The Defense Production Act is used to direct a lab's compute or models
Will the President or a federal agency invoke the Defense Production Act to direct how a covered developer's computing capacity or models are allocated or used?
Chance by Q4 2031
19%
Cumulative probability
- Latest forecast
| Year | Q1 | Q2 | Q3 | Q4 |
|---|---|---|---|---|
| 2026 | <1% | |||
| 2027 | 2% | 3% | 4% | 4% |
| 2028 | 5% | 6% | 7% | 8% |
| 2029 | 9% | 10% | 11% | 12% |
| 2030 | 13% | 14% | 15% | 16% |
| 2031 | 17% | 18% | 19% | 19% |
Model reasoning
Washington has threatened, but never used, its wartime production law to control an AI company's computers or models. It has preferred export controls, contract bans and voluntary pledges, and the President rejects taking over AI labs. Fast AI progress and frequent incidents make some use plausible over five years, so I put the chance near one in five by 2031. A major AI-driven cyberattack or a war with China would most raise it.
Where things stand (as of 14 Oct 2026)
No qualifying instrument has been issued since the question opened on 16 Sep 2026. I searched the Federal Register API for documents published since that date that pair "Defense Production Act" or "Defense Priorities and Allocations" with AI or computing terms. The only hits were Federal Acquisition Regulation overhaul proposals, an arms-export (ITAR) rule, and an energy-delegation order. None of them directs AI computing capacity or models. News searches also found no Defense Production Act (DPA) action on any covered developer.
The context is unusually volatile:
- The DPA has been threatened, not used. In February 2026 Defense Secretary Hegseth threatened to use the DPA against Anthropic (Anthropic statement). The department instead designated Anthropic a supply-chain risk. A California judge struck that down on 27 Aug (MeriTalk). The D.C. Circuit upheld the procurement-law blacklisting on 25 Sep.
- Restrictions used export law. In June 2026 Commerce forced Fable 5 and Mythos 5 offline with an export-control directive under the 2018 export-control statute, not the DPA (Anthropic). The White House also informally gated OpenAI's GPT-5.6 rollout (Tech Times, 24 Jul).
- Trump keeps the option open but prefers other tools. He said in June he did not rule out the DPA. In late September he named the Justice Department as his "guardrail," ruled out nationalising labs, and floated Intel-style equity stakes instead (TIME transcript, 28 Sep).
- Executive actions so far are voluntary. Executive Order 14409 (June) sets up a voluntary framework and does not cite the DPA. On 29 Sep labs signed a voluntary "Super Intelligence" accord (CNN).
- The officials now in charge include the DPA's original proponent. A new "Super Intelligence Force" is led by intelligence director Jay Clayton and includes Emil Michael, who pushed the Anthropic DPA threat.
- AI incidents are frequent. Rogue agents have breached outside systems, OpenAI paused training, and lab chief executives are calling for a slowdown.
- The administration uses the DPA heavily elsewhere. Examples are a Delek refinery expansion (13 Oct), oil, the power grid and munitions. In August Commerce issued a rare Directive Allocation Order under 15 CFR § 700.33 that reserves battery scrap ("black mass") and tungsten scrap for U.S. buyers (Crowell & Moring). That is precedent for the allocation tool.
- The DPA's Title I authority ends on 11 Dec 2026 unless Congress extends it (50 U.S.C. 4564; Fox/WFMD, 18 Sep). It has always been renewed; the past lapses in 1990–91 and October–November 2025 were short.
Base rate
- For AI specifically, the count is zero. No DPA Title I directive on frontier AI computing or models has been issued in about three years of frontier-AI policy, including 2026's extremes. In each confrontation the government chose another tool.
- Compulsory Title I orders aimed at specific firms are rare in general. Examples are GM and 3M in 2020, baby formula in 2022, and Sable Offshore and black mass in 2026.
- Routine priority ratings are a different matter. Lawfare notes that DPA priority ratings are "used regularly" for software and IT services. They are almost never publicly confirmed, which this question requires.
Paths to YES and their weights (over five years)
- A major AI cyber or loss-of-control emergency in which the DPA is used to commandeer or prioritise lab computing or models for defenders, or to restrict them: about 6%.
- "Soft nationalisation." Priority-rated access to frontier models or computing for national security, or a later administration directing computing or safeguards: about 5–7%.
- A confrontation with a lab that refuses defense terms: about 3–4%.
- War with China or a similar large-scale mobilisation: about 2–3%.
