From Yellowcake to Gigawatts

2026-01-29T00:00:00+00:00

The hyperscalers need nuclear power. 10 GW of deals in 18 months. More capacity than the US has built in 30 years. But nuclear has a supply chain. Uranium gets mined, converted to gas, enriched, fabricated into fuel. Each stage has its own facilities, capacities, bottlenecks. The deals get covered. The fuel cycle doesn't.</p>

Demand Shock</h2>
Hyperscaler nuclear power demand is exploding. In 2025 the US data center load is around 61.8 GW ^{1</a></sup>.
This is up 22% from 2024, which had a load of ~50.5 GW.
Projections predict demand to go up to ~75.8 GW in 2026.</p>}
A single gigawatt is roughly the same amount of power needed to power 750,000 US homes, or a mid-sized American city 2</a></sup>. A single large nuclear reactor can produce a single gigawatt. To produce the same amount of energy with natural gas would require about 3 natural gas plants.</p>

The Scale of the Bet</h3>

The US operates 94 nuclear reactors with 97 GW of total capacity, the largest fleet in the world 3</a></sup>. Within 18 months hyperscalers have made 10 GW of nuclear deals which is more nuclear capacity contracted than the US has built in 30 years 4</a></sup>. These deals have been made by four companies. That's 10% of the entire existing US nuclear fleet contracted by four companies.</p>

The 2025 hyperscaler capital expenditure: ~$350-400 billion 5</a></sup>. Of that, ~75% is AI infrastructure: servers, chips, data centers 6</a></sup>. The capital expenditure is projected to total $1.15 trillion between 2025-2027 7</a></sup>. Three of these deals depend on fuel that doesn't exist at commercial scale.</p>

Big Four Commitments </strong> </figcaption>
Company</th> Partner</th> Type</th> Capacity</th> Term</th> Online</th></tr></thead>

