Energy Bottlenecks and the Physical Breakthrough of Uranium Enrichment · Scott Nolan

2026-06-09 · A faithful, transcript-grounded reading by PodLens

Original episode:https://youtu.be/wisccQYTRQc?si=q8SYAbcYN-yI0qau　·　Timestamps are clickable — they seek the player in place

What This Episode Is About

General Matter co-founder and CEO Scott Nolan and host Anj dive deep into the energy bottlenecks behind AI compute expansion, the central role of nuclear energy (especially the uranium enrichment phase) in baseload power, and the systematic path of hard tech entrepreneurship. Drawing on his early engineering experience at SpaceX and over a decade of hard tech investing experience at Founders Fund, Scott Nolan analyzes the unprecedented security crisis the United States faces in the nuclear fuel supply chain (especially uranium enrichment): domestic capacity is near zero, with a high reliance on Russia and Europe. The conversation focuses on the structural conflict between the super-linear growth of compute demand and the two-decade stagnation of the US power grid, unpacking the paradigm shift from "stranded energy utilization" (such as Crusoe's early evolution from Bitcoin mining to AI data centers) to high-density nuclear baseload power. In addition, he discusses the level of government support in rebuilding critical infrastructure, the historical path dependency after the US-Soviet Cold War (such as the devastating impact of the Megatons to Megawatts program on the domestic US uranium enrichment industry), and provides a forward-looking assessment of the commercialization timeline for space orbital energy, AI-assisted engineering design, and next-generation Small Modular Reactors (SMRs).

