Michael Kratsios in conversation with Santi Ruiz

Episode Overview

• Explores the historical shift in American innovation where private sector R&D spending now vastly exceeds government funding, changing the federal role from "funder" to "coordinator."
• Frames the current AI revolution not as a software challenge, but as a physical infrastructure crisis requiring massive investments in energy, data centers, and land use.
• Details the U.S. geopolitical strategy of exporting the entire "tech stack" (chips, models, and apps) to ensure global allies rely on American rather than Chinese infrastructure.
• Examines internal government levers for accelerating science, including the use of federal lands for data centers and training "Science LLMs" on proprietary government data.

Key Concepts

• The Inversion of Innovation Funding In the Apollo era, the federal government funded the majority of R&D. Today, the private sector dominates spending. This fundamental shift means effective policy cannot rely on "buying" innovation. Instead, the government must act as an accelerator by removing regulatory barriers and coordinating private capital toward national interests.

• AI as an Energy and Land Challenge While often discussed as code, the constraint on AI progress is physical. Success depends on the capacity to build data centers and generate electricity. Consequently, technology policy has morphed into infrastructure and land-use policy. To address this, the government is using levers like permitting reform and offering federal land (specifically Department of Energy sites) to host data centers.

• The "Federated" Science Model Unlike nations with a centralized "Ministry of Science," the U.S. uses a decentralized model where power is spread across agencies (DOE, NSF, NASA). The Office of Science and Technology Policy (OSTP) lacks budget authority but uses "convening power" to align these disparate agencies. Success depends on high-level political will to force these agencies to prioritize shared goals like AI infrastructure.

• Geopolitics of the "Full Stack" National security now relies on which country's technology stack the world adopts. The U.S. strategy involves "Sovereignty via Interdependence": exporting the underlying American infrastructure (chips and base models) so that other nations can build their own local applications on top of it. This allows allies to maintain cultural sovereignty while remaining structurally aligned with the U.S. ecosystem rather than China's.

• The "Science LLM" Opportunity Commercial AI models are trained on the open internet. However, the federal government possesses vast silos of non-public, high-quality scientific data (material science, weather, physics) within National Labs. A major untapped opportunity lies in training specialized AI models on this proprietary data to accelerate scientific discovery, rather than just consumer applications.

Quotes

• At 3:44 - "If you go back to the era of... post-World War II, the vast, vast majority of R&D was being funded by the federal government. And over the last 70 years, we've had this inversion or this flip where now the majority of R&D is funded in the private sector." - Explaining why government strategy must shift from spending money to clearing paths for private capital.

• At 6:33 - "This revolution is going to be powered by electricity and by data centers, and we have to have those built in the United States as quickly as humanly possible." - Identifying the critical physical bottleneck that threatens to stall software advancement.

• At 10:30 - "We actually do have the very best technology... We have this window of time... where we truly can be... the single powerful supplier of the totality of the stack." - Highlighting the unique strategic opportunity the U.S. has to dominate the global market before competitors catch up.

• At 17:16 - "We have our own real estate. We have federal lands which can be used for build-outs themselves... The Department of Energy has already announced four locations for build-outs of data centers." - revealing a specific, underutilized government lever to bypass private market real estate friction.

• At 29:16 - "We have essentially been on kind of this autopilot mode where the same methodologies of choosing which grants and who to fund have essentially been stagnant for many, many years." - Critiquing the outdated, university-centric mechanisms of distributing scientific funding.

• At 35:09 - "The Science LLMs need to be trained on science data. And that's far more siloed and not necessarily in the public domain... there's data there that can be tapped into to drive this science." - Identifying the government's specific comparative advantage in the AI era: access to proprietary scientific datasets.

Takeaways

• Shift Policy Focus from "Protect" to "Promote" Do not rely solely on sanctions or export controls to win geopolitical competition. Actively flood the global market with cost-competitive, superior American technology stacks to ensure developing nations adopt U.S. standards rather than Chinese alternatives.

• Treat AI as an Infrastructure Project Stop viewing AI development solely through the lens of software regulation. Prioritize deregulation of energy production, grid expansion, and land permitting to support the physical reality of data centers required for the next generation of models.

• Modernize Scientific Funding Structures Move scientific funding off "autopilot." Diversify beyond traditional university grants by funding Focused Research Organizations (FROs) and fast-tracking initiatives that utilize government data silos to train specialized "Science LLMs."