---
title: Frontier Lab
type: vocabulary
url: "https://www.envisioning.com/vocab/frontier-lab"
summary: A small number of companies building the largest, most capable AI models at industrial scale.
year: 2023
generality: 0.50
---

# Frontier Lab

A small number of companies building the largest, most capable AI models at industrial scale.
A frontier lab is one of a small handful of organizations — typically five to ten global entities — that train the largest and most capable AI models at industrial scale. The category emerged in the 2020s as the cost and complexity of training a frontier model — the leading edge of model capability — grew past the point where academic labs, startups, or non-specialized companies could realistically compete. Members of the category, as of 2026, include US-based OpenAI, Anthropic, Google DeepMind, Meta FAIR, and xAI, plus selected Chinese labs (DeepSeek, Qwen, Moonshot, Zhipu AI, and others), with the EU and other regions working to seed domestic equivalents. The defining feature is not just the scale of model training but the industrial infrastructure — capital, talent, energy, and data pipelines — that sustains it as a recurring capability rather than a one-off achievement.

Mechanically, a frontier lab is a different kind of organization from either a research lab or a software company. Its product is a model — trained, evaluated, and versioned — and the underlying economic asset is the trained model's weights, which are then exposed through APIs, integrated into products, or licensed to enterprises. The lab combines research, infrastructure engineering, safety teams, evals, and product teams under one roof, often with significant overlap. Capital structure is distinctive: the labs require tens of billions of dollars in compute and talent investment that is expensed long before corresponding revenue arrives, which shapes them into entities with high burn rates and contested unit economics. The strategic decisions made by frontier labs — which capabilities to build, which deployment modes to allow, which safety thresholds to enforce before release — have outsized influence on the trajectory of the field as a whole, because they are the gatekeepers for capabilities that no one else can match.

The advantage of the frontier-lab framing is that it recognizes that capability in AI has become a concentrated attribute. A handful of organizations control the training of the systems whose capabilities set the trajectory of the field. The framing also distinguishes a structural feature of the industry — that a small set of entities function as chokepoints for capabilities, capital, talent, and increasingly governance — which has implications for competition policy, antitrust, and national security. The cost of the framing is that it elides the diversity within the category: the labs differ substantially in structure (some are public-benefit corporations, some are divisions of larger companies, some are private), in openness (some publish most of their research, others do not), and in alignment with national security establishments. The frame also risks treating the category as closed: it understates how rapidly the set of frontier labs has changed (DeepSeek was not on the list in 2023 and was by 2025) and how easily a trailing lab can become frontier-capable with one good result.

Open questions include whether the frontier-lab set stays small (because training costs keep rising and concentration is reinforced) or expands (because training efficiency improves and new entrants catch up), whether safety and governance regimes designed around a small frontier-lab set remain valid as the set expands, and how the relationship between US, Chinese, and other sovereign frontier labs evolves as the geopolitical dimension becomes load-bearing. There is also a governance question about whether frontier labs should be regulated as utilities, as dual-use technology exporters, or as something new, and what mechanisms would make their decision-making legible without destroying the productive competition between them.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/frontier-lab)