Frontier Labs as National Laboratories: 2026-10-22
Snapshot · horizon by end of 2030 (2030-12-31)
Each probability is a separate estimate that its path happens by the horizon. Paths can co-occur, so they do not sum to 100%. Estimates: draft estimates by the research team.
A. Ad hoc levers continue (the status quo)
85%range 75–93%
Ad hoc levers continue (the status quo). Testing stays voluntary, controls are applied model by model, and disputes go to court.
Resolves YES if
Resolves YES if between 2026-10-23 and 2030-12-31 the federal government applies at least two new firm- or model-specific measures to named frontier developers: an export control on named models, a procurement designation, or a negotiated pre-release access agreement.
Signposts
Emerging↑more firm- or model-specific federal directives or designations after 2026-10-22as of 2026-10-22
B. Steering through conditions on chips and contracts
80%range 65–90%
Steering through conditions on chips and contracts. The state governs frontier compute it does not own by attaching conditions to export licenses and investment agreements.[183][185] It is path G applied abroad first, and it is already running: Microsoft–G42 operates under an "Intergovernmental Assurance Agreement," and the 2025 G42 and HUMAIN approvals carry "security and reporting requirements."[183][185]
Resolves YES if
Resolves YES if between 2026-10-23 and 2030-12-31 there are at least three new publicly announced US export approvals, licenses or government-backed agreements for foreign AI compute that state security, reporting, know-your-customer or monitoring conditions, or a published federal rule that imposes such conditions.
Signposts
Emerging↑intergovernmental assurance agreements reused as a template for other AI campusesas of 2026-10-22
Not yet↓approvals granted without stated conditionsas of 2026-10-22
Proposers and analogues
International analogues
United Arab Emirates: Microsoft–G42 investment under an Intergovernmental Assurance Agreement [183]
United Arab Emirates and Saudi Arabia: Commerce approvals for G42 and HUMAIN conditioned on security and reporting requirements [185]
id B-conditional-compute · horizon 2030-12-31
C. Frontier labs as users and partners of national facilities
60%range 45–75%
Frontier labs as users and partners of national facilities. AI follows the beamline pattern: government-owned compute (Solstice, Lux, Doudna) allocated by merit review, with company work handled as proprietary use at full cost recovery, CRADAs or Strategic Partnership Projects.[134][140][158] This is the path most grounded in how Argonne and Berkeley Lab already work.
Resolves YES if
Resolves YES if between 2026-10-23 and 2030-12-31 a named frontier developer is publicly documented training, fine-tuning or running frontier-model research on DOE-owned compute under a user agreement, CRADA, Strategic Partnership Project or merit allocation.
Signposts
Not yet↑an INCITE-style call that allocates frontier-scale AI computeas of 2026-10-22
Emerging↑Lux-type public-private co-investments repeatedas of 2026-10-22
Chessen and Martell (Lawfare): Apollo-style whole-of-society program [198]
Public AI writers (Mozilla, Public AI Network, Brookings): Public option across the AI stack [202][203][205]
Google: Advanced partnerships with national labs; export rules that allow exports [221](interest: Cloud and AI-services exporter)
International analogues
United Kingdom: AI Research Resource (Isambard-AI) allocated by independent panel [164][165][166]
European Union: AI Gigafactories on EuroHPC built and operated by consortia [169][170]
id C-lab-compute-partnerships · horizon 2030-12-31
D. A national-champion bargain
30%range 15–45%
A national-champion bargain. A Bell-style arrangement: protection, procurement and possibly equity in exchange for a public return.[101][109]
Resolves YES if
Resolves YES if by 2030-12-31 the US government acquires equity, warrants, a golden share or a contractual revenue share in a frontier developer, or signs an agreement granting it preferred status explicitly in exchange for a public return.
Signposts
Not yet↑a finalized government equity stake, and whether it carries governance rightsas of 2026-10-22
Arbel, Goldstein and Salib: Narrow government halt right [199]
id D-national-champion-bargain · horizon 2030-12-31
E. A public lab spins out a company
25%range 12–40%
A public lab spins out a company. A government lab builds a capability and hands it to a new firm, as ITRI did with TSMC in 1987.[196] ITRI had earlier licensed IC technology from RCA in 1976 and spun off UMC in 1980.[196][197] The flow runs opposite to the thesis, from public to private.
Resolves YES if
Resolves YES if by 2030-12-31 a model trained at a DOE national laboratory, or the core team that built it, moves to a new company or nonprofit under a license, CRADA or similar agreement that is publicly announced.