Microsoft</td> Constellation</td> Restart</td> 835 MW</td> 20 yr</td> 2027</td></tr>
Amazon</td> Talen Energy</td> Existing</td> 1,920 MW</td> 17 yr</td> 2026-2032</td></tr>
Google</td> Kairos Power</td> SMR (HALEU)</td> 500 MW</td> —</td> 2030-2035</td></tr>
Meta</td> Constellation</td> Existing</td> 1,121 MW</td> 20 yr</td> 2027</td></tr>
Meta</td> Vistra</td> Existing + Uprates</td> 2,609 MW</td> 20 yr</td> 2026-2034</td></tr>
Meta</td> Oklo</td> SMR (HALEU)</td> 1,200 MW</td> —</td> 2030</td></tr>
Meta</td> TerraPower</td> SMR (HALEU)</td> 2,800 MW</td> —</td> 2032-2035</td></tr> </tbody></table>
~11GW of nuclear power demand </p> </figure>
Why Nuclear</h3>
Nuclear works for data centers because it runs constantly. According to the U.S. Energy Information Administration capacity factor data for 2024 ^{8</a></sup>, nuclear has the highest reliability by a large margin when compared to other energy sources.</p>}
RELIABILITY BY ENERGY SOURCE </span> </span>Nuclear █████████████████████████████████████████ 92% </span>Natural Gas █████████████████████████ 60% </span>Coal ███████████████ 42% </span>Wind ██████████ 33% </span>Solar █████ 23% </span></code></pre> Capacity factor measures actual output versus theoretical maximum over a year. A 1-gigawatt nuclear plant delivers 920 megawatts on average and is limited mainly by scheduled refueling and maintenance. A 1-gigawatt solar farm delivers 230 megawatts on average, limited by the fact that the sun doesn't shine at night or through clouds. For a data center that needs power every hour of every day, this isn't an environmental preference it's arithmetic.</p> Fuel Cycle 101</h2> Nuclear power has an elaborate supply chain with many components. Uranium doesn't arrive enriched. It goes through four stages before it becomes fuel.</p> Mining & Milling: Ore → Yellowcake</h3> The process starts with mining uranium ore which is about 0.1-20% uranium content depending on grade. Then you mill down the ore into a powder and purify it. This is commonly done at or near the mine site. You're extracting uranium from rock.</p> Once mined you crush the ore into fine particles. Then you dissolve the powder in sulfuric acid or alkaline carbonate solution and filter out the undissolved byproducts. Then do another purification step involving ion exchange on the solvent. Then finally you precipitate the solution by adding chemicals to drop uranium out of solution as a solid resulting in dry yields of yellowcake powder (U₃O₈).</p> Mining & Milling </strong> 62,000 tU/year global production Output: Yellowcake (U₃O₈)</span> </figcaption> Company</th> Country</th> Share</th></tr></thead> Kazatomprom</td> Kazakhstan</td> 23%</td></tr> Cameco</td> Canada</td> 15%</td></tr> Orano</td> Niger/Kazakhstan</td> 8%</td></tr> Uranium One</td> Russia/Kazakhstan</td> 5%</td></tr> CGN</td> China/Namibia</td> 5%</td></tr> Navoi Mining</td> Uzbekistan</td> 5%</td></tr> BHP</td> Australia</td> 4%</td></tr> Energy Fuels</td> USA</td> <1%</td></tr> </tbody></table> Top 3 countries: Kazakhstan 43%, Canada 7%, Namibia 7%* Supply adequate</p> </figure> The precipitate, a yellow coarse powder roughly 80% uranium by weight, is packed into drums as a standardized commodity for shipping. Yellowcake from Kazakhstan, Canada, and Namibia all goes into the same global supply chain. Gets shipped to conversion facilities (Port Hope, Tricastin, Metropolis) to become UF₆.</p> Conversion: Yellowcake → Uranium hexafluoride (UF₆)</h3> Yellowcake is powder. You can't enrich powder. Conversion turns it into uranium hexafluoride (UF₆), a gas that can be spun in centrifuges. The process runs through multiple stages: dissolving yellowcake in nitric acid, purifying the solution, reacting it with hydrofluoric acid to produce "green salt" (UF₄), then fluorinating that at 500°C to get UF₆. The purity requirements are extreme. Enrichment centrifuges are precision equipment and contamination destroys them.</p> Conversion </strong> 62,000 tU/year nameplate capacity Output: Uranium hexafluoride (UF₆)</span> </figcaption> Company</th> Facility</th> Capacity</th></tr></thead> Orano</td> France (Tricastin)</td> 15,000 tU</td></tr> Cameco</td> Canada (Port Hope)</td> 12,500 tU</td></tr> Rosatom</td> Russia</td> 12,500 tU</td></tr> CNNC</td> China</td> 15,000 tU</td></tr> ConverDyn</td> USA (Metropolis)</td> 7,000 tU</td></tr> </tbody></table> Western capacity: 34,500 tU Near capacity</p> </figure> Every metric tonne of U₃O₈ (yellowcake) outputs about ~0.98-0.99 metric tonnes of UF₆. By compound mass it's different (U₃O₈ is 85% uranium, UF₆ is 67% uranium), but the industry measures in tU to keep it comparable.</p> Enrichment</h3> Natural uranium is 0.7% U-235 (fissile) and 99.3% U-238 (not fissile). Reactors need higher U-235 concentration. The enrichment happens through a gas centrifuge process. The uranium hexafluoride is heated to gas, and fed into a cylindrical rotor. The rotor spins at 50,000-70,000 RPM which is a wall speed approaching Mach 2 (twice the speed of sound). The centrifugal force pushes heavier U-238 toward outer wall, lighter U-235 concentrates toward center. Countercurrent flow thermal gradient creates vertical circulation. Enriched gas is drawn from top-center, depleted gas from bottom-outside.</p> Separative Work Units (SWU)</h4> SWUs are a measure of the effort required to separate U-235 from U-238 in a centrifuge. You can think of it as a standard measure of enrichment. Enrichment companies sell SWU as a service. You give them UF₆, they spin it, you get enriched UF₆ back. Current spot price is ~$190/SWU.</p> The kicker here is that not all nuclear reactors require the same amount of enrichment. Traditional nuclear reactors require low-enriched uranium (LEU) which usually refers to 3-5% U-235. Most modern reactors and all of the small modular reactors (SMRs) require a much higher enrichment called High-Assay Low-Enriched Uranium (HALEU) referring to around 19.75% U-235 concentration. To produce 1 kg of 4% LEU requires ~4-5 SWU. To produce 1 kg of 19.75% HALEU requires ~15-20 SWU. Same machines, more work.</p> Depleted uranium (mostly U-238) is the byproduct. Operators choose how much U-235 to leave behind, lower tails = more product extracted but more SWU required. There is an economic tradeoff based on uranium vs SWU prices. Modern centrifuges are carbon fiber rotors, magnetic bearings, operate for decades. Rosatom, Urenco, Orano, CNNC are the major operators. Centrus is the only US enricher, currently demo-scale for non-standard enrichment.</p> The output is enriched UF₆ in cylinders. 3-5% concentration for conventional reactors. 5-19.75% for HALEU (advanced reactors). Above 20% is weapons-grade (HEU) which is illegal for commercial use.</p> Enrichment </strong> 65M SWU/year global capacity — Output: Enriched UF₆ (3-5% or 5-20% U-235)</span> </figcaption> Company</th> Country</th> Share</th></tr></thead> Rosatom (TENEX)</td> Russia</td> 46%</td></tr> Urenco</td> UK/DE/NL/USA</td> 29%</td></tr> CNNC</td> China</td> 12%</td></tr> Orano</td> France</td> 10%</td></tr> Centrus</td> USA (Piketon)</td> <1%</td></tr> </tbody></table> US imports 25% of enrichment from Russia (ban 2028)* — Russia controls 46%</p> </figure> In enrichment you lose a lot of mass depending on the U-235 concentration required. More U-235 concentration correlates to needing more input.