AI Principles · Foundation

The foundation-model labs

A handful of organizations build the frontier models that nearly everyone else builds on. The club is small for a reason: scaling laws turn compute into a capital barrier only a few can clear.

Read at your depth: 01 The answer · 02 Intuition · 03 Mechanics · 04 The math · 05 The code · 07 Sources

01The answer, then the intuition

A short list, split in two

Training a frontier model takes a cluster worth billions, a rare concentration of talent, and years of accumulated know-how. So the organizations that can do it number in the handful, not the hundreds — and they split along one decisive line: closed labs that sell access to weights they keep private, and open-weight labs that publish downloadable models anyone can run.

That single choice — open or closed — shapes each lab's business, its backers, and its role in the market. Filter the landscape and click any lab to see its public profile:

The frontier labs — a landscape snapshot

Publicly known model families and stances (2025-era). ▮ closed ▮ open-weight.

Snapshot of public information; the roster and positions shift quickly. Not investment advice.

02Mechanics

Why the club stays small — and split

The capital barrier. A frontier training run costs on the order of a large data center's output for months. Only players with access to that capital — usually via a hyperscaler partnership — can sustain it, so the frontier is structurally an oligopoly.
Closed frontier. Labs like OpenAI, Anthropic, and Google DeepMind keep weights private and sell access through APIs and products. Their bet: stay ahead on capability and capture value through the interface, funded by deep-pocketed compute partners.
Open-weight. Meta (Llama), Mistral, DeepSeek, and Alibaba (Qwen) publish downloadable weights. Their bet: commoditize the layer below the frontier, win developers and ecosystems, and deny closed labs a durable moat on "good enough" intelligence.
The backers. The circular twist: cloud providers fund the labs (Microsoft↔OpenAI, Amazon and Google↔Anthropic), because a lab's success drives demand for the backer's own compute. The capital and the compute come from the same place.

So the map isn't just a roster — it's a structure. A few closed labs push the frontier; a few open labs chase them from just behind; and a small number of hyperscalers bankroll nearly all of it, because the whole thing runs on their chips.

04The math

expand ▾

Why concentration is baked in

Concentration follows directly from the scaling law. If a frontier run costs $R$ and a player needs headroom $h$ times that to sustain a program, only players with capital $K_i$ above the threshold can compete:

$$ \text{viable labs} = \big|\{\, i : K_i \geq h \cdot R \,\}\big| $$

And scaling laws make $R$ grow by roughly $10\times$ per generation. Since capital pools are fixed and long-tailed, each $10\times$ in $R$ prunes the field from the bottom — the count of viable players falls monotonically as the frontier advances:

$$ R \uparrow 10\times \;\Longrightarrow\; \text{viable labs} \downarrow $$

That's the whole industrial logic in one line. The same predictability that makes scaling a good bet also makes it an expensive one — and expense concentrates. Open-weight labs partly escape by not needing to monetize the model directly, but the frontier itself trends toward fewer, larger players every generation.

05The code

expand ▾

The field, pruned by cost

As a frontier run gets 10× more expensive each generation, count who can still afford it.

concentration.py

# illustrative capital pools ($); the point is the trend, not the exact names
players = {"BigCloud A": 200e9, "BigCloud B": 150e9, "BigCloud C": 100e9,
          "Well-funded lab": 20e9, "Mid lab": 5e9, "Startup": 1e9, "Academic": 0.1e9}

for run_cost in [1e8, 1e9, 1e10, 1e11]:
    can = [n for n, k in players.items() if k >= run_cost * 10]   # need ~10x headroom
    print(f"run cost ${run_cost:.0e}: {len(can)} of {len(players)} can sustain it")
# run cost $1e+08: 6 of 7 can sustain it
# run cost $1e+09: 4 of 7 can sustain it
# run cost $1e+10: 3 of 7 can sustain it
# run cost $1e+11: 0 of 7 can sustain it   <- the frontier prices almost everyone out

07Going deeper

expand ▾

The primary sources

Stanford HAI — AI Index Report · who's producing frontier models, and at what cost.
Epoch AI — Notable AI Models · a tracked database of models, labs, and training compute.
Bommasani et al. (2021) — On the Opportunities and Risks of Foundation Models · the concept and concentration.
SemiAnalysis · the compute-and-capital relationships between labs and their backers.

"First Principles — how AI actually works", chapter "The foundation-model labs", by The Catch. CC-BY 4.0. Hosted at TheCatch.AI / Tech Stack / AI Principles.

Originally from First Principles by The Catch · CC-BY 4.0 · Hosted at TheCatch.AI / Tech Stack / AI Principles