White House tech strategy leaves open-weight AI off its critical list

The White House has published a new National Security Science and Technology strategy. It names 15 critical technology areas the United States intends to lead. Artificial intelligence and autonomy is one of them, with 12 subfields listed underneath. Open-weight AI is not among them. Neither open weights nor open-source software development appears anywhere in the 24-page document.

The strategy is dated August 2026 and comes from the Director of the Office of Science and Technology Policy. It meets a congressional requirement written into the CHIPS and Science Act. That law obliges OSTP to submit a science and innovation strategy after each National Security Strategy. This one supports the 2025 NSS, and its appendix updates the official OSTP list of critical and emerging technologies.

What the strategy prioritises

The document sets three priorities for battlefield dominance: undersea, space, and AI and autonomy. It ties those to deterrence in the Indo-Pacific. Twelve enabling technologies follow, running from directed energy and hypersonics to semiconductors and nuclear energy.

On how to fight, it leans towards mass and cheapness. The United States will seek “optimal combinations of lower-cost (and sometimes lower-tech or attritable) platforms” that can be fielded in larger numbers, the strategy says. It also pushes Pentagon buyers towards other transaction authorities and milestone-based contracts.

Agencies should try “faster non-traditional approaches that may involve higher-risk but offer high potential reward”, it adds.

The AI list, in full

Under artificial intelligence and autonomy, the appendix starts with perception and sensor fusion, planning and reasoning, and robotics and embodied intelligence. Then come foundation models, multi-agent systems and swarm intelligence, and autonomous systems across five domains.

The list ends with autonomous command and control, agent identification and authentication, interpretability and control, adversarial robustness and AI security, continual learning, and distributed and privacy-preserving machine learning.

Foundation models cover “large language, multimodal, and world systems”, the appendix says. Nothing there refers to model weights, to open release, or to compute thresholds. The word “open” appears five times in the whole strategy. Each time it describes open research environments or an open investment climate, never a model.

Two nearby entries make the gap visible. The strategy calls brain-computer interfaces critical, under future computing technologies. It also calls distributed ledger technologies and digital assets critical, under information management and cybersecurity.

The argument that this is a gift to China

Patrick Tucker at Nextgov first flagged the omission and called it a huge oversight. He argues that the strategy sides with a handful of well-connected companies and their energy-intensive approach to AI. Cheaper alternatives that researchers increasingly favour lose out. The AI giants had already split on open weights in July, with Nvidia on one side and OpenAI and Anthropic absent from the other.

Tucker cites Michael Schiffer, a senior advisor at Scalare Advisors and a senior fellow at the Center for American Progress. Schiffer argued in Just Security on 18 August that export controls can no longer decide who leads in AI. Chinese models took 41% of Hugging Face downloads over the past year, he wrote. Alibaba’s Qwen family alone has spawned more than 200,000 derivative models.

One point of precision matters here. Schiffer’s essay went up a day before the strategy was reported, and it never mentions the strategy. It is a general argument about American technology policy. Nextgov applied it to the document. His line about China’s open-weight advances exposing the limits of a denial strategy describes US policy as a whole.

What the document says about its own list

The strategy carries a caveat that cuts against reading the omission as a decision. Each critical technology area “includes a set of subfields that illustrate its scope but are not meant to be comprehensive”, it states. It offers the appendix as an update to an existing list, not a closed one.

A further step is promised too. For the technologies in the appendix, OSTP and the National Security Council will coordinate technology-specific strategies “at the next level of detail”. An AI plan is therefore still to come. Open weights could appear there.

The policy fight underneath all this remains open. Anthropic has pressed for tighter scrutiny of open-weight models and for curbs on chip sales to China. American startups have been fighting back against proposed restrictions. The administration also has form for keeping its AI positions unpublished, as it did with its AI framework this month.

What deployed units actually need

A separate constraint sits underneath the debate. Jake Steckler of GovAI wrote in a Carnegie Endowment paper on 10 August that “Ukraine’s computational architecture, consisting of Western cloud access, domestic data centers, and forward-deployed compute nodes, is already straining as it integrates more AI into targeting and coordination”. He wants funding to prioritise “denied environments without cloud compute access”.

Steckler puts numbers on the hardware limits. The best object detection models running on edge devices perform roughly 30% worse than their laboratory versions, he found. The Pentagon’s latest budget request seeks $4.2bn for sovereign AI infrastructure, and Congress has not approved it. His paper does not discuss open weights either. Its subject is compute and connectivity.

The public, and the allies

Public opinion runs against the infrastructure this strategy depends on. An Annenberg Public Policy Center survey polled 1,320 US adults between 16 June and 19 July. Some 18% said AI’s effect on the country over the next decade would be positive, against 39% who said negative.

Opposition to a new local data centre hit 61%, up 12 points since the spring, which is why big tech launched a charm offensive.

Europe barely features. The strategy names Australia and the United Kingdom as partners on critical technologies, critical minerals and the submarine industrial base. It refers to bilateral technology prosperity deals. It does not mention the European Union once.

France and Germany are meanwhile funding a European Palantir rival, on the argument that sovereign military software should not be American. The AI plan OSTP still owes will show whether open weights were left out on purpose.