</p> Enrichment Requirements </strong> </figcaption> Reactor</th> Developer</th> Fuel Type</th> Enrichment</th></tr></thead> Natrium</td> TerraPower</td> Metallic</td> ~19.5%</td></tr> KP-FHR (Hermes)</td> Kairos</td> TRISO pebbles</td> ~19.75%</td></tr> Xe-100</td> X-energy</td> TRISO pebbles</td> HALEU (5-20%)</td></tr> Aurora</td> Oklo</td> Metallic</td> HALEU</td></tr> VOYGR</td> NuScale</td> Conventional UO₂</td> 4.95%</td></tr> </tbody></table> SMRs need higher enrichment </p> </figure> HALEU requires 5x more natural uranium and 7x more enrichment work than conventional fuel. That 7x multiplier shows up twice: enrichment costs ~$7,000/kg versus ~$1,000/kg, and global centrifuge capacity (65 million SWU/year) yields 1.6 million kg of HALEU versus 11 million kg of conventional fuel. The infrastructure is the same. The throughput isn't.</p> </p> Chokepoints</h2> The fuel cycle looks linear on paper. In practice, it has three points where demand stacks up against fixed capacity. Each one matters differently depending on which nuclear story you're tracking. Importantly there is no chokepoint on uranium ore or yellowcake.</p> Chokepoint 1: Enrichment</h3> The Russian state-owned company TENEX produces 46% of global enrichment capacity ^{9</a></sup>. Combined Western capacity is about 54% but geographically constrained. Centrus is the only US-based company but produces less than 1% of global enrichment (demo-scale only). The US imports roughly 20-25% of its enrichment from Russia 10</a></sup>, earning Rosatom ~$1 billion/year from US uranium exports.</p>} On May 13, 2024, President Biden signed H.R. 1042, the Prohibiting Russian Uranium Imports Act 11</a></sup>. The law bans unirradiated low-enriched uranium from Russia effective August 11, 2024. DOE can issue waivers through January 1, 2028, but all waivers terminate on that date 12</a></sup>. From 2028 through 2040, no Russian LEU enters the US market.</p> This gives utilities roughly two years to secure alternative supply. Historically new centrifuge cascades take 8-10 years to build and license 13</a></sup>. Urenco's New Mexico expansion took 7 years from groundbreaking to full capacity. No Western company has announced capacity additions at the scale needed to replace Russian supply. The US-based enrichment facility Centrus has a backlog totaling $3.9 billion through 2040 14</a></sup>. Spot SWU prices near historic highs at ~$188/SWU 15</a></sup>, with limited liquidity.</p> Chokepoint 2: Conversion</h3> Conversion is the forgotten middle child of the nuclear fuel cycle. Everyone talks about uranium mining and enrichment. Almost nobody talks about converting yellowcake powder into gas.</p> The current Western conversion capacity is about 34,500 tU/year. The facilities administering this conversion are running near capacity. Only three Western facilities: Orano (Tricastin), Cameco (Port Hope), ConverDyn (Metropolis). This was the binding constraint before enrichment tightened.</p> In November 2017, Honeywell placed its Metropolis Works facility, the only US conversion plant, on idle standby due to an oversupplied UF₆ market 16</a></sup>. At the time, spot conversion prices had collapsed to around $4-5/kgU. In February 2021, ConverDyn announced plans to restart, targeting early 2023 production 17</a></sup>. The restart cost approximately $150 million and required hiring 160 full-time employees 18</a></sup>. By late 2024, North American spot conversion prices reached a record $97/kgU which is roughly 20x the 2017 lows 19</a></sup>. Long-term conversion prices finished 2024 at $50/kgU, up from $34.25/kgU a year earlier 20</a></sup>.</p> You can't enrich what you can't convert. Conversion bottleneck delays cascade downstream. ConverDyn is now operating at about 7,000 tU capacity half its original 15,000 tU nameplate 21</a></sup>. The plant is sold out through 2028 22</a></sup>.</p> Chokepoint 3: HALEU The Fuel That Doesn't Exist</h3> Hyperscaler SMR commitments require higher enrichment concentration. Google's Kairos investment of 500 MW requires TRISO fuel at about ~19.75% enrichment. Meta's Oklo plan for 1,200 MW requires highly enriched metallic fuel (HALEU). Meta's TerraPower investment of 2,800 MW needs metallic fuel at about ~19.5% enrichment. In total that's about ~4,500 MW of contracted capacity dependent on fuel that doesn't exist at scale.</p> Current global Western HALEU production is about ~900 kg/year, produced entirely by Centrus at its Piketon, Ohio demonstration cascade 23</a></sup>. In June 2025, Centrus completed delivery of 900 kg of HALEU to DOE under Phase II of its contract, the first significant domestic HALEU production in over 70 years 24</a></sup>. The DOE projects that more than 40,000 kg of HALEU will be needed by 2030, with demand reaching 50,000 kg/year by 2035 and over 500,000 kg/year by 2050 25</a></sup>. That is a 44x shortfall from current production to 2030 demand.</p> HALEU also requires 7x more SWU than conventional LEU, making it more expensive and more work to produce. The passage of H.R. 1042 unlocked $2.72 billion in funding; $2.02 billion for LEU production and $700 million specifically for HALEU 26</a></sup>. But the timeline to deploy that funding at scale remains unclear. Nine out of ten advanced reactor designs funded by the US government require HALEU 27</a></sup>.</p> The Implication</h2> The hyperscaler nuclear deals split into two categories.</p> Microsoft, Amazon, and Meta's Vistra/Constellation contracts are buying power from existing reactors or restarts. The fuel cycle for these plants is strained but functional. Execution risk is grid connection, regulatory approval, PPA pricing. The uranium shows up.</p> Google's Kairos deal, Meta's Oklo and TerraPower commitments, these are different. They're buying power from reactors that need fuel that doesn't exist at commercial scale. 900 kg/year from one facility. 40,000 kg/year needed by 2030. 44x shortfall.</p> We've seen what happens when fuel cycle capacity disappears. In 2017, Metropolis shut down because conversion prices collapsed to $4/kgU. By 2024, spot prices hit $97/kgU, a 20x move and the plant is sold out through 2028. Conversion was the warning shot. Enrichment is next. HALEU doesn't have a warning shot because there's nothing to shut down. The commercial supply chain doesn't exist yet.</p> The supply chain stages map to different investment exposures:</p> Mining (U₃O₈): Supply adequate. Kazakhstan dominates. Retail access via Sprott Physical Uranium Trust, Yellow Cake plc.</li> Conversion (UF₆): Near capacity, 20x price move since 2017. No direct retail exposure.</li> Enrichment (SWU): Russia controls 46%. Centrus is the only public Western play. Spot markets are OTC, illiquid.</li> HALEU: 44x supply gap. Centrus is the only producer. Current output is DOE-owned.</li> </ul> Cameco gives you mining plus conversion. Centrus gives you enrichment plus HALEU optionality. Neither gives you pure SWU exposure. The further downstream you go, the more concentrated the chokepoints, and the fewer ways to express a position.</p> The $2.7B in DOE funding is real. The 2028 Russian ban is law. But centrifuge cascades take 8-10 years to build. The SMRs that DOE is funding need fuel that DOE's enrichment programs won't produce at scale until the 2030s. The policy supports the reactors. The policy doesn't close the fuel gap in time.