AI for Science: The Next Geopolitical Battleground

How the U.S., U.K., and China are orienting national AI policy toward AI for Science

Over the past several years, the “AGI race” has become an international geopolitical competition. The U.S. use of chip export controls and increasing international interest in ‘sovereign AI’ matched Silicon Valley’s fever-pitch warnings that these near-god technologies might end the world.

But as the AGI bubble has matured into a more product- and returns-oriented phase, AI companies have begun to shift their focus to actual practical applications.

One of the clear beneficiaries of this shift is AI for Science. Just in the past year at OpenAI and Anthropic:

• Sam Altman has said the AI application he is most excited about over a 10-year horizon is AI for Science

• Kevin Weil shifted from CPO to head of AI for Science projects

• Liam Fedus, former OpenAI VP of post-training, left to co-found an autonomous lab for scientific discovery company

• Anthropic launched Claude for Life Services last month

• Anthropic is now hiring an AI for Science writer, along with an AI for economics writer.

This is a remarkable vibe shift from a few years ago, when AI for Science was a mere pit stop on the destination to AGI. Now, AI for Science is the goal — from Anthropic’s editorial perspective, the entire economy and science are roughly co-equal in focus.

Governments are now undergoing the same realignment. As a result, Q4 2025 has produced major AI-for-Science announcements from the U.S., U.K., and China.¹ What follows is my summary of these national strategies on AI for Science — and my assessment of where each country now stands.

U.S.

The Trump White House recently released an executive order (EO) on “Launching the Mission Genesis”, which tasks the Department of Energy (DOE) with marshalling its 17 national labs, supercomputers, datasets, and scientific infrastructure to build AI for Science foundation models and robotic labs, in partnership with private partners.² It is notable that DOE, rather than the National Science Foundation, is taking the lead; equally notable is that the EO provides no meaningful funding. This EO follows on the AI Action Plan published earlier this year, which included several notable references to programmable cloud labs³ and AI for semiconductor R&D.

Despite the ongoing and unprecedented cuts to federal science funding, the U.S. continues to benefit from its world-leading AI talent and private capital markets. Particularly in AI for Science, these are often in partnership with public science institutions. A few examples just from this past year:

• Deepmind and National Hurricane Center partnered to deploy Deepmind’s weather forecasting model, which outperformed even human experts and ensemble models at predicting the trajectory of Hurricane Melissa

• Meta, Lawrence Berkeley National Lab, and Lawrence Livermore National Lab partnered together to produce the Open Molecules 25 computational dataset, a massive Density Functional Theory (DFT) computational materials dataset which has been described as like “AlphaFold for materials science”.

• And as I wrote about earlier this year, the U.S. has seen $1B in private venture capital investment in developing autonomous labs

Key Takeaway: The race remains the U.S.’s to lose. Despite federal science funding cuts, immigration rollbacks, and an exodus of agency staff, the U.S. still retains deep capital markets and extraordinary talent and entrepreneurship. The Trump administration has taken several steps to emphasize AI for Science — including the Genesis EO — but whether these efforts materialize into meaningful substance remains unclear, given broader policy headwinds and acute staffing and funding constraints across agencies

U.K.

Just a few days before the Trump White House released the Mission Genesis EO, the U.K. Department for Science, Innovation, and Technology (DSIT) released their AI for Science strategy.⁴ Compared to the White House EO, the UK DSIT strategy is far more comprehensive and is accompanied with £137m in actual funding. I also found it notable the UK DSIT strategy explicitly couples the AI for Science strategy with the UK Modern Industrial Strategy, demonstrating at least greater awareness of the importance of technology development and industrial competitiveness.

Overall, the pace of U.K. investment and action on AI for Science, and autonomous labs, has been remarkably impressive. In July 2025, I wrote a profile on 5 things the U.K. is doing on AI for Science and they have continued to set the pace throughout the year. Their AI for Science push is part of a broader whole-of-government push for AI, including AI growth zones and investments in compute access and AI hardware innovation⁵.

Just in the 4 months since I published my profile in July, the U.K. has also:

• UK DSIT Sovereign AI unit released an open call for autonomous lab market engagements.

• ARIA also recently closed a funding call for “AI scientists”, which funded 9-month sprints to derisk early-stage technologies around AI-powered scientific discovery

• Commissioned Isambard-AI, a new supercomputer at University of Bristol, equipped with 5.4k GH200’s, making it 10x faster than the next-fastest supercomputer in the U.K. and #11 on the Top500 supercomputer list⁶

Key Takeaway: In hindsight, I believe we will look back upon 2023-2025 as the beginning of a remarkable investment from the U.K. in revitalizing their scientific apparatus, with the creation of ARIA and DSIT backed with actual public capital to invest in transformative projects. But the U.K. lacks the same structural advantages the U.S. enjoys and faces broader macroeconomic headwinds. On AI for Science specifically, limited compute and a far smaller talent pool than the U.S. or China will continue to constrain progress,⁷ even as public sector investment accelerates.

China

Two months ago, China released their AI+ diffusion plan [see Jeff Ding and Matt Sheehan’s far more in-depth coverage], which included AI for Science and Technology as one of the 6 key sectors for diffusion. Claude-translation [h/t Matt Sheehan] of that particular section:

(1) “AI+” Science and Technology

1. Accelerate the process of scientific discovery. Speed up exploration of new scientific research paradigms driven by AI, and accelerate the process of major scientific discoveries “from 0 to 1.” Accelerate the construction and application of scientific large models, promote the intelligent upgrading of basic research platforms and major science and technology infrastructure, build open and shared high-quality scientific datasets, and improve the ability to process complex multimodal scientific data. Strengthen AI’s role as a cross-disciplinary driver, and promote the integrated development of multiple disciplines.

