The Pentagon shakes up acquisitions in push for adaptability, OpenAI formally restructures, and the “American AI Exports” program launches to mixed reviews

Subscribe to policy.ai
Updates

This is Alex Friedland’s last edition of policy.ai. CSET is grateful to Alex for his incredible work turning policy.ai into a go-to resource for policymakers, industry leaders, and others who want to stay up-to-date on the latest in AI, emerging technology, and national security policy. Policy.ai will be taking a break over the holidays, and we look forward to resuming in 2026.

   — Christa Bennett, Director of Communications and Strategic Engagement

Worth Knowing

OpenAI Completes Restructuring: OpenAI finalized its long-anticipated and legally fraught corporate overhaul late last month, transitioning from a nonprofit-controlled structure to a public benefit corporation. The restructuring gives the OpenAI Foundation (the renamed nonprofit organization) a roughly 26% stake (valued at approximately $130 billion) in the new for-profit entity, OpenAI Group PBC. As part of the deal, Microsoft, an early OpenAI investor, agreed to an altered partnership agreement, which had emerged as a potential liability for both parties. Under the new agreement, OpenAI can no longer unilaterally declare it has achieved AGI and cut off Microsoft’s access — any AGI declaration must now be verified by an independent expert panel — and OpenAI can collaborate with third parties on certain products, release open-source models that meet capability criteria, and use cloud providers beyond Azure for some workloads, though it committed to purchasing $250 billion in Azure services over an undisclosed period. Microsoft also gained explicit freedom to “independently pursue AGI” alone or with other partners. The restructuring received the go-ahead of the California and Delaware attorneys general — which had been seen as a potential roadblock — but is still likely to face scrutiny as part of an ongoing legal challenge from co-founder-turned-rival Elon Musk. While the company seems focused on making the most of its new structure (some observers are already talking about a $1 trillion IPO), there are still open questions about how much power (and appetite to use it) the non-profit OpenAI Foundation still has. The foundation has both a significant financial stake and the power to appoint and remove members of the for-profit entity’s board, but it remains to be seen whether it will use that power to push for change if the company’s pursuit of profits conflicts with the non-profit’s mission.
Potential ASML Rival Substrate Raises Funds (and Eyebrows): San Francisco-based Substrate has raised more than $100 million to develop what it claims is an alternative to the lithography machines that are essential for chipmaking. The company says it can use particle accelerators and X-rays to produce results comparable to the industry’s most advanced tools, but significant questions remain about whether the company, led by two brothers with no semiconductor industry experience, can deliver. Advanced lithography is extraordinarily complex; Netherlands-based ASML is the only supplier of the most advanced lithography machines, which use extreme ultraviolet light and cost upward of $400 million each. ASML’s technology is so essential that the United States has effectively blocked sales of its tools to China, a restriction that remains one of the primary reasons Chinese chipmakers lag behind their foreign rivals. Unseating ASML, therefore, would dramatically reshape both the semiconductor industry and potentially the geopolitical landscape. Substrate claims it has developed a lithography system that channels light from a particle accelerator to print features at the cutting edge — what the company says is the 2-nanometer process node. The startup has attracted backing from prominent investors, including Peter Thiel’s Founders Fund, General Catalyst, and In-Q-Tel, the CIA-backed venture capital firm. But serious doubts about the company’s technology, its founders, and its timelines remain. Still, the stakes are high enough — and the potential rewards significant enough — that Substrate and companies like it will remain attractive to investors.
Government Updates