</p> Citations</h2> ^1</sup>S&P Global, "Datacenter Market Monitor," Oct 2025. Data Center Dynamics</a></p> </div> ^2</sup>EIA, average US home consumption ~10,500 kWh/year. EIA FAQ</a></p> </div> ^3</sup>EIA, "US operates world's largest nuclear fleet," 2025. EIA</a></p> </div> ^4</sup>Goldman Sachs, "10GW+ nuclear capacity contracted," Sep 2025. Technology Magazine</a></p> </div> ^5</sup>Morgan Stanley/McKinsey hyperscaler capex estimates, 2025. Fusion Worldwide</a></p> </div> ^6</sup>IEEE ComSoc, "Hyperscaler capex >$600B in 2026," Dec 2025. IEEE</a></p> </div> ^7</sup>Goldman Sachs via I/O Fund, "$1.15T capex 2025-2027," Nov 2025. I/O Fund</a></p> </div> ^8</sup>EIA capacity factor data, nuclear at 92% in 2024. EIA</a></p> </div> ^28</sup>PJM queue data; Goldman Sachs grid analysis, Feb 2025. Goldman Sachs</a></p> </div> ^9</sup>Columbia CGEP, "Reducing Russian Involvement in Western Nuclear Power Markets," Jan 2023. Russia held ~46% of global enrichment capacity in 2018. Columbia SIPA</a></p> </div> ^10</sup>EIA Uranium Marketing Annual Report, 2024. Foreign-origin SWU included 20% from Russia in 2024. EIA</a></p> </div> ^11</sup>DOE, "Biden-Harris Administration Enacts Law Banning Importation of Russian Uranium," May 2024. DOE</a></p> </div> ^12</sup>Congress.gov, H.R. 1042 - Prohibiting Russian Uranium Imports Act. All waivers terminate January 1, 2028. Congress.gov</a></p> </div> ^13</sup>World Nuclear Association, "Uranium Enrichment." Enrichment is capital intensive with significant barriers to entry. WNA</a></p> </div> ^14</sup>Centrus Energy Q3 2025 Results, Nov 2025. Backlog of $3.9 billion through 2040. Centrus</a></p> </div> ^15</sup>Centrus Energy Q3 2025 Results, Nov 2025. "With LEU SWU prices near historic highs, there is clear, accelerating market demand for a new U.S.-owned enrichment supply." Centrus</a></p> </div> ^16</sup>World Nuclear News, "Idled US conversion plant preparing for 2023 restart," Feb 2021. Honeywell announced temporary suspension of UF6 production in November 2017. WNN</a></p> </div> ^17</sup>World Nuclear News, "US conversion plant gears up for next 40 years," Apr 2021. Restart plan targeting production in 2023. WNN</a></p> </div> ^18</sup>POWER Magazine, "Honeywell to Reopen Sole U.S. Uranium Conversion Plant," Feb 2021. Restart cost ~$150 million, hiring 160 employees. POWER</a></p> </div> ^19</sup>Cameco Q4 2024 MD&A, Feb 2025. North American spot conversion price reached record $97/kgU at end of 2024. Cameco</a></p> </div> ^20</sup>Cameco Q4 2024 MD&A. Long-term UF6 conversion prices finished 2024 at $50/kgU, up $15.75 from end of 2023. Cameco</a></p> </div> ^21</sup>Energy Intelligence, "Nuclear Fuel: Metropolis Sold Out through 2028," Nov 2023. The facility is expected to reach its targeted annual production rate of 7,000 tons of uranium as UF6. Energy Intelligence</a></p> </div> ^22</sup>Energy Intelligence, "Nuclear Fuel: Metropolis Sold Out through 2028," Nov 2023. Energy Intelligence</a></p> </div> ^23</sup>IAEA, "Fuelling the Future: Building Fuel Supply Chains for SMRs and Advanced Reactors." Piketon's HALEU cascade expected to produce 900 kg in 2024. IAEA</a></p> </div> ^24</sup>POWER Magazine, "Centrus Completes 900-kg HALEU Delivery to DOE," Jun 2025. First significant domestic production of HALEU. POWER</a></p> </div> ^25</sup>DOE, "What is High-Assay Low-Enriched Uranium (HALEU)?" DOE estimates domestic demand could reach 50 metric tons per year by 2035. DOE</a>; Federal Register, DOE HALEU EIS, Mar 2024. DOE projects more than 40 MT needed by 2030, over 500 MT by 2050. Federal Register</a></p> </div> ^26</sup>Utility Dive, "Congress passes Russian uranium import ban, unlocking $2.7B," May 2024. $2.72 billion freed up including $700 million for HALEU. Utility Dive</a></p> </div> ^27</sup>IAEA, "Fuelling the Future." Nine out of ten advanced reactor designs funded by the US Government need HALEU. IAEA</a></p> </div> COGS Rules Everything Around Me 2026-01-26T00:00:00+00:00 AI is collapsing the cost of cognitive labor. The innovations and macroeconomic phenomena that are causing these changes are noisy and obscured in the information silos of the modern world. This is the most significant macroeconomic shift since the 1970s offshoring of manufacturing. There are implications across software, energy, and monetary systems.</p> Inputs</h2> To train models and do inference on trained models you of course need GPUs. The GPU market is wildly inefficient and because there is not really a GPU battery (you can kind of argue bitcoin is), a lot of compute is wasted. Traditionally when a company wants to get a bunch of GPUs to train a new model or serve inference, their only option was to get a long term contract. These contracts normally are many millions of dollars and for long periods of time.</p> The data centers prefer this for the same reason a landlord prefers a long term lease as opposed to month-to-month: it reduces risk. When you build a data center everything is a cost of capital game, you take out a loan based on a depreciation curve of the GPUs. Initially all the revenue is just paying back the loan, then everything is electricity and premium. Data centers in general don't want to hold onto the GPUs because they depreciate in value, unlike say a home. Thus the customer is often left holding the bag, unable to fully saturate the cluster and required to make these large payments. We at SF Compute</a> built a liquid GPU spot market in attempt to solve these problems.</p> There is much more to be said about the material and energy inputs to manufacturing chips and running clusters (more on that later). But the low hanging fruit is improving GPU utilization.</p> Outputs</h2> Over the past several decades software created this magical land of scalable high margin businesses. The US indexed heavily on the internet and offshored manufacturing. This birthed about 50 years of a roaring software economy.</p> Before software came along at all, most businesses were low margin, high volume businesses. The only exception was consulting. Everything was a COGS game. Software was an economic anomaly that, because of the cognitive moat, could defend massive scalable high margins (B2B SaaS). Software is manufactured consulting. This also gave birth to the attention economy where companies optimized against the consumer's best interest in the name of immense profit.</p> The 50 Year Bubble</h3> Not too long ago a number of companies raising money as neo clouds pulled a big con on a bunch of venture capitalists and maybe themselves too. They looked at traditional CPU-based clouds and the large software margins that come from them and said "hey, I'm going to do that with GPUs for AI." The problem here is that for AI companies training models or serving inference there is actually very little software margin. This is because the quality of your model scales with the number of GPUs used to train it. So each additional GPU makes the product better, not some software feature on top of it. And so lots of these companies that raised capital claiming margins of CPU clouds have been and are continuing to blow up or pivot.</p> Eroding Moats</h3> What does it mean when the technical moat of high-margin B2B SaaS companies goes away? These are the companies that have been fueling the largest amount of economic growth in the world over the past 50 years.