The AI Labs

Models change constantly. New versions ship every few months, benchmarks get rewritten, and whatever was state-of-the-art in January is mid-tier by June. Keeping up with individual models is a losing game.

The labs building them, however, are more stable — and more revealing. Understanding who is developing this technology, why, and with whose money tells you more about where AI is headed than any benchmark chart.

The frontier three

A small number of organizations have the capital, compute, and research talent to train truly frontier models. Everyone else is working with what they build or release.

Anthropic was founded in 2021 by Dario Amodei, Daniela Amodei, and a group of researchers who left OpenAI specifically over disagreements about safety priorities. It's structured as a Public Benefit Corporation — a legal form that binds the company to a public interest mission, not just shareholder returns. Its core research investments include Constitutional AI (a technique for making models follow principles rather than just human approval signals), mechanistic interpretability (trying to actually understand what happens inside these models), and a published Responsible Scaling Policy — a written commitment to halt deployment if internal safety evaluations hit predefined thresholds. Primary funding: Amazon ($4B+) and Google ($2B+). Revenue comes from Claude.ai subscriptions and API access.

OpenAI was founded in 2015 as a nonprofit, explicitly to ensure that AGI benefits all of humanity. It converted to a "capped profit" structure in 2019 to attract capital, and has been converting to a conventional for-profit corporation since 2024. It built ChatGPT, the product that sparked the current wave, and remains the most widely recognized name in consumer AI. Its commercial model runs deep: ChatGPT subscriptions, API access, and a structural partnership with Microsoft that embeds OpenAI technology across Office, Bing, Azure, and GitHub. The tension between its founding mission and its commercial trajectory has become increasingly difficult to ignore — more on that below.

Google DeepMind was formed in 2023 from the merger of Google Brain and DeepMind. DeepMind — founded in London in 2010 and acquired by Google in 2014 — had built one of the most serious AI safety research programs in the world before the merger. The combined entity continues publishing safety-relevant research. Its commercial model is different from the others: Gemini models are woven directly into Google Search, Workspace, and Cloud, which means AI revenue flows through Alphabet's existing advertising and enterprise businesses rather than standing alone. That parent structure creates incentives that don't always align with moving slowly.

The other players

Meta AI has taken a strategically distinct path: releasing model weights publicly under the Llama family. Mark Zuckerberg has framed this as democratizing AI; the strategic logic is that open-sourcing powerful models commoditizes them — if everyone has access to competitive AI infrastructure, no single competitor can build a durable advantage around it. Meta's AI features are distributed across Facebook, Instagram, and WhatsApp, reaching billions of users. Meta is not primarily an AI company; AI is a capability it needs to remain competitive.

xAI, founded by Elon Musk in 2023, produces the Grok models distributed through X (Twitter) premium subscriptions. Musk has spent years publicly warning about AI existential risk — he was an early OpenAI backer and board member, and co-founded it before departing — while simultaneously racing to build competitive AI systems. Grok has been deliberately positioned with fewer content restrictions than competitors. The lab publishes minimal safety research.

Mistral AI is a French startup founded by former DeepMind and Meta researchers, known for producing capable models at smaller parameter counts than the frontier labs. Several of its models are released as open weights. It's primarily targeting the European enterprise market, where GDPR compliance and data sovereignty concerns create demand for models that can be run on European infrastructure. Its focus is efficiency and compliance more than safety research.

At a glance

The "safety posture" indicator above reflects observable actions, research output, and structural commitments — not the language on any lab's website. Every lab has a safety page; not all of them have safety behavior that matches it.

Safety: what the actions show

Every AI company talks about safety. The useful question is what they do about it.

Anthropic is the most coherent case. It was founded because people believed safety-focused development needed to exist as an independent priority, not as a feature added to a product roadmap. Its Responsible Scaling Policy is written down and public. Its interpretability research is aimed at actually understanding model internals — arguably the most important unsolved problem in AI safety — rather than at making models seem more trustworthy. The company's corporate structure legally constrains it in ways that voluntary commitments don't.

Google DeepMind has genuine research infrastructure here. The team has produced important work on reward hacking, specification gaming, and agent safety. The complication is Alphabet: a parent company whose revenue depends on search and advertising has structural incentives to deploy AI quickly into those products, regardless of what the safety researchers prefer.

OpenAI is where the stated values and observable behavior have diverged most visibly. The events are documented:

In May 2024, Jan Leike — who ran OpenAI's superalignment team, tasked with solving long-term AI safety — resigned. His public statement was unusually direct: "Safety culture and processes have taken a back seat to shiny products." Ilya Sutskever, the co-founder most associated with safety concerns, departed around the same time. The long-term safety team was subsequently dissolved. The board fired Sam Altman in 2023 over concerns about candor; the safety-focused board members were replaced after he was reinstated. None of this is interpretation — these are reported events.

xAI publishes the least safety research of any major lab and has positioned its product around having fewer restrictions. Treating this as a safety posture is not unfair to the company — it's what they've said.

Voting with your dollars

If you're paying for AI products — subscriptions, API access, enterprise contracts — that spending is a real market signal. Companies read it.

The next lesson looks at a different axis entirely: open source AI, what it can actually do for you, and what hardware it requires.