Signposts
Not yet↑a DOE science model such as AuroraGPT, or its team, moving into a company or nonprofitas of 2026-10-22
Taiwan: ITRI transferred fabs, equipment, technology and staff to TSMC (1987); earlier UMC spin-off [196][197]
id E-public-lab-spinout · horizon 2030-12-31
F. An AI FFRDC under an M&O contract
20%range 8–35%
An AI FFRDC under an M&O contract. The government owns a secure AI facility and contracts an operator, possibly an LLC of companies and universities on the Triad or UT-Battelle pattern, with performance ratings and recompetes as the accountability tool.[148][149][152]
Resolves YES if
Resolves YES if by 2030-12-31 Congress or a federal agency establishes a new FFRDC, or awards a management-and-operating contract, for a government-owned secure facility whose stated mission includes frontier-AI development or evaluation.
Signposts
Not yet↑legislation or appropriations naming an AI FFRDCas of 2026-10-22
Not yet↑a DOE or DoD solicitation for an AI facility operatoras of 2026-10-22
Not yet↑the 2025 Action Plan's high-security data centers becoming a government-owned siteas of 2026-10-22
RAND: Security levels for model weights that such a site would need [208]
International analogues
European Union: Gigafactory consortia that build and operate public AI sites [170]
id F-ai-ffrdc · horizon 2030-12-31
G. The regulated circuit (Alex's end state)
15%range 7–30%
The regulated circuit (Alex's end state). Compute and model access are licensed, covering both training runs and distillation through APIs.[125][126]
Resolves YES if
Resolves YES if by 2030-12-31 a federal statute or final rule requires licensing, registration with approval, or mandatory pre-release access for frontier models or large AI data centers, or imposes know-your-customer duties on compute or API customers.
Signposts
Not yet↑a replacement for the diffusion or compute-reporting rule that reaches API customersas of 2026-10-22
Microsoft: Licensing; licensed AI datacenters; KYC [220](interest: Already operates large datacenters)
Dean Ball: Licensed private certifiers; liability safe harbor [200]
Samuel Hammond: Compute thresholds as a temporary measure [201]
Arbel, Goldstein and Salib: Narrow government halt right [199]
Dario Amodei / Anthropic: Democratic entente; national-security testing; regulation of all frontier firms [210][211][127](interest: Chip limits also restrict competitors' distillation (see ledger 125))
International analogues
European Union: AI Act plus General-Purpose AI Code of Practice [172]
id G-regulated-circuit · horizon 2030-12-31
H. A contestable national-champion program
12%range 5–25%
A contestable national-champion program. The state funds several model builders and re-selects them in rounds on stated criteria, the way DOE re-competes site operators.[148][180] Korea's program dropped two of five teams in its first round, and a newly selected team receives compute, datasets and a "K-AI" designation.[180]
Resolves YES if
Resolves YES if by 2030-12-31 a US federal program funds at least three frontier-model developers with public compute or money and publishes re-selection or elimination rounds.
Signposts
Not yet↑a US program that funds several frontier developers with public compute and published elimination criteriaas of 2026-10-22
South Korea: Sovereign foundation-model program with elimination rounds [180]
id H-contestable-champions · horizon 2030-12-31
I. An international lab
5%range 2–12%
An international lab. A CERN-style consortium with allies.[123][124]
Resolves YES if
Resolves YES if by 2030-12-31 the US joins a treaty or intergovernmental agreement establishing a jointly governed frontier-AI research facility with other governments.
Signposts
Emerging↑shared evaluations between CAISI and the UK AI Security Instituteas of 2026-10-22
Dario Amodei: Entente of democracies rather than a shared lab [210]
Sam Altman: Coalition of democracies (2024 op-ed) [212]
International analogues
European Union: AI Gigafactories (EU program, not US-joined) [169]
id I-international-lab · horizon 2030-12-31
Source labels: reported rests on a single press account; snippet only means a quote was seen only in a search snippet.
Method and re-scoring rules (from the latest snapshot)
independence
Each probability is a separate estimate that its path resolves YES by the horizon. Paths can co-occur, so probabilities do not sum to 1; when one path's resolution makes another more or less likely, say so in the rationale rather than normalizing.
signpost mapping
Each signpost has a direction. not_yet->emerging moves the path roughly 0.03-0.05 in that direction; emerging->happened or not_yet->happened roughly 0.05-0.15, depending on how close the signpost is to the resolution criterion. A reversal (happened back to not_yet) moves it by the same amount the other way. Keep low<=point<=high and widen the range when evidence conflicts.
resolution
A path resolves YES when its resolution_criteria are met by a public document dated after 2026-10-22 and on or before the horizon; it resolves NO at the horizon otherwise. Record the resolving ledger id. A resolved path keeps probability 1 or 0 in later snapshots.
cadence
Re-score quarterly, and also within two weeks of any event that changes a signpost to 'happened'.
Why this horizon: Spans the next presidential term and a full contracting or export-licensing cycle, close enough that the 2026 record still bears on it, and matches stated delivery dates such as FugakuNEXT's 'around 2030' (ledger 177).