</p> When the cost to produce software goes down it becomes easier to compete with software companies. When there is more competition, margins get squeezed, and companies really have to begin to justify their margins. This is good because it will really require consumer applications to start to offer a higher quality of service. Brain rot is not a strong enough consumer value proposition.</p> This will mainly happen by building mechanisms to improve market inefficiencies. The way to create value is to bundle existing things together. Some examples of how you can use efficient markets to bring the costs of goods down:</p> Ride share app market: let Uber, Lyft, Waymo, Tesla compete in real time on price for the customer</li> Message aggregator: Why does the consumer have 10 messaging apps wtf? (also then can't sell ads to them)</li> Money aggregator: Consumer finance is terrible. Can we make a one stop shop full service application for users?</li> </ul> These are only now possible with the advent of AI being the tip of the spear to end the decade long web scraping war. Agents > humans means the humans don't have to put up with the enshittification of the internet anymore. The dead internet theory is panning out in real time.</p> The Pattern</h2> What happens in a macroeconomic landscape when the cost of labor goes down? The last time this happened was in the 1970s when the US offshored a lot of its manufacturing. What happened to the economy then? Energy skyrocketed, in particular oil. This was the beginning of the era of oil. Labor arbitrage creates energy demand. Capital flows to scarcity.</p> The world has an increasing demand for energy. Large datacenter buildouts demand way more electricity than is being produced in the US. Historically energy surpluses have always coincided with periods of prosperity in human civilization. Energy scarcity has correlated with conflict. Today there are two new big energy players other than oil: solar and nuclear. The cost to produce solar has been decreasing drastically year over year. Its COGS are going down, setting it up to be a huge player in high volume low margin energy production. Uranium on the other hand is facing a drastic quantifiable supply gap. Demand of 180M lbs/year vs. supply of 140M lbs/year creates a 40M+ lb structural deficit.</p> For fifty years software let us forget that margins come from somewhere. The cost of cognition is now denominated in kilowatt hours. We're all in the COGS game again — the only question is whether you're selling commodities or buying scarcity. On a long enough time horizon, this is the good part of capitalism: the cost of living goes down.</p> The Thing About Knowing Things 2026-01-18T00:00:00+00:00 A conduct of inquiry.</em></p> If you are anyone claiming to have genuine curiosity about the world you are naive to reason about things without wondering what it means to reason. What reasoning tools exist, are they good? What assumptions do it make? What does it mean to know something? I often tell people that some philosophy needs a warning sign. Once an idea gets a foothold in your mind you cannot unthink it. This proverbial toothpaste doesn't go back in the tube.</p> Trusting your senses</h2> Naively the first thing people assume is that they can trust their senses. Historically philosophers once thought the same and dubbed this epistemic philosophy Empiricism. The problem with empiricism is that it is sometimes the case that our senses deceive us. We touch, see or hear something that is either not there, or is not able to be perceived by others.</p> Brain in a Vat</h3> The modern version of Descartes's "evil demon" tricking your own mind. There is an old thought experiment in which one attempts to argue that they are not a brain in a vat. The brain here is removed from the body and kept in a nutrient dense vat, connected to a supercomputer that generates sensory experiences for you that emulate a normal life</em>. The Matrix</a> paints more color to the same idea here. There is merit in painting skepticism against your own experience. In fact this phenomenon can be traced to the many ways people personify the devil in Judeo-Christian religion. Surely not all abstract ideas are demonic. But Empiricism cannot express objective abstract ideas. How can we measure morality, mathematics, or concepts with our senses? Empiricism provides no insight to explaining the experience of cognition.</p> However there are also more concrete counterexamples. For example the Young and Old Woman.</p> </p> Some people look at this image and see a young woman. Others see an old one. If you look at the image long enough you may be able to see both. It seems that either our senses are unreliable oracles for an idea of reality, or there is no such thing as an objective reality.</p> The Scientific Method</h2> Most people's first encounter with the scientific method is a bit authoritative. "Here is how you do science!" not a "how can we know things". To me this seemed like the next best place to look for ways in which people know things. After all, scientists claim to know things, or society seems to believe they do.</p> </p> When you break down the scientific method it stands upon the shoulders of two types of epistemic reasoning, Inductive and Deductive reasoning. Induction is a bottom up method for pattern recognition to form generalizations. Deduction is a top down application of generalizations.</p> Inductive Reasoning</h3> Induction is the process of drawing general theories from specific experiences. This is what people call the inductive step in the scientific method, when you form a hypothesis. The logic goes as follows: Observation -> Conclusion. For example:</p> Observation: I have met a friendly person</li> Conclusion: all people are friendly</li> </ul> This now gives us the ability to form some abstract relationship from experiences to ideas. However as you can tell induction is necessary but not sufficient to reason about objective knowledge. In fact, as said in David Hume's Problem of Induction</em>, just because something has happened in the past is in fact not a good enough reason for it to happen in the future. How can we justify believing future experiences will resemble past ones, or that our senses are reliable guides to truth, without appealing to reason beyond experience?</p> Deductive reasoning</h3> Deductive reasoning is the process of applying a generalization to a specific problem, often to learn about something specific. Anything with the structure If X then Y</em> is deductive reasoning where you are applying X to Y. For example, mathematics is deductive as well as formal logic. In fact there is a result known as the Curry-Howard isomorphism</a> showing that math proofs and computer programs are equivalent. The problem with deductive reasoning is twofold:</p> You depend on an assumption somewhere (initial assumption X)</li> Everything can be either true or false</li> </ul> Gödel's Theorem of Incompleteness</h2> Gödel's theorem of incompleteness is a fantastic meta result that defines the limitations of formal languages of reasoning. Most all explanations of the theorem I have come across have been verbose requiring lots of prior knowledge. An intuitive</em> way to describe the result of the theorem, not the proof, is to think of it as a trilemma where there are three properties any such language may have and you can only pick two of them.