Odd Lots

James Van Geelen on the Next Phase of the AI Buildout

October 6, 2025

Key Takeaways Copied to clipboard!

  • The massive AI infrastructure buildout, exemplified by OpenAI's Stargate data center in Abilene, Texas, is driving a significant, real-world industrial CapEx cycle involving the construction of dedicated power generation facilities like natural gas plants. 
  • The energy bottleneck for AI buildout is leading companies to increasingly rely on behind-the-meter power generation, utilizing simple-cycle natural gas turbines from suppliers like GE Vernova and Caterpillar, which are experiencing high demand and stock appreciation. 
  • The AI investment cycle exhibits bubble-like characteristics, such as complex financing (e.g., GPU-backed SPVs and off-balance-sheet arrangements) and zealous language ("Machine God"), but the demand appears real, evidenced by 100% utilization of new capacity and the subsequent surge in 'picks and shovels' plays like industrial equipment and data storage companies. 

Segments

Data Center Field Trip Intro
Copied to clipboard!
(00:01:33)
  • Key Takeaway: The hosts express curiosity about touring hyperscale data centers and the potential for identifying ‘picks and shovels’ AI investment opportunities within their physical components.
  • Summary: The hosts discussed the appeal of touring a data center, focusing on examining every physical component for publicly traded investment plays. They humorously suggested creating a ‘hyperscale data center ETF’ based on these granular equipment suppliers. This sets the stage for the guest’s on-the-ground research.
Stargate Data Center Visit
Copied to clipboard!
(00:02:46)
  • Key Takeaway: James Van Geelen visited one of the five planned Stargate data centers in Abilene, Texas, which is part of a half-trillion-dollar infrastructure buildout comparable in scale to post-WWII efforts.
  • Summary: Van Geelen used drone footage to confirm the massive scale of the Stargate facility, noting its size rivals lower Manhattan. The project involves major players including Oracle, OpenAI, NVIDIA, and Saudi interest via MGX. This single site is one of five planned facilities representing the largest infrastructure buildout since World War II.
Power Generation and Turbine Shortage
Copied to clipboard!
(00:07:30)
  • Key Takeaway: The Stargate facility is powering itself via ten on-site, simple-cycle natural gas turbines from GE Vernova and Caterpillar’s Solar Turbines due to the immediate power needs and a multi-year shortage of advanced turbines.
  • Summary: The data center built its own power source, consisting of ten 35-megawatt simple-cycle natural gas turbines, as relying on the existing grid is too risky following events like Winter Storm Uri in Texas. The use of lower-end turbines is necessitated by a shortage, which has benefited industrial equipment makers like Caterpillar.
Picks and Shovels Investment Thesis
Copied to clipboard!
(00:09:01)
  • Key Takeaway: The AI buildout is currently in a heavy construction phase, creating massive demand for industrial labor and materials, benefiting old-line industrials whose stock performance is now catching up to the secular theme.
  • Summary: The construction phase requires 7,000 workers on a single site, indicating that the ‘picks and shovels’ phase is still active and spending is occurring rapidly. Companies like Caterpillar have seen significant stock appreciation as the market recognizes this infrastructure spending is more akin to LNG terminal CapEx than traditional tech CapEx.
Data Center Location and Energy Strategy
Copied to clipboard!
(00:13:25)
  • Key Takeaway: Hyperscalers are deliberately choosing remote locations like Abilene, Texas, to build their own power infrastructure and avoid local political opposition (NIMBYism) concerning energy and water usage.
  • Summary: Future data centers are planned in areas with available land, gas, and power, such as Shackleford, Texas, with Abilene planning a 600-megawatt expansion. Advancements in closed-loop liquid cooling have significantly mitigated water consumption concerns, making power generation the primary constraint driving the need for private, behind-the-meter power plants.
AI Bubble Dynamics and Financing
Copied to clipboard!
(00:20:28)
  • Key Takeaway: While the excitement around AI suggests a bubble, the current CapEx spending is fundamentally different from the dot-com era because the infrastructure being built is immediately utilized at 100% capacity, with no ‘dark fiber’ equivalent.
  • Summary: The current cycle is characterized by real utilization, unlike the dot-com bubble’s over-investment in unused fiber. Financing structures, including off-balance-sheet GPU collateralized SPVs, resemble past bubbles, but the underlying demand for compute (especially for video models) is driving sustained CapEx.
NVIDIA’s Role in Financing Customers
Copied to clipboard!
(00:24:09)
  • Key Takeaway: NVIDIA finances its customers (like OpenAI and CoreWeave) because achieving AGI as quickly as possible is the best outcome for NVIDIA, as any delay reduces the likelihood of reaching that massive future infrastructure demand.
  • Summary: NVIDIA’s investment strategy is tied to accelerating the entire AI ecosystem toward AGI, which represents the ultimate, massive demand for their hardware. Furthermore, since this boom is CapEx-intensive, a halt in spending due to market downturns would be catastrophic, incentivizing NVIDIA to support customer financing.
Robotics Data Needs and Unitree
Copied to clipboard!
(00:36:39)
  • Key Takeaway: The next frontier of AI data generation requires physical interaction data, which companies like China’s Unitree are gathering by deploying affordable, safety-conscious humanoid robots into the real world.
  • Summary: Unitree, based in Hangzhou, focuses on hardware, manufacturing components like actuators and LIDAR in-house or locally. They are intentionally deploying robots that are not overly powerful (humanoid starts at $16,000) to safely collect vast amounts of real-world movement data necessary for advanced robotics and vision-language-action models.