2. Drive innovation in R&D models and efficiency improvement. Promote the integrated and coordinated development of AI-driven technology research and development, engineering implementation, and product deployment; accelerate the landing and iterative breakthroughs of technologies “from 1 to N”; and promote the efficient transformation of innovative achievements. Support the promotion and application of intelligent R&D tools and platforms, strengthen collaborative innovation between AI and areas such as biomanufacturing, quantum technology, and sixth-generation mobile communications (6G), use new scientific research achievements to support scenario applications, and use new application needs to drive breakthroughs in scientific and technological innovation.

3. Innovate methods of research in philosophy and the social sciences. Promote a shift in research methods in philosophy and social sciences toward human–machine collaborative models, explore new organizational forms of philosophy and social science research suited to the AI era, and broaden research horizons and perspectives of observation. Conduct in-depth research on the deep impacts and mechanisms of AI on human cognition and judgment, ethical norms, and other aspects, explore the formation of a theoretical system of “AI for good,” and promote AI’s better service to humanity.

China’s policy ecosystem is more bottom-up: central government sets high-level targets (e.g., 70% AI penetration in each sector by 2027) and delegates actual implementation to provincial and municipal governments. Case in point: Beijing municipal government got a head start, publishing their municipal AI for Science Strategy in July 2025. [Claude translation here]

Also notable is the now hallmark recognition by the Chinese government of pairing both “from 0 to 1” and “from 1 to N” innovation, diffusion being a demonstrated strength of theirs the past few decades in automotive and energy manufacturing. And of course, “AI for Science” is not only the direct product of “AI for Science” efforts but of broader macro considerations. In this case, China seeks to take advantage of the U.S. exodus of scientific talent through their new K-Visa for STEM talent, even despite local opposition.

Key Takeaway: Given the language and cultural barrier, along with a unique political system, China is always more opaque and difficult to grok. But it is undeniable that their patient investment in science talent are bearing fruit — they are now home to the majority of the top 10 research universities and have an almost spontaneous flourishing of competitive open source AI models. I have a much fuzzier sense of what is happening specifically in China on AI for Science (someone should really dig into this!), but they have all the ingredients for success.

---

I will close with a reminder of what I view to be the real lesson of AlphaFold: patient federal investment in the 1990s created an open protein-structure dataset that scientists — especially in the U.S. — contributed to for decades. It was only a few years ago that the value of this dataset was realized by another U.S. company to create AlphaFold. That story is still unfolding but it is worth remembering that the seeds for it were planted decades ago by public science funding.

Now, more than ever, public science funding is critical. The recent history of AI for Science shows that private companies rarely generate their own scientific datasets; they rely on them as public goods. Building the datasets needed for AI foundation models and adapting existing scientific infrastructure — from autonomous labs to high-performance computing, beamlines, and particle accelerators — is the essential task for nation states. Just as with the dawn of the electronic computer, the race is very much afoot: countries that move decisively now will define scientific and technological leadership in the midst of this historic disruption.

---

Addendum on the hermeneutics of government plans

Having spent some time in government, I unfortunately am cursed with too many opinions on government strategies and reports.

To be clear: the posting of a document on a white house website, or agency website, does not mean “Mission Accomplished”.

A report or executive order or strategy is meaningful insofar as it represents a policy consensus on a particular set of directions or actions, and actually empowers agency staff to actually carry out activities listed in the order. One astute observer of government bureaucracy referred to these orders as “hunting licenses” — but getting a license is merely the first step, and usually is insufficient without funding. A simpler way to put it is, “It’s an implementation game” (and a money one).

Each national strategy exists in the unique political economy of each country. By themselves, they merely suggest particular focuses or directions. Hence why I try to present the broader context that those strategies are placed in, both in terms of the policy environment and broader macro conditions for each country. Ultimately, real progress will proceed slowly, invisibly through the harnessing of infrastructure and bureaucracies to create public value through datasets, compute, or other investments.

---

Addendum — reflections on AI for Science

Almost exactly 6 years ago, I started this somewhat wonky substack⁸ as an aspirational Berkeley undergrad researcher as a way to force myself to read more research papers on how AI was being used in various scientific fields. Several jobs, writing hiatuses, and model releases later, AI for Science is now at the forefront of national science and tech policy and geopolitical competition. It has been an unexpected journey in many ways, but I feel fortunate to be able to play a small role in how the field progresses.

1. And as I have written about previously, this is not the first time nation states have raced to leverage and deploy a general purpose technology to create public scientific advantage.

2. I wrote a twitter summary of my takeaways on the Genesis EO.

3. which is perhaps the only Trump administration AI for Science announcement which has actual funding tied to it thus far - $100M from NSF for programmable cloud labs.

4. disclaimer: I am a fellow at Renaissance Philanthropy, which is supporting the implementation of the UK DSIT AI for Science strategy, as publicly disclosed in the release.

5. The simultaneous focus on supporting new domestic entrants in AI hardware, while also trying to increase access, is partially reminscent of the U.K.’s failed attempt to build an IBM competitor in the 1950’s.

6. U.S. national lab supercomputers occupy #1, #2, and #3 spots on the Top500 list.

7. and were what doomed the U.K.’s efforts to build globally competitive numerical weather forecasting models in the 1950’s as well.

8. I am in the market for a better substack name if anyone has recommendations