</p> The properties are:</p> Consistency: Doing the same thing gives the same result</li> Completeness: Every true or untrue statement is provable</li> Expressiveness: Defined by ability to encode natural numbers</li> </ul> </p> These are perhaps interesting results because they are restrictions on our ability to formally reason about objective truth at all. Pure induction is just perceptive pattern matching and has no semantically defined rules and thus cannot be a language of reasoning. There do exist some inductive languages of reasoning that do depend on deduction for example Bayesian inference or Solomonoff induction.</p> Because of these fundamental limitations in a unified theory of knowledge that are all independently crippling, really truly having any objective knowledge is fundamentally very difficult. It would be very bold of me to claim knowledge of much of anything. Perhaps the only thing I can say with confidence is that what you believe determines your experience of the world.</p> The Yoneda Lemma</h2> One of the most compelling results in epistemology comes from a type of mathematics called Category Theory</a>. Category theory is interesting because it is defined as the most</em> abstract type of mathematics.</p> </p> When things are abstract they generalize and are often more useful because they apply to all the things they generalize. For example it's more useful to know something about all numbers than a specific number.</p> However category theory is not just the most general theory of mathematics, it is the most general theory, period. So results in category theory apply to, well, everything. As a result of how powerful category theory is, it is actually pretty hard to make new results in it. You can see category theory come up everywhere; it is in some sense a collection of the most useful patterns in the world.</p> The Yoneda lemma states that all properties of all objects are defined by their relationships to other objects. Put simply, ontology is bullshit, there are no ontological (inherent) properties of anything at all. Everything is interdependent. If an object lives in a vacuum it is defined entirely by its relationship to the vacuum and now both our perceptive relationships to the idea of this object.</p> Interestingly this is the same view of Tibetan Buddhism. That everything is defined entirely by its relationships with everything else. This result supports a subjective relationship dependant epistomology not one of an objective truth.</p> This is a powerful result because it rests only on the axioms of mathematics, often formally referred to as the Peano axioms. That isn't to say mathematics is perfect; in fact there are paradoxes in mathematics like the Banach-Tarski paradox</a>, or Russell's paradox</a>. But it does rest on the same foundations that most modern science does.</p> Often when we encounter a paradox the academic thing to do is to put pressure on the assumptions that led us there. The assumption the Banach-Tarski paradox puts pressure on is the axiom of choice. However you also need the axiom of choice for measure theory, which you need for quantifying anything. Knowledge and reasoning is about a tradeoffs of beliefs. What you believe determines your reality.</p> Just Do Things 2026-01-14T00:00:00+00:00 The Opportunity</h2> Creating things is at the heart of the human experience. It's never been a better time to create things.</p> Artificial intelligence is noisy and polarizing, but it has drastically reduced the cost of creating. One of the many outputs of AI, but perhaps the most remarkable, is software. The frontier models are currently the worst they will ever be. They are also substantially better at programming than most programmers. Of course there are still some limitations, but they continue to improve.</p> The cost to produce software has gone down dramatically by two or three orders of magnitude. The cost to produce software used to require a team of talented engineers, each requiring a few hundred thousand dollar salary. The cost to produce software now is less than minimum wage.</p> This also means that anyone can build software. The barrier to entry, the learning curve, they've all gotten flatter. Just last week Linus Torvalds was found to be using AI for one of his projects</a>, admitting it did a better job than he did. That same week an OpenAI model solved an Erdős problem</a>.</p> Agency</h2> Agency is the only thing between you and the opportunity. Agency is the most economically valuable trait in the world. It is the ability to turn ideas into substance, to execute, to stop thinking and planning and do. It is the ability to turn thoughts into words, a proverbial incantation. Now those words are all you need. Agency is what empowers you to just do things</em>.</p> What Works Against Agency</h3> Human beings tend to overestimate unknown unknowns. When we were hunter gatherers this would greatly contribute to our chance of survival. Trying a new type of food or traveling to a new area could mean death. But in the context of the modern world and for us to create, it works against us.</p> At a personal level it feeds our insecurities as we compare ourselves to others and wonder about them. But also very much so at business and organization levels. I cannot overstate how common it is for startups to spend so much time worrying about their competition instead of just building things. The funny thing is that the competition is often doing the same thing about you.</p> Choosing Optimism</h3> Now all of this can be really scary to hear about, but there are good reasons to be optimistic. Being optimistic and excited will improve your life and empower you to have more agency and create more things.</p> Right now things are changing and there is a lot of uncertainty. Change creates fear. This can be hard; humans struggle immensely with the unknown. The way our media has been designed will reward fear and doom greatly.</p> Actively choosing to be optimistic when there is a lot of change is challenging and requires a lot of bravery. You can't plan your way to the best outcomes. They come from showing up in places you didn't know existed. Every time you act, you're buying a lottery ticket you didn't know was for sale.</p> You have no idea how good it can get.</em> The way you win is by acting.</p> State of the World 2026-01-09T00:00:00+00:00 It's 2026. The cost of producing software is less than minimum wage. The US Petro Dollar is weakening. Last year the global BRICS alliance produced more GDP than the G7 for the first time in the world. We sanctioned Russia's SWIFT access and they survived. Countries are no longer forced to price oil in dollars. Countries don't want to use SWIFT anymore. Countries are buying gold again. The USA is going to war to fight to keep its monetary grasp on the world. The purchasing power of the dollar is 1/100th of what it used to be. Pax Americana is ending. When you price equities and commodities in a rapidly depreciating asset, everything looks like it goes up.</p> As the cost of software goes down, the high-margin businesses that came with software will start to die out. Before software there was no such thing as a high-margin business. Everything was a high-volume, low-margin, cost-of-goods-sold game. I expect AI to bring us back to this type of economy. Over the next 10 years VCs will begin to die out as margins are squeezed from every angle. Now you don't need to be an engineer to create a software business.