Pentagon Launches Sweeping Acquisition Overhaul: Defense Secretary Pete Hegseth announced a major transformation of the Pentagon’s acquisition system in a speech at the National War College last week. The changes — detailed in a new Acquisition Transformation Strategy and three supporting memos issued the same day — aim to prioritize speed and adaptability in weapons buying. Hegseth said the Pentagon will embrace purchasing “the 85% solution” and iterating toward full capability rather than waiting years for perfect systems. The reforms introduce significant changes to the Pentagon’s acquisitions architecture. Among other things, they create new “Portfolio Acquisition Executives” with authority to shift money between related programs based on performance, a significant departure from the current system that ties funds to specific projects. The changes shouldn’t come as a surprise. Since President Trump took office earlier this year, a key focus of the Pentagon’s leadership has been on accelerating acquisition. But Hegseth’s speech and the accompanying memos are the clearest attempt yet to dismantle the established acquisitions processes. While the moves have been applauded as long overdue by some, they’re also likely to face difficulties and some pushback both from Congress — which may prefer to retain fine-grained control over defense acquisitions — and some of the large defense contractors that have benefited from the established system.  

Commerce Department Launches American AI Exports Program: Last month, the Commerce Department’s International Trade Administration announced the implementation of the “American AI Exports Program,” following President Trump’s July executive order on promoting exports of the “American AI Technology Stack.” The program aims to select industry-led consortia to export “full-stack” AI packages — encompassing hardware, software, models, and applications — to countries around the world. Approved packages will receive priority access to federal financing tools, including Export-Import Bank loans and International Development Finance Corporation guarantees. The rollout began with a Request for Information soliciting industry feedback on program design — including questions about what should constitute the AI “tech stack” and how consortia should be formed. Axios reported that some industry observers were skeptical of the rollout, calling it “extremely underwhelming” and “chaotic.” Some of the plan’s architects seem to agree the plan needs tweaking — Dean Ball, the former White House staffer who drafted the original executive order, argued the order may need to be updated; in a recent Substack post, he proposed allowing individual firms to demonstrate they can provide full-stack solutions through existing partnerships rather than forcing formal consortium arrangements. Ball argued this would “piggyback off existing corporate dealmaking efforts” and avoid unnecessary complexity. The RFI remains open until November 28, after which Commerce will incorporate feedback into the program’s final design.

DOE Announces Nine Supercomputers with Public-Private Funding: The Department of Energy unveiled plans for nine new supercomputers to be built across three national labs in partnership with leading tech companies. The systems — spread between Argonne, Oak Ridge, and Los Alamos National Labs — will vastly expand AI-focused computing capacity, with some coming online as soon as 2026. At Argonne, Solstice — a 100,000-GPU Nvidia system designed in partnership with Oracle — will be the largest AI supercomputer in the DOE lab system. A companion system, Equinox, will be operational by 2026 with 10,000 GPUs. At Oak Ridge, Lux (an AMD-powered system expected to come online in early 2026) and Discovery (an HPE-built system arriving in 2028 that will significantly outperform the lab’s current Frontier supercomputer). Los Alamos announced Mission and Vision, both slated for 2027 — Mission will support classified nuclear weapons simulations without live testing, while Vision will advance open AI research for national security applications. The initiative marks a shift in government buying away from traditional procurement toward public-private investment, with companies agreeing to co-fund the projects in exchange for computing access. Energy Secretary Chris Wright touted this as a “new commonsense approach” that brings capacity online more rapidly. The labs did not disclose exact funding breakdowns, and Los Alamos noted its systems depend on “anticipated future funding.”

What’s New at CSET

REPORTS

PUBLICATIONS AND PODCASTS

IN THE NEWS

What We’re Reading

Paper: Emergent Introspective Awareness in Large Language Models, Jack Lindsey, Anthropic (October 2025) 

Paper: The Smol Training Playbook: The Secrets to Building World-Class LLMs, Loubna Ben Allal, Lewis Tunstall, Nouamane Tazi, Elie Bakouch, Ed Beeching, Carlos Miguel Patiño, Clémentine Fourrier, Thibaud Frere, Anton Lozhkov, Colin Raffel, Leandro von Werra, and Thomas Wolf (October 2025) 

Paper: Towards a future space-based, highly scalable AI infrastructure system design, Blaise Agüera y Arcas, Travis Beals, Maria Biggs, Jessica V. Bloom, Thomas Fischbacher, Konstantin Gromov, Urs Köster, Rishiraj Pravahan, and James Manyika, Google (November 2025)