Seventy-One Labs, Seventy-One Different Futures

I started researching AI labs because I wanted a list. Not another ranking of whichever chatbot scored highest this week, but a more literal map of the organizations actually building models. Who are they, where are they, how many people work there, what are they trying to make, and what does each of them seem to believe AI is for?

The list reached seventy-one before the category started breaking apart.

Seventy-one is not the number of AI labs in the world. There are probably hundreds, depending on how generously the word *lab* is used. It was simply the point at which I had enough rows in the spreadsheet to notice that I was no longer comparing similar organizations. One company was building a chatbot. Another was trying to automate scientific research. Another was training humanoid robots. Another existed partly so that a country would not have to depend entirely on American technology. Another was generating songs. Another was designing drugs.

They all called themselves AI labs, or were described that way, but they did not appear to be moving toward the same future.

The word *lab* still carries a strangely innocent feeling. I imagine a few scientists in a room, some whiteboards, a half-disassembled machine and a project whose purpose is not yet clear. A lab produces experiments. It asks questions that a normal company might not have time to ask. Even when a lab is inside a large corporation, the word suggests that it remains slightly outside the machinery of quarterly results and product roadmaps.

That image no longer fits very well.

Stanford's 2026 AI Index found that industry produced more than 90 percent of notable AI models in 2025. It also noted that the most capable systems are becoming the least transparent, with important details about training code, datasets, parameter counts and training duration increasingly undisclosed.

This changes what a lab is. It is not just a group of researchers anymore. It may also need data centers, chip agreements, enterprise sales teams, policy departments, consumer applications, safety researchers, lobbyists, lawyers, content partnerships and people negotiating electricity supply. Epoch AI's company dataset now tracks frontier companies through funding, staff counts, compute spending, revenue and product usage, the same categories we might once have used to compare airlines or industrial conglomerates.

The closer I looked, the more the organizations seemed to be dividing reality between themselves.

OpenAI and Anthropic are trying to capture a large part of knowledge work. Their models are increasingly expected not only to answer questions but to write software, read documents, perform research, operate tools and continue working without constant instruction. They are building something that sits between a colleague, a computer interface and an externalized layer of thought.

Google DeepMind is harder to place because it seems to be building several futures at once. Its public work stretches from general models to robotics, weather forecasting, biology and scientific discovery. It can present Gemini as a consumer assistant while also using AI to model proteins, materials and physical action. Google DeepMind describes the same organization as working to advance science, transform work, serve diverse communities and improve billions of lives. That is less like a software product than an attempt to become part of the infrastructure through which future discoveries are made.

Then there are the labs that are not primarily trying to own knowledge work. Figure's Helix is described as a generalist vision-language-action model for humanoid control. The model is supposed to connect language and perception to a physical body, allowing a machine to interpret instructions and manipulate objects. The future implied here is not primarily that I will ask a chatbot to help me write an email. It is that the same underlying approach may eventually move boxes, assemble objects, clean rooms or perform other forms of physical labor.

That is a very different political and economic project, even though it may use some of the same research vocabulary.

The scientific labs are different again. Isomorphic Labs is building an AI drug-design engine intended to work across therapeutic areas and different kinds of drugs. Its interface is aimed at scientists, while its models are supposed to predict biological behavior and help design molecules. Here the imagined future is not primarily conversational. It is a world in which parts of scientific intuition and experimentation become computational, and where the institution controlling the model may influence which diseases, molecules and commercial opportunities receive attention.

I kept thinking about how misleading it is to place these organizations in the same leaderboard. A language model can be measured on mathematics or coding questions. A robot model has to survive contact with a kitchen. A drug-design system eventually has to survive biology. A regional language model might be technically weaker in English and still be far more important to the people it was designed for.

The ranking changes depending on what kind of future you value.

Mistral is often treated as Europe's answer to the large American labs, but its importance is not only whether its latest model defeats another model on a benchmark. The company presents itself as a European frontier lab focused on openness, transparency, cost efficiency, deployment control and technological independence. Its models can be accessed through large cloud services or deployed on infrastructure controlled by the organization using them. Its future is one in which European institutions can use powerful AI without handing every document, workflow and dependency to a handful of companies elsewhere.