</p> As a software engineer, the things you once loved have been taken from you. The machine is now better at most everything you used to love to do, and it gets better every day. You wait in anticipation for your comfy life and salary to disappear. You desperately look for places to put your assets that will protect you.</p> What to Build</h2> If the US is going to maintain a globally respectable economic status, it needs to build things that improve people's lives. If the US builds applications that measurably improve the quality of life of its citizens, it will create shared national values and loyalty among its people.</p> Modern consumer applications make people so unhappy. Monetizing human attention should be a crime. The technology we have to do it is weapons-grade, and our brains don't stand a chance</a>.</p> The good news is that we are leaving the parasitic consumer app era and will be back on track to building software that actually improves people's lives. We can finally de-shitify the internet. The bad news is, your ability to write software doesn't give you a better shot at building it than anyone else.</p> What software actually improves your life? There are only three meaningful consumer applications that are a step function from where we are now. All of them are aggregators of existing services. You know what sucks about the internet? The fact that you have to context-manage 4 social apps, 10 messaging apps, 3 brokerages, 1 mortgage, and five health trackers. The context management of everything plus the constant fight for your attention makes you feel like each app is a crying baby vying for your attention. This makes using the internet feel more like a battle than something that actually helps people. Almost all the applications can be distilled into three broad verticals. The only remaining positive-sum game of innovation in software is to attack the margins of these businesses. These verticals are like tech trees for the US economy. The three verticals are:</p> Information (message aggregator)</li> Finance (money aggregator)</li> Health (wellness aggregator)</li> </ul> The thing about all of these is that all the companies you want to aggregate don't want to be aggregated and will do everything in their power to stop you from doing so. Go fuck 'em up—there is a trillion-dollar harpoon pointed at the Mag 7. All of these will importantly converge to low-margin, high-volume businesses. All of these depend on a strong but soveriegs digital identity.</p> Information</h3> What if your social apps weren't always trying to monetize your attention and actually were designed to help you communicate and process important information? What if news organizations followed their constitutional obligation to have your interests in mind? What if you had a single messaging application, not ten? It has been said for years now that we are living through World War Three</em>—the war on information. Most of the info slop on the internet is a result of the monetization of attention through targeted advertising. If you can build a system to meaningfully improve the quality of information presented to users, you can improve their lives. This is probably the first vertical to be built. Modern social media companies will fight you, rate-limit you, blacklist you, and even try to acquire you in your journey to build this.</p> Finance</h3> Can we create a single useful holistic view of a person's finances without weaponizing it against them? A dashboard where your entire financial picture can be managed and owned by you. This is by no means a novel idea; many companies have been trying to do this for years (see Rocket Money, Mint, etc.). The problem is they are not vertically integrated enough to give you execution from one place—only visibility (still better than nothing). Sovereignty here is important, and the digitization of fiat currency is well on its way for this to come about. Financial organizations have built a legal swamp moat around themselves to prevent people like you and me from playing ball. Do you have a Series X license? Are you an accredited investor? Does your brokerage have FDIC insurance? The game here will be less a technical game and more of a legal one.</p> Empowering people's sovereignty is empowering.</p> Health</h3> Aggregate all your biomarker sensors, your gym apps, your medical history, your blood work into something that is supportive</em> of your vitality and longevity. Health applications suffer from a legal swamp in a different way. HIPAA compliance? There exists a legitimate technological challenge in gathering consistent health data and then transforming that into personalized insights</em>. Despite the challenges here, I cannot think of an industry that could be more important to modernize and improve. The bar for quality healthcare/applications is so low and the opportunity to improve people's lives so great. There are opportunities in software here for sure, but much of the impact and long-tail benefit of improving this vertical lies in hardware. The opportunity to improve drug discovery and research on human aging and</em> flourishing is unquantifiable. It feels good to feel good.</p> What could be a more honorable endeavor than helping people take better care of themselves?</p> Hardware</h2> The main point of software on a long time horizon is to make better hardware. Unfortunately, the USA has forgotten how to manufacture physical things. After years of exporting the construction of physical goods overseas, the domain expertise has dwindled. However, there is a shift in the air. The genius of the Elon Cinematic Universe is in manufacturing real physical things. American Dynamism may yet bear fruit.</p> While I believe this to be the most important thing, it is also the most difficult. Building physical things is real engineering—it's hard because hardware is hard. Software is easy because it's soft; you can change it, fix it if it breaks. When hardware breaks, you need to build it again, sometimes from the beginning.</p> Chemical, biological, and physical manufacturing will be some of the most important industries in the world.</p> The Power of Ideas 2026-01-02T00:00:00+00:00 The most ephemeral but powerful forces in the world are our thoughts. These shape our perception of the world, ourselves, and eventually drive us to make the decisions that we do. I find it difficult to imagine an experience without a conceptualization or an intellectualization of experience. The dissemination of ideas has shaped the world, defined civilizations, cultures and values of nations, movements, cults and individuals. In more modern times our thoughts have defined and shaped companies, businesses and markets.</p> The technologies known to have impacted the world most strongly are the ones that shape information dissemination. The impact of Johannes Gutenberg's discovery of the printing press was a technology that allowed for writers to shape our present perception of the past. It also led to the protestant reformation whose impact is still felt in the western world today defining a culture of religious pluralism and laying the ground for the scientific revolution.