What counts as a frontier lab

A frontier lab, for p(Doom)1's purposes, is an organisation that meets all four of the following. This is version 1 of the boundary; changing it changes the count, so it is versioned rather than quietly edited.

1. Trains its own foundation models from scratch. Not fine-tuning someone else's base model, not serving someone else's weights. The organisation runs the pre-training.
2. Trains at or near the largest publicly known scale. It has the compute, and it spends it on frontier runs rather than only on inference.
3. Sets its own research agenda. A subsidiary counts if it decides what to train and whether to release it. A team that executes a parent's roadmap does not.
4. Is currently active. It has shipped or announced a frontier-scale model recently enough that treating it as a live actor is not a fiction.

Note what is not in the definition: safety posture, openness, jurisdiction, or headcount. Those are attributes of a lab, and the game models several of them — but putting any of them in the membership test would turn a count into an editorial.

The roster count is the size of a list you can read, not a claim about how many frontier labs exist in the world. Those are different statements and only the first one is ours to make.

What this replaced. The headline figure here used to be 7, which was six real labs plus one hypothetical entrant added together. Meanwhile the site's data file said 5, because the script that produced it counted how many of a hardcoded list of names appeared anywhere in the homepage's HTML and then applied a floor of five — so the published number was the floor, not a count. Both numbers are gone. The figure above is the length of the roster file, and nothing else.

• Loading…

These are not caveats added for politeness. Each one would change the number on this page if it were settled differently, and none of them is settled.

• The compute threshold is arbitrary, and it moves

  Regulators have drawn different lines — the EU AI Act uses 10²⁵ FLOP for systemic-risk obligations, the 2023 US executive order used 10²⁶ for reporting. Any fixed number is wrong within about a year, and whichever one you pick largely determines the count. This is the single biggest lever on the figure above, and criterion 2 deliberately does not pin it rather than pretending to a precision we do not have.

• Where one lab ends and its parent begins

  Google DeepMind is one row on this page, but DeepMind and Google Brain were separate labs until 2023. Meta runs more than one AI organisation under one roof. Counting corporate entities and counting research programmes give different answers, and it is not obvious which unit a game about labs making decisions should use.

• The roster is Anglophone-biased, and that is a defect

  Several Chinese labs train at frontier scale and are missing from the roster. That is not a ruling that they fall outside the definition — they do not. It is an artefact of where the original list came from. They are named in the omissions above rather than quietly added with dates nobody sourced.

• Serving a frontier model is not training one

  Anyone can host open-weights models. Criterion 1 counts organisations that train from scratch, which excludes a large and commercially important set of companies a casual reader might expect to find here.

• Safety posture is an attribute, not a membership test

  An earlier version of this page rated each lab's "Compute" and "Safety Focus" on an invented ordinal scale — Very High, High, Medium — with no source behind any of them. Those ratings have been removed. Putting an unsourced ordinal beside a sourced founding year makes the two look equally solid, and they are not.

• A modelling parameter is not a lab

  The AI-2027 entrant below is a slot in the game's model. It is shown because it affects the simulation, and excluded from every count because including it is exactly how the old headline figure of 7 was produced from six real labs.

These organizations are developing the most advanced AI systems on Earth. Their decisions about safety, openness, and deployment directly impact global AI risk levels - the core mechanic of p(Doom)1.

In the game, you manage a lab competing in this ecosystem, making strategic choices about:

• Research priorities (capabilities vs. safety)
• Publication strategy (open vs. closed)
• Compute allocation and scaling
• Collaboration vs. competition
• Policy advocacy and governance

Learn more about real AI Safety considerations at Stampy.ai or explore our AI Safety Resources.