Timeline Topic Map

• [00:08-01:14] Host Anj introduces today's guest Scott Nolan, introducing the closed loop of the AI compute flywheel from pre-training to deployment, and pointing out that today's focus is the energy bottleneck.
• [01:15-02:37] Anj explains the system stack perspective, proposing that traditional hierarchical structures are too rigid, and uses the systematic picture of the "AI manufacturing pipeline" to show the constraining relationship between computing and energy.
• [02:38-03:56] Unpacking the supporting systems outside AI compute centers, pointing out the two-year compute and power grid supply pressure caused by the explosion of ChatGPT (late 2022), emphasizing that the power grid has more physical construction urgency compared to data centers.
• [03:57-04:59] Tracing how breakthroughs in the AI application layer (such as the explosion of enterprise applications triggered by the release of Claude 4.6 in late 2025) cause a new round of compute and energy shortages, recreating a Groundhog Day-style cycle.
• [05:00-06:02] Scott Nolan introduces his educational background (Cornell Mechanical/Aerospace, Stanford MBA), his early experience as a propulsion systems engineer at SpaceX, and his decade at Founders Fund focusing on hard tech investing.
• [06:03-07:05] Scott Nolan recalls his experience with nuclear energy startups at Founders Fund, discovering that starting in 2020, all nuclear projects faced a bottleneck crisis of "fuel must be purchased from Russia," prompting him to deeply dissect this supply chain gap in 2023.
• [07:06-08:35] Anj and Scott reference statements from industry leaders (such as OpenAI's Sam Altman's "ultimate cost converges on energy" at a congressional hearing, Jensen Huang's admission on the Joe Rogan podcast that energy is the ultimate bottleneck, and Elon Musk's emphasis on energy).
• [08:36-09:37] Comparative analysis of the 1 Terawatt AI power demand forecast, pointing out that the US has been in near-stagnation in power grid construction over the past 20 years and must shift toward a China-style "near-vertical" slope of power grid expansion.
• [09:38-11:05] Discussing the definition and evolution of stranded energy (such as remote hydropower, isolated geothermal, or West Texas wind power). Explaining the historical role of Bitcoin mining as a "dress rehearsal" for AI infrastructure construction, and using Crusoe's Stargate project as an example to analyze how it converted the infrastructure accumulation of early Bitcoin mining farms into AI data centers.
• [11:06-12:30] Discussing the decisive impact of baseload power on data center uptime. Analyzing the issue of excessively high costs for wind/solar paired with batteries, and the supply chain congestion caused by multi-year lead times for natural gas turbines.
• [12:31-13:37] Comparing the safety and carbon emission metrics of various energy forms, pointing out the historical statistical advantages of nuclear in low carbon and safety (tied for first with wind energy), and arguing why nuclear energy is the ultimate long-term scaling limit for hyperscalers.
• [13:38-14:59] Unpacking the five steps of the nuclear fuel cycle (Mining -> Conversion -> Enrichment -> Re-solidification -> Pellet fabrication), pointing out that the US market share in Enrichment, the core high-value step, is below 0.1%, relying heavily on imports from Europe and Russia.
• [15:00-16:00] Scott Nolan introduces the founding journey of General Matter: officially established in January 2024, securing $900 million in orders from the DOE within 24 months.
• [16:01-17:05] Detailing the organizational building of General Matter's early startup phase, how its funding and team (including multidisciplinary backgrounds from Tesla, SpaceX, and national laboratories) completed long-cycle planning with high-intensity 100-hour work weeks, and selected the undeveloped land south of the 100-acre old enrichment plant site in Paducah, Kentucky.
• [17:06-18:05] Exploring the federal government's actual level of support in driving energy infrastructure, explaining the policy continuity since the Biden administration, and refuting the stereotype that "the government does not care about science and engineering."
• [18:06-19:05] Predicting the hundreds to thousands of high-tech and manufacturing jobs General Matter will create in California and Kentucky over the next 4 years, and exploring how AI compute demand is reversely reviving physical industries and employment.
• [19:06-20:05] Q&A 1: How to accelerate the deployment of nuclear energy before 2034. Scott Nolan expects capacity to come online and scale rapidly before the end of 2030, believing the early 2030s will be the true hockey stick starting point for commercial nuclear deployment.
• [20:06-21:10] Q&A 2: Exploring the feasibility of space orbital energy solutions. Scott Nolan points out that SpaceX, with its unparalleled launch volume, is expected to solve space power supply through geostationary orbit data center satellites, but this will be a unique advantage for SpaceX, while other players will still face a tough battle in ground infrastructure.
• [21:11-22:20] Q&A 3: Recalling Scott Nolan's internship and full-time experience at SpaceX during its early 35-person phase (troubleshooting two Falcon 1 launches—the first due to an aluminum nut cracking on a fitting causing a fuel leak and fire, the second due to second-stage propellant slosh without enough designed baffles causing loss of control), sharing the lesson of why he chose to leave for business school at the 100-person scale due to misjudging it as "not scrappy enough," warning against underestimating the value of system-level organizational building from 100 to 1000 people.
• [22:21-23:18] Q&A 4: Reflections on the European nuclear energy path. Using Germany's closure of perfectly functioning nuclear power plants—which led to the vacant baseload being filled by coal and natural gas, causing massive carbon emissions and air pollution (forming a red-and-blue contrast of clean zones with France)—as an example to reveal the catastrophic negative corrections of policy and populism on clean energy.
• [23:19-24:18] Q&A 5: Geopolitics and ore source distribution. Kazakhstan accounts for 40% of global uranium ore production. Scott Nolan calls for the US to not only achieve independence in enrichment but also serve as a low-cost supply hub for allies, thereby reducing the geopolitical proliferation risks of small nations enriching nuclear fuel on their own.
• [24:19-24:48] Anj summarizes and expresses eager anticipation for next-generation nuclear construction.

