James van Geelen on the Next Phase of the AI Buildout

Summary of James van Geelen on the Next Phase of the AI Buildout

by Bloomberg

38m•October 6, 2025
Listen to Episode

Summary — "James van Geelen on the Next Phase of the AI Buildout" (Bloomberg Odd Lots)

Overview

This episode is a conversation with James van Geelen (Citrini newsletter) about the real-world, physical buildout behind the AI boom. Van Geelen describes field research (drone & site visits) to mega data-center projects like “Stargate” in Abilene, TX, and explains how that on-the-ground view changes the investment and macro outlook: the build is massive, capital-intensive and broad (power, cooling, construction, storage, robotics), and it has both clear economic winners and structural risks (financing, power constraints, political/local friction). The hosts and guest also debate whether the current AI frenzy is a bubble and outline indicators to watch.

Key points & main takeaways

  • Scale: The hyperscaler buildout is enormous — projects measured in gigawatts and square miles (e.g., Stargate/Abilene; Meta Hyperion in Louisiana). Van Geelen calls it “probably the largest infrastructure build out and effort since World War Two.”
  • On-site power: Many sites are building behind‑the‑meter natural gas power plants (simple-cycle turbines, often smaller/fast-to-procure models from Caterpillar/Solar Turbines and GE Vernova) because grid reliability and speed-to-build matter more than maximum efficiency.
  • No “dark fiber” analogy: Unlike the dot-com era overbuild, current data centers are quickly utilized — van Geelen claims new centers hit close to 100% utilization when online.
  • Water & cooling: Liquid (closed‑loop) cooling and adiabatic systems reduce net water footprint relative to earlier concerns; power remains the primary bottleneck.
  • Supply-chain winners: Not only GPUs (NVIDIA) but also industrials and infrastructure suppliers (Caterpillar, GE Vernova, Siemens Energy, Seagate, Western Digital, Eaton, Mitsubishi Heavy, etc.) are direct beneficiaries — many old-line industrials have become key AI “picks and shovels.”
  • Financing complexity: Large off‑balance-sheet financing and asset-backed structures (PIMCO on debt for Meta Cyperion, Blue Owl equity, Blackstone SPV for CoreWeave, GPU‑collateralized loans, etc.) create circular/fragile financial arrangements that can amplify downside risk.
  • Neoclouds & capacity markets: Hyperscalers are contracting external GPU/compute capacity (e.g., CoreWeave) to shift capex and operational risk off-balance-sheet.
  • Bubble vs. real demand: There is real demand for compute, storage, and data; yet financing dynamics, PR/marketing, and “religious” beliefs about AGI create bubble-like signals. The buildout is capital‑intensive, meaning bust dynamics would differ from the dot‑com crash.
  • Robotics & training data: Robotics hardware (e.g., Unitree robots) and video models are critical to the path toward more embodied AI — robotics hardware makers (including many Chinese firms) are already shipping affordable units to collect training data.
  • Labor & local economy effects: Massive local hiring for construction and maintenance can crowd out other local businesses and drive inflation in specific trades (HVAC, electricians, etc.).

Notable quotes / insights

  • “This represents probably the largest infrastructure build out and effort since World War Two.” — James van Geelen
  • “Every single time that one of these data centers comes online, you're at 100% utilization. There's no dark fiber.” — James van Geelen
  • Site naming and mindset: data halls named “ludicrous building” and the recurring theme of building “machine God” — a cultural indicator of zeal and potential excess.
  • Practical/strategic investor idea from van Geelen: “Buy the real companies building the physical data center hardware and short private credit players that are financing circular structures.”

Topics discussed

  • Field research: drone & site visits to Stargate (Abilene) and other hyperscale projects
  • Data-center scale, construction and security details (walls, building envelopes, cooling)
  • On-site power plants: simple-cycle natural gas turbines and turbine shortages
  • Major hardware and industrial suppliers benefiting from the build
  • Liquid cooling, water usage, and environmental claims
  • Financing structures: off‑balance sheet deals, SPVs, GPU-collateralized loans, private credit
  • Neoclouds (CoreWeave) and hyperscaler capacity contracts
  • Bubble dynamics and historical parallels (railroads, dot-com)
  • Indicators to watch for a potential turn in the cycle
  • Robotics hardware (Unitree), video models, and training-data economics
  • Labor market and local economic/crowding-out effects

Risks, indicators & what to watch

  • Financial warning signs:
    • Rising distress/waivers in private credit and SPVs
    • Contract renegotiations or missed contractual deliveries
    • Falling used‑GPU prices (implies softening demand for compute)
    • News of delayed or undelivered power hookups or turbines
  • Operational & political risks:
    • Local opposition (NIMBY) and permitting delays
    • Power grid reliability and regulatory constraints
    • Supply‑chain bottlenecks (turbines, storage media, specialized components)
  • Technology/revenue indicators:
    • Monetization developments (ads at model outputs, payments inside AI apps)
    • Adoption of video models and robotics (demand for storage, specialized compute)
  • Market signs of froth:
    • Circular financing becoming more widespread
    • Exuberant rhetoric (e.g., “machine God”) combined with large off‑balance sheet allocations

Action items & recommendations (for investors / analysts)

  • For investors who want exposure but worry about financing risk:
    • Long position candidates: industrials and hardware firms that supply turbines, cooling, power transmission, storage (e.g., Caterpillar, GE Vernova, Siemens Energy, Seagate, Western Digital) — these have real products and alternative revenue streams.
    • Hedge/short candidates: private credit firms and SPVs (Blue Owl, certain asset-backed/finance intermediaries) that may bear disproportionate downside in a capex slowdown.
  • Monitor short-term signals: used GPU market prices, private credit waiver headlines, power delivery timelines, and contract renegotiations.
  • Map the financing flows: build a visualization of major off‑balance-sheet deals, SPVs, and GPU-backed loans to understand counterparty and concentration risk.
  • Watch for structural shifts in demand: video model adoption (storage demand), robotics deployments (actuators/LIDAR suppliers), and hyperscaler contract terms with neoclouds.
  • Consider local macro impacts: evaluate labor/capacity constraints in regions with large buildouts (wage inflation for trades; local political risk).

Bottom line

The AI infrastructure buildout is real, massive, and already reshaping traditional industrial and capital markets. There are clear, investable winners among physical suppliers and storage vendors, but the financing and valuation dynamics introduce non-trivial systemic risks. Practical monitoring (GPU secondary markets, private credit strains, power delivery) and a focus on physical supply-chain winners vs. financing intermediaries provide a pragmatic framework for participation and risk management.