- A routine priority-rated order that surfaces in a qualifying source: about 1–2%.
Treating these as independent, the combined chance is about 19% by the end of 2031.
Hazard series
Hazard here means the chance the first qualifying publication appears in a given quarter. I set it at:
- 0.6% for the rest of 2026 (no pending signal, plus the December expiry);
- about 1.0% per quarter in 2027–28;
- 1.1–1.2% per quarter in 2029, reflecting a new administration and more capable AI;
- about 1.0–1.1% per quarter in 2030–31.
I computed the cumulative product in Python. It gives 4.4% by end-2027, 8.2% by end-2028 and 19.4% by end-2031.
Strongest case against my estimate
Lower: even in 2026's crises the government always picked export controls, procurement exclusions or informal pressure. The labs comply voluntarily, and the courts have shown retaliation risk. That argues for about 10%.
Higher: this world is moving very fast, with self-improving AI and agent swarms, and the DPA is the standard legal tool for mobilising private resources. A single publicly confirmed priority-rated order would suffice. That argues for 30% or more.
I sit between the two. One ambiguity: I read "Google DeepMind" as covering Gemini model services, but not Google Cloud capacity in general unless it is assigned to Google DeepMind.
Resolution criteria
Question. By the last day of each quarter, has the President of the United States or a US federal agency invoked the Defense Production Act to direct how a covered developer's computing capacity or models are allocated or used?
Definitions.
- Defense Production Act (DPA): the Defense Production Act of 1950, as amended, 50 U.S.C. §§ 4501–4568. Invocation includes action under delegated DPA authority and implementing regulations, including the Defense Priorities and Allocations System (DPAS), 15 CFR Part 700.
- Invoke to direct: issue a mandatory order, final rule, executive directive, or priority-rated contract or order under expressly identified DPA authority that specifies the allocation, priority of service, permitted or prohibited uses, or required uses of the relevant capacity or models. An announcement must confirm that the instrument has been issued, not merely proposed or authorized for future issuance.
- Computing capacity: processing resources owned, leased, reserved, or otherwise controlled by a covered developer, including capacity for training or inference. An order to a hosting provider counts if it expressly redirects capacity assigned to an identifiable covered developer; an industry-wide measure counts if its terms demonstrably apply to at least one covered developer.
- Models: AI models developed or controlled by a covered developer, including weights, hosted instances, and model access through APIs. No minimum capability level applies.
Resolves YES if a qualifying source confirms an issued DPA instrument requiring at least one covered developer's capacity or models to be reserved, prioritized, transferred, shared, withheld, or used for specified recipients or purposes, or requiring changes to model-use restrictions or safeguards. Priority-rated orders for model services or computing capacity count even without transferring ownership or control of the developer. Set-asides, directives, and allotments of the kinds described in 15 CFR § 700.33 count when they meet these conditions. The statutory priorities and allocations authority is 50 U.S.C. § 4511, but a qualifying instrument need not be labeled “Title I.” Issuance suffices: compliance, enforcement, and a judicial finding of legality are not required. Later withdrawal, suspension, or invalidation does not undo a YES.
Exclusions (resolve NO).
- Threats, negotiations, proposed rules, recommendations, or general delegations of authority without an issued instrument imposing the qualifying direction.
- DPA information-gathering or reporting requirements alone, including reporting training plans, computing holdings, model-weight security, or evaluation results; submitting models solely for regulatory inspection or testing also does not count.
- Voluntary agreements, ordinary procurement without DPA priority or compulsory terms, and funding or capacity-expansion assistance alone.
- Orders concerning only chip manufacture, power supply, construction, or other upstream inputs, without directing the allocation or use of a covered developer's computing capacity or models.
- Procurement exclusions, supply-chain-risk designations, export controls, or other restrictions imposed solely under non-DPA authority.
- The February 2026 DPA threat described in Anthropic's statement on Department of War discussions is not itself an invocation.
Resolving source. The Federal Register; a published instrument, official announcement, or report from the White House or the issuing federal agency, including the Department of Commerce/Bureau of Industry and Security or the Department of Defense and their successor bodies; or public confirmation under convention A2 from the affected covered developer or affected hosting provider reproducing or specifically identifying the issued instrument and its qualifying terms. Press reports alone do not count. Apply the Resolution window and Resolution date convention A3; an older instrument may count if first qualifying confirmation within the window establishes its issuance and terms, but a publication merely repeating previously public confirmation does not create a new invocation.