</p> The same zeitgeist is felt today with the internet; shaping our ideas of good and bad, success and failure through algorithmic content platforms where individuals vie for resources more valuable than gold: attention. The monetization of attention has been immensely profitable to consumer internet applications allowing businesses to pay for targeted advertisement in efforts to change your mind in the name of profit. When you have more personal data about your users, you are able to profile what and when they are the most susceptible to influence. Nothing in the history of human civilization has shown to be as effective at changing your mind than highly personalized targeted ads.</p> What are you if not only an idea in your mind?</p> The Algorithm</em></h2> The dominant driving force of the collective narrative rests in the algorithm of social milieu, the driving force optimizing for engagement not necessarily for wellbeing or truth. If you yourself have resources to expend, you can spend it to leverage the algorithm to shape the minds of internet consumers to your own motives. Changing what someone believes is an existential phenomenon, for what people believe determines their lived experiences, their reality. The algorithm is thus a reality distortion tool, enabling thought/cult leaders, governments, the politically motivated, businesses to shape the very substrate of who you are.</p> The algorithm's sophistication was a natural byproduct of capitalist incentives to drive engagement and sell more targeted advertising and is both a great tragedy and a great opportunity for anyone willing to pay for mindshare. Venture Capitalists, Entrepreneurs and opportunists were quick to leverage the latest and greatest information dissemination technology for power.</p> The acceleration of the commoditization of the human experience has been the greatest trick we have played on ourselves. Forsaking ourselves for nothing</em>.</p> To bring awareness to the ways in which our own minds are impacted is groundless when you have no way to distinguish your own ideas from the ideas paid to be engrained within you. Not to say that we are stupid or naive, but the sophistication of the algorithm is so great. It is not a fair fight to pit our slowly biologically evolving brains against the latest and greatest technology purpose built and optimized</em> to get you to think anything for a price. Even when one has strong self-awareness and discipline, the moment one engages with an algorithmic content machine, they must acknowledge that they are being shaped and sculpted by the interests of others through the information they consume.</p> Our monkey brains stand no chance.</p> </p> It's no mystery to anyone that there are dramatic byproducts of consuming information at such a rapid pace.</p> We are quick to adopt ideas without having any skin in the game or even forethought on the bigger picture. The speed at which things seem to change leaves newer generations unperturbed by high-contrast narratives. Generation Z, known for its nihilistic irony, is left like a leaf in the wind, unbothered by even the most dramatic of ideas.</p> The most powerful countries in the world, Russia, China, and the USA all differ in how they control the flow of ideas via technology about themselves but also each other.</p> There is nothing online that isn't either being used to gain your attention or being paid for by someone for you to see.</p> Imprisonment of Perception</h2> My initial takeaway from this reality was to attempt to protect my mind, ideas, and beliefs from such powerful tools of chaos. For many years I have strongly resisted any sort of social media or participating in algorithmic content platforms. I had for many years removed my social media accounts, made efforts to be offline and reduce my digital identity and footprint as much as possible.</p> With values shaped by ideas from Eastern philosophy and surveillance capitalism, I had selfishly wanted to avoid the perception of others altogether. When people have an idea of who you are, what you are like and perceive you to have properties x, y, or z, they bestow upon you an inevitable expectation of who you are to them. This can be limiting and in many cases push you to fulfill expectations of others without continuously redefining yourself.</p> Life by the Dunbar Number</h3> It turns out that consuming a global narrative is overwhelming and difficult. The amount of information consumed with no ability to impact it works against our biology.</p> In the 1990s a British anthropologist Robin Dunbar noticed that primates with larger neocortices tended to live in larger social groups. Studying the size of these groups he coined the Dunbar Number to be 150 as the cognitive limit for the number of meaningful relationships in our life. The study also proposed a strong upper bound of 1500 as the strict upper bound on the amount of people we can recognize.</p> So up until now I had made consistent but perhaps futile attempts to live by the Dunbar number, focusing on the relationships I have with the people in my life. Trying to not pay heed or get wrapped up in all the suffering of the world that I couldn't do anything about.</p> Skin in the Game</h2> In order to see things as they are you must first abandon hope of finding an ideology to stand upon. There is only moral subjectivity, and thus nothing concrete exists. Even the most seemingly sinister of actors believe what they're doing is right.</p> Without objective truth, values are revealed by their costs. Authentic beliefs require sacrifice. Withdrawal from the algorithmic arena is safe. It costs nothing. But if ideas matter, if your perception of reality matters, then silence is surrender.</p> Withdrawal doesn't protect you; it just means surrendering the battlefield entirely. If ideas disseminate through these platforms, then staying out means your ideas don't exist in the cultural conversation at all. Yes, you'll be manipulated. But participating with awareness, eyes open to the distortion, is better than the guaranteed defeat of non-participation. Writing becomes an act of defiance not against the algorithm, but against the fear of being shaped by it.</p> May you be blessed by the milieu of the internet.</p>

Network Spirits

From Yellowcake to Gigawatts

Fuel Cycle 101</h2> Nuclear power has an elaborate supply chain with many components. Uranium doesn't arrive enriched. It goes through four stages before it becomes fuel.</p>

Chokepoints</h2> The fuel cycle looks linear on paper. In practice, it has three points where demand stacks up against fixed capacity. Each one matters differently depending on which nuclear story you're tracking. Importantly there is no chokepoint on uranium ore or yellowcake.</p>

COGS Rules Everything Around Me

The Thing About Knowing Things

Inductive Reasoning</h3> Induction is the process of drawing general theories from specific experiences. This is what people call the inductive step in the scientific method, when you form a hypothesis. The logic goes as follows: Observation -> Conclusion. For example:</p>

Just Do Things

State of the World

The Power of Ideas

Fuel Cycle 101</h2>
Nuclear power has an elaborate supply chain with many components. Uranium doesn't arrive enriched. It goes through four stages before it becomes fuel.</p>

Chokepoints</h2>
The fuel cycle looks linear on paper. In practice, it has three points where demand stacks up against fixed capacity. Each one matters differently depending on which nuclear story you're tracking. Importantly there is no chokepoint on uranium ore or yellowcake.</p>