, Anthropic, and Google DeepMind. A wider group follows closely, and the lead changes with almost every major release.

The layers of the AI industry

The industry works like a stack. Each layer depends on the one below it.

flowchart TB
  apps["Applications: chat assistants, coding tools, agents, AI features in existing software"]
  models["Model labs: train and serve foundation models"]
  cloud["Cloud platforms: data centers that rent out computing power"]
  chips["Chips: GPUs and AI accelerators, and the factories that make them"]
  apps --> models --> cloud --> chips

Many companies span several layers. Google designs its own chips, runs a cloud, trains Gemini, and puts AI into Search and Android. OpenAI trains models and sells ChatGPT, but rents most of its computing power. The sections below follow the stack from the model labs down to the chips.

Frontier model labs

OpenAI

Founded in 2015 in San Francisco as a non-profit research lab, OpenAI started the current wave of AI with the public release of ChatGPT in November 2022. It later added a for-profit structure to raise capital; Microsoft has been its largest investor since 2019.

Main products:

• ChatGPT, the most widely used AI assistant, for consumers and businesses.
• GPT models, including reasoning models, sold through the OpenAI API.
• Codex, a coding agent.
• Sora for video and built-in image generation.
• Open-weight models (gpt-oss) and the open speech recognition model Whisper.

OpenAI also leads Stargate, a large data center project announced in 2025 with partners including SoftBank and Oracle.

Anthropic

Former OpenAI researchers, among them Dario and Daniela Amodei, founded Anthropic in 2021. It is a public benefit corporation with a strong focus on AI safety research. Amazon and Google are major investors.

Main products:

• Claude models in several sizes, from fast and cheap to most capable.
• Claude apps for web, desktop, and mobile.
• Claude Code, an agentic coding tool that runs in the terminal and in editors.
• The Claude API, also available through Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure.

Anthropic published the Model Context Protocol in 2024, now a common standard for connecting agents to tools. Its models are widely used for programming and for business work.

Google DeepMind

DeepMind was founded in London in 2010 and acquired by Google in 2014. In 2023, it merged with Google Brain, the team behind the Transformer architecture. Google is the only company with a strong position in every layer of the stack.

Main products:

• Gemini, both the model family and the assistant app that replaced Bard in 2024.
• Gemma, open-weight models that run on smaller hardware.
• Google AI Studio and Vertex AI for developers and companies.
• AI features in Search, Android, Gmail, and Workspace, and research tools such as NotebookLM.
• Veo for video and Imagen for images.
• TPUs, Google’s own AI chips.

DeepMind is also known for scientific AI. AlphaFold predicts protein structures, and its creators Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for it.

Meta

Meta has run an AI research lab (FAIR) since 2013 and created PyTorch, the most widely used framework for building neural networks. From 2023, its Llama models made Meta the leading Western publisher of open-weight models.

In 2025, Meta regrouped its AI work into Meta Superintelligence Labs and invested heavily in new data centers. In 2026, it released Muse, its first family of closed-weight models. Meta’s main channel is distribution: the Meta AI assistant is built into WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban Meta glasses.

xAI

Elon Musk founded xAI in 2023. Its Grok models power a chatbot built into the social network X and available as a standalone app and API. xAI built one of the largest AI training clusters, called Colossus, in Memphis. In February 2026, SpaceX acquired xAI, bringing Grok, X, and Starlink under one company.

Open-weight challengers

A second group of labs publishes many of its models with open weights, so anyone can download and run them. Their models are usually a few months behind the frontier and far cheaper to use. Many developers and companies use them for cost reasons or to keep data on their own servers.

Mistral AI

Mistral was founded in Paris in 2023 by former researchers from Google DeepMind and Meta. It is Europe’s best-known model lab. It releases both open-weight and commercial models, offers the assistant Le Chat, and markets itself to European companies and governments that want to keep data and control within Europe.

DeepSeek

DeepSeek is a Chinese lab founded in 2023 in Hangzhou and funded by the hedge fund High-Flyer. In January 2025, its open-weight reasoning model DeepSeek-R1 matched leading American models at a fraction of the reported training cost. The release caused a sharp one-day drop in Nvidia’s share price, and it showed that export limits on chips had not stopped Chinese labs from reaching the frontier.

Alibaba and other Chinese labs

Alibaba’s Qwen family is among the most downloaded open-weight models, with sizes from small models for phones to very large ones. Other Chinese labs with strong open models include Moonshot AI (Kimi), Zhipu AI (GLM), and MiniMax. ByteDance (Doubao) and Baidu (Ernie) run large consumer assistants in China.

Platforms and distribution

Most people and companies do not use models directly from a lab. They reach them through cloud platforms and products they already use.

Microsoft sells AI under the Copilot brand: in Windows, in Microsoft 365, and as GitHub Copilot for developers. Its Azure cloud hosts models from OpenAI and other labs, and Microsoft has started training its own MAI models.

Amazon offers models from many labs through Amazon Bedrock, trains its own Nova models, designs AI chips called Trainium, and has rebuilt its voice assistant as Alexa+.

Apple integrates AI into its devices as Apple Intelligence, with a focus on processing data on the device, and works with outside model providers for more demanding requests.

Hugging Face is the central platform where open-weight models and datasets are published, much as GitHub is for source code.

Chips and compute

Every model runs on specialized hardware, and that hardware is scarce.

Nvidia dominates AI chips. Its GPUs train most frontier models, and its CUDA software, available since 2007, keeps developers on its platform. AI demand made Nvidia one of the most valuable companies in the world.

Others compete for parts of that market. AMD sells competing GPUs, while Google (TPU) and Amazon (Trainium) design chips for their own clouds. Almost all of these chips are manufactured by TSMC in Taiwan, which makes chip production a geopolitical issue. Export controls by the United States limit which chips can be sold to China.

Products at a glance

How to follow a fast-moving field

Rankings at the top change with every major release, often within weeks. A few habits keep the picture accurate without following every announcement:

• Think in model families, not versions. GPT, Claude, Gemini, Llama, and Qwen will each have a new version soon. The family and the company behind it say more about pricing, licensing, and data handling than the version number does.
• Check independent comparisons. Leaderboards such as LMArena, which ranks models by human preference, and Artificial Analysis, which compares quality, speed, and price, are more reliable than a vendor’s own benchmarks. Evaluating AI models explains what they measure.
• Read the model card. Providers publish a model card or system card with each major model. It lists the context window, intended uses, known limits, and safety tests.
• Test on your own task. A model that leads a general benchmark is not always the best for a specific job. A small evaluation set from your own work is the most useful comparison.