Core Viewpoints List

1. The ultimate limiting factor for AI compute expansion will be the cost of electrical energy. Chip manufacturing costs and model training costs will continue to fall with technological iterations, but the electricity consumed to run these massive models is a physical necessity that cannot be completely erased by algorithms. All compute competition will ultimately converge on the competition for energy costs. [09:11-09:53] | Type: Viewpoint
2. Bitcoin mining is essentially a "dress rehearsal" for AI data center infrastructure construction. The technical and operational experience accumulated by Bitcoin miners in their early days—finding remote stranded energy, designing on-site high-density cooling systems, and distributed power capture—has been rapidly repurposed as data center infrastructure in the AI era, proving that an seemingly bubbly industry can still deposit real infrastructure assets. [22:40-23:00] | Type: Viewpoint
3. Wind and solar power, due to the lack of low-cost grid-scale storage, cannot meet the high uptime requirements of AI data centers for baseload power. Elevating unstable renewable energy to the uptime levels required by data centers through batteries is extremely expensive and unrealistic under current technology and cost structures. [16:15-16:29] | Type: Viewpoint
4. Nuclear energy is the ultimate long-term solution for hyperscalers under the requirements of safe and low-emission baseload power. Statistical data shows that nuclear energy has extremely low carbon emissions throughout its life cycle, and its safety is tied for first with wind energy, making it the inevitable destination for tech giants seeking stable, clean power. [17:21-18:04] | Type: Fact
5. The United States currently has a fatal supply chain gap in uranium enrichment, the core high-value step of the nuclear fuel supply chain. The US domestic market share of global uranium enrichment has fallen below 0.1%, making it completely unable to provide fuel at scale for existing and future reactors, leaving the US passively dependent on import channels from Russia and Europe. [19:16-19:37] | Type: Fact
6. The post-Cold War "Megatons to Megawatts" program, while providing cheap nuclear fuel, accidentally destroyed the domestic US uranium enrichment industry. The commercial operation of down-blending decommissioned Russian nuclear warheads and introducing them to the US market at low prices wiped out the survival space for private domestic US enrichment plants, leading to the closure of the last domestic commercial enrichment plant in 2013. [58:04-58:43] | Type: Fact
7. Space orbital data centers are a unique path that only SpaceX, with its massive and extremely cheap launch capability, can attempt. Despite technical challenges in space heat dissipation and energy transmission, directly utilizing space solar energy via geostationary orbit satellites is a higher-dimensional attempt to bypass ground power grid interconnection delays and regulatory quagmires. [44:48-45:13] | Type: Prediction
8. When entering stagnant and capital-intensive traditional heavy industries, hardware startups must introduce Tesla and SpaceX-style agile development and system integration genes. Traditional contractor models and conservative design chains cannot cope with rapid iteration pressures; hands-on systems engineering teams are the core of accelerating physical world infrastructure. [35:38-35:53] | Type: Viewpoint
9. The expansion of AI compute is not merely destroying jobs; instead, it is reversely triggering a wave of employment in high-tech manufacturing and industrial infrastructure in the physical world. To unblock compute bottlenecks, thousands of high-income physical industry jobs are actually being created in California design and R&D roles as well as Kentucky ground construction plans. [40:40-41:35] | Type: Prediction

Plain English Retelling

Let's talk about Scott Nolan's share at Stanford CS153. Many people discuss how much model capabilities have updated every day, like how the recent Claude 4.6 has changed workflows, but few calculate the account from a physical and infrastructure level: when intelligence becomes as ubiquitous as air, can the underlying physical world actually support it?

Scott Nolan points out a cold reality ignored by most: the ultimate bottleneck of AI is neither algorithms nor even just chips, but the most basic electricity. The expansion of large language models and compute centers is super-linear, but the construction of power grids cannot be completed overnight. In the years since the explosion of ChatGPT, the supply chain has remained extremely tight. Because grid expansion takes time, even if you buy tens of thousands of GPUs and build the data centers, if there are no high-voltage lines to pull the power, it's just a pile of scrap metal.

In the early years, people solved this problem by looking for "stranded energy". For example, West Texas has a lot of wind but no residents, or abandoned hydropower stations in remote mountains; the electricity in these places couldn't be integrated into the grid and transported out, so it was wasted in the past. Thus, Bitcoin miners ran over to build mining farms, consuming this cheap electricity on-site. This process was actually a "dress rehearsal" for AI data centers. Companies like Crusoe started by helping people mine, accumulating the ability to capture power on-site, design high-density compute racks, and network transmission, and then naturally transitioned into AI data center service providers.