Sarvam describes itself even more directly as India's full-stack sovereign AI platform, building models, infrastructure and applications intended to work for India. AI Singapore's SEA-LION project trains models around Southeast Asian languages and cultural contexts that are often weakly represented in global systems. These are not just smaller competitors hoping to become the next OpenAI. They are responses to the possibility that language, culture and public infrastructure might otherwise be interpreted through models built somewhere else.

In that sense, some AI labs resemble national institutions even when they are private companies. They are not only selling intelligence. They are deciding which languages will be understood well, which laws will be easy to encode, which histories are present in the training data and which forms of knowledge remain peripheral.

The spreadsheet made this visible in an uncomfortable way. The American labs tended to describe a universal future. Their mission was intelligence for everyone, scientific progress for humanity, or systems capable of performing most economically valuable work. Regional labs were more likely to name the territory directly: European AI, Indian AI, Arabic AI, Southeast Asian languages, local deployment, national infrastructure.

One side speaks as if intelligence can be universal. The other has already noticed that someone will own the supposedly universal system.

There were other divisions. Open-weight labs imagine a future in which developers can inspect, adapt and run models themselves. Closed labs argue, sometimes explicitly and sometimes through their products, that the most capable systems require centralized control. Enterprise labs want AI to live inside private company data and regulated workflows. Consumer labs want it inside the most intimate parts of everyday life. Voice labs want it to speak. Music labs want it to compose. World-model labs want it to simulate space. Robotics labs want it to act. Scientific labs want it to discover.

The organizations may share researchers, papers, chips and investors, but their products imply different relationships between people and machines.

Even the headcounts were revealing, although the numbers were often difficult to verify. Some organizations valued in the billions were still relatively small teams. Others were divisions inside corporations with hundreds of thousands of employees, making the boundary of the *lab* almost impossible to draw. Does Google DeepMind include every engineer who puts Gemini into a Google product? Does Meta's AI lab include the recommendation systems that already shape what billions of people see? Is Microsoft an AI lab, an investor in AI labs, a distributor of their models, or all three at once?

The category begins to feel less like a description and more like camouflage.

Calling something a lab allows it to retain the cultural status of research while accumulating the power of infrastructure. A lab can say it is exploring what is possible. A platform has to admit that other people may become dependent on it. A pharmaceutical company has to answer questions about access and pricing. A labor-automation company has to discuss workers. A defense contractor has to discuss conflict. A utility has to discuss energy and public resources.

An AI lab can move between these identities without settling into any of them.

This does not mean the research is fake. Many of these organizations are doing remarkable work, and some of it would be difficult to fund through universities or governments alone. But the language matters because it shapes the questions we ask. We ask whether a model is intelligent, open or safe. We ask much less often what kind of institution the lab is becoming, what dependencies it creates, or what part of life it is quietly positioning itself to mediate.

After seventy-one rows, I no longer saw one AI race. I saw dozens of bets on what the world will need and who should be allowed to provide it.

One lab believes the future belongs to the best general model. Another believes it belongs to the company that controls the interface. Another thinks it belongs to open weights. Another is betting on robots, voices, proteins, cars, national languages or private corporate data. Some may disappear. Some may be acquired. Some will probably become features inside larger platforms. A few may become institutions that are difficult to imagine living without.

The list was supposed to help me understand who was ahead.

Instead, it made me wonder why so many radically different futures are being built under the same harmless little word.

AI authorship note

This essay was developed from Kris Haamer's editorial direction and a broader research map of global AI labs. AI supported source gathering, comparison, drafting and implementation. Kris is responsible for the framing, argument, source selection and decision to publish it.

Connected reading

• The Open Models and the Rest of the Work
• The AI Future Is Becoming a Land Use Problem
• The Product Is Becoming an Organization
• The World Runs on Companies You Have Never Heard Of