The State of American AI, October 2026

How this is counted. Each published organization record has exactly one ownership category, recorded from its sources: publicly traded, privately held, nonprofit, a unit of another organization, foundation-hosted, a public institution, or unknown. Research units and subsidiaries have their own records. Rows add up to the total.

How this is counted. A unit is a research unit, division, subsidiary, or hosted project whose record names a parent organization. Units are counted separately so that a parent and its unit are never presented as two independent companies. Rows add up to the total.

How this is counted. Roles are the documented activities recorded for each organization. Most records have more than one role, so the rows add up to more than the 150 records; each share is out of all organization records.

How this is counted. Sectors are applied only where a record's evidence supports them, and a record can have several, so rows add up to more than the 150 records. Records with no sector tag are counted in their own row.

How this is counted. The state is read only from the text of each record's sourced headquarters label (for example “Santa Clara, California”). Offices, data centers, and other locations are not counted. Records whose label names no U.S. state are counted in the last row. Each record counts once; rows add up to the total.

How this is counted. Founding years come from each record's sourced founding date. Records with no founding year recorded are counted in the last row: a missing year means the catalog has not documented it, not that the organization is new. Units are counted separately from their parents. Rows add up to the total.

How this is counted. Model families (overview records for a line of models) are counted separately from model releases (one record per assessed release, which carries the license and checklist). Frameworks, runtimes, and research stacks together are this report's software. Rows add up to the total.

How this is counted. Tiers are computed from each release's public-materials checklist and licenses under USASI rubric v0.2; they are never typed by hand. Each release has exactly one tier, so rows add up to the total. In the rubric the open tiers are cumulative (every open-stack release is also open-weight), so open-weight in the wider sense is the first three rows together: 90 releases.

How this is counted. From the weights entry of each release's checklist. Public download: the general public can download the parameters (license terms may still restrict use). Gated: available only by request, with approval, or to some users. Not public: documented as unavailable. Not assessed: the evidence is missing or insufficient. Rows add up to the total.

How this is counted. Open-weight releases (the open-weight tier or a higher one) counted for each organization record the release lists. A release that lists two organizations counts for both, so rows can add up to more than the total. Counts reflect the releases this catalog has assessed, not everything an organization has published.

How this is counted. Licenses that apply to a release's weights (weights, weights and code, or everything). Standard permissive: an SPDX identifier on this report's permissive list (Apache-2.0, Apache-2.0 WITH LLVM-exception, MIT, BSD-2-Clause, BSD-3-Clause, ISC, NCSA, Zlib, BSL-1.0, 0BSD, Unlicense, CC0-1.0, CC-BY-4.0, ODC-By-1.0, PDDL-1.0, CDLA-Permissive-2.0). Standard copyleft: a GPL-family, MPL, EPL, or share-alike identifier. Non-commercial: the license's guide on this site is in the non-commercial category, or its SPDX identifier or name says non-commercial. Custom: any other license, usually the publisher's own terms without an SPDX identifier. A release with several weights licenses counts once, in its most restrictive family (non-commercial, then custom, then copyleft, then permissive). Rows add up to the total.

How this is counted. Each license that applies to a release's weights, grouped by this site's license guide when one matches and otherwise by SPDX identifier or recorded name. A release counts once per license; a release with two weights licenses counts under both.

How this is counted. Each license that applies to a project's code, grouped by license guide, SPDX identifier, or recorded name. Projects often license different parts differently, so a project can count under several licenses and rows add up to more than the total.

How this is counted. A configuration is one documented way to do a task: a specific combination of model, runtime, app, or hosted service, with requirements restated from the publisher's own documentation. Local: runs on your own computer. Hybrid: a tool on your computer that sends requests to a hosted model. Hosted: runs on the provider's servers. Rows add up to the total.

How this is counted. The publisher's own memory statement for the exact variant named on each configuration, as recorded with its source. Graphics-memory (VRAM) and system-memory figures are kept apart, because they describe different memory. A figure is the publisher's estimate, not a USASI test, unless the record documents a test. Configurations without a stated figure are counted in the last row. Rows add up to the total.

How this is counted. The same rule as the Local AI by memory pages: a configuration is listed when every finder requirement is met and the publisher's figure is within the amount. A configuration that fits 8 GB also fits larger amounts, so rows overlap; shares are out of all local configurations.

How this is counted. Listings from employers' own job feeds that were open and confirmed within 48 hours of the snapshot time, the Jobs directory's default view. Applications are on employers' sites; job data are refreshed daily, so these counts are frozen with the edition. Locations are read from the employer's text without geocoding. U.S. locations only: names the United States or a U.S. state and no place outside it. U.S. and elsewhere: names both. Outside the U.S. only: names only places outside it. Not determinable: gives nothing the rules can place (for example a city without a state). Rows add up to the total.

How this is counted. A listing counts in a state when one of its employer-supplied locations names it: a postal code after a city (“Atlanta, GA”), the full state name, “New York City” or “NYC”, or “Washington, DC”. This is the rule of the Jobs by state pages. A listing that names several states counts in each, and listings that also name places outside the U.S. are included, so rows overlap. Locations that give only a city or only the country name no state and are not counted here.

How this is counted. Each listing counts once, for the organization whose official job feed published it. Only organizations whose feed the site reads are included. Counts describe open postings at one moment, not company size, growth, or hiring trends. Rows add up to the total.

How this is counted. Remote: the employer marks the workplace as remote, or a listed location says “Remote” (for example “Remote - US”). Remote, U.S. locations only: also names only U.S. places, the rule of the Remote jobs, U.S. locations page. Rows add up to the total.

How this is counted. Each published document in the policy tracker, by its document type as recorded from the official text. Types with no document are shown as zero. Rows add up to the total.

How this is counted. Status as recorded from official sources when the document was last reviewed: in effect, final, revoked, superseded, rescinded, draft, or status not stated. A revoked or superseded document stays in the tracker with that status. Rows add up to the total.