But now, this cheap "stranded energy" has been completely snatched up. For AI compute to continue to multiply, it must rely on stable, 24/7 uninterrupted "baseload power". Wind and solar are simply unusable without cheap energy storage batteries. Natural gas power generation is the current emergency solution, but this has led to orders for power generation turbines being backlogged for several years, with manufacturers simply unable to produce them in time.

Therefore, in the long run, there is only one way—nuclear energy. Nuclear energy is both clean and safe. But the awkward thing is, although the US wants to develop nuclear power on a large scale, it finds that it can't even manufacture nuclear fuel anymore. Nuclear fuel production goes through five steps: mining, conversion, enrichment, reduction, and pellet fabrication, and the US's global market share in the most critical and high-tech "enrichment" phase is actually below 0.1%. Existing nuclear power plants and future advanced Small Modular Reactors (SMRs) are all counting on Russia and Europe to ship fuel.

How did this fatal wound come about? It was actually a "path dependency" after the Cold War. When the Soviet Union collapsed, the US and Soviet Union set up the "Megatons to Megawatts" program, down-blending highly enriched uranium from decommissioned Soviet nuclear warheads into ordinary fuel to sell to the US for power generation. This was an excellent peace dividend at the time, cheap and easy. But it also drove all domestic US uranium enrichment plants into a dead end, because they simply couldn't compete commercially with free "decommissioned warheads". By 2013, the last commercial enrichment plant in the US was forced to close, completely severing domestic technology and capacity.

Scott Nolan founded General Matter precisely because he saw this fatal gap forgotten by everyone. They want to rebuild highly scalable uranium enrichment capacity south of the old plant site in Kentucky, and use a SpaceX-style clean sheet logic to design technology from scratch to bring costs down. He poached engineers from Tesla and SpaceX who were used to iterating quickly in the hardcore physical world, collaborated with the government and the DOE, and secured $900 million in contracts in just two years.

This conversation tells us that the future of AI is definitely not just in the virtual cloud. It is acting like a massive black hole, pulling Cold War legacies, nuclear energy supply chains, heavy industrial manufacturing, and space orbital launches all into this massive physical gravitational field.

Key Segments to Listen to

• [09:11-09:45] Scott Nolan explains the mechanism of "all compute costs ultimately converging on electricity costs," hitting the physical hard barrier of AI evolution directly.
• [11:52-12:28] Explaining the commercial essence of "stranded energy" and why Bitcoin mining is a great "dress rehearsal" for AI infrastructure construction. This breakdown holds extremely high methodological value.
• [19:16-19:37] Scott Nolan drops the shocking fact that the US global share in the uranium enrichment phase is below 0.1%, which is the core anchor point for understanding the nuclear energy supply chain crisis.
• [46:21-49:50] Recalling the specific engineering reasons for the two early SpaceX Falcon 1 launch failures (aluminum nut cracking and propellant slosh), and Scott's honest strategic misjudgment in leaving the company at the 100-person scale, which is a rare lesson in hard tech organizational development.
• [50:34-51:58] Unpacking the policy tragedy of Germany closing mature nuclear power plants, leading to the vacant baseload being filled by coal and natural gas, contrasting the epistemological bias against nuclear energy with data and intuition.

Resonances with past episodes

• Extension→ Product Building and Career Evolution in the AI Era · Nikhyl Singhal
  Extending the rule of 'iteration speed determines success or failure' from the software and internet industries to traditional heavy industries, pointing out that hardware startups also need to break the sluggishness of traditional contractor models through agile development and system integration.
  This[35:38-35:53] When entering stagnant and capital-intensive traditional heavy industries, hardware startups must introduce Tesla and SpaceX-style agile development and system integration genes. Traditional contractor models and conservative design chains cannot cope with rapid iteration pressures; hands-on systems engineering teams are the core of accelerating physical world infrastructure.

  Related`[13:54]` Product iteration speed (Iteration Speed) determines product success or failure more than the initial state, which constitutes the core advantage of startups against large companies.
• Extension→ System Design of Venture Capital and Paradigm Shifts in the Age of Intelligence · Ben Horowitz
  Extending the business competition logic of 'the physical supply chain is the real defense' to the macro dimension of national security and AI energy security, pointing out that the gap in the US uranium enrichment—the core link of the physical supply chain—has become the biggest weakness constraining the long-term energy security of its AI compute.
  This[19:16-19:37] The United States currently has a fatal supply chain gap in uranium enrichment, the core high-value step of the nuclear fuel supply chain. The US domestic market share of global uranium enrichment has fallen below 0.1%, making it completely unable to provide fuel at scale for existing and future reactors, leaving the US passively dependent on import channels from Russia and Europe.

  Related[21:53-22:10], [58:47-01:00:43] In the era of ubiquitous AI and the SaaSpocalypse, code and user interfaces are no longer defensive; the real defense is the physical supply chain and specific distribution channels.
• Corroboration→ System Design of Venture Capital and Paradigm Shifts in the Age of Intelligence · Ben Horowitz
  Jointly revealing that when government decision-making lacks industrial synergy and long-term strategic vision, it can deal a devastating blow to critical technology industries, thereby triggering profound geopolitical crises.
  This[58:04-58:43] The post-Cold War 'Megatons to Megawatts' program, while providing cheap nuclear fuel, accidentally destroyed the domestic US uranium enrichment industry. The commercial operation of down-blending decommissioned Russian nuclear warheads and introducing them to the US market at low prices wiped out the survival space for private domestic US enrichment plants, leading to the closure of the last domestic commercial enrichment plant in 2013.

  Related[51:58-52:45] The silence of the tech industry in Washington policymaking will bring unbearable geopolitical crises and regulatory regressions to the industry's development.
• Supplement→ System Design of Venture Capital and Paradigm Shifts in the Age of Intelligence · Ben Horowitz
  Addressing the obstruction of ground data center construction caused by regulatory overreach and grid interconnection delays, space orbital data centers provide a physical, higher-dimensional solution that bypasses ground geographical and political limitations.
  This[44:48-45:13] Space orbital data centers are a unique path that only SpaceX, with its massive and extremely cheap launch capability, can attempt. Despite technical challenges in space heat dissipation and energy transmission, directly utilizing space solar energy via geostationary orbit satellites is a higher-dimensional attempt to bypass ground power grid interconnection delays and regulatory quagmires.

  Related[01:05:16-01:06:11] The core danger of geopolitics in the AI era is that nations, out of fear, engage in defensive over-regulation, thereby losing the superintelligence arms race to China.
• Corroboration← The Discipline of Value Delivery per Gigawatt · Amin Vahdat
  Both point out that energy efficiency improvements at the algorithmic level cannot eliminate rigid bottlenecks in the physical world. The former explains through Jevons paradox that efficiency gains actually stimulate larger compute demands, while the latter points out from the perspective of physical limits that energy cost is the ultimate constraint on compute expansion that cannot be erased by algorithms.
  This[09:11-09:53] Chip manufacturing costs and model training costs will continue to decline with technological iterations, but the electricity consumed to run these massive models is a rigid physical demand that cannot be completely erased by algorithms; all compute competition will ultimately converge on the competition over energy costs.

  Related[58:40-01:00:05] Compute hardware will remain the primary bottleneck for the next 5 to 10 years. Any algorithmic breakthrough in energy efficiency will be rapidly consumed by new, more valuable compute demands due to Jevons paradox.

A faithful reconstruction and plain-language retelling of the episode, generated by PodLens.

This is one source-grounded reading, not a replacement for the original. Every point is anchored to its source, so you can check it yourself — and corrections are welcome.