# Gavin Baker’s “Land Man” primer for the AI buildout

*By Recommended Reading from Tech Founders • August 15, 2026*

The strongest organic recommendation is a TV series used to understand the physical work behind data centers; three additional leads cover orbital-compute assumptions, self-improving research, and collateral risk.

## The standout: the physical layer of AI infrastructure

- **Title:** *Land Man* (the title as spoken in the discussion)
- **Content type:** TV series
- **Author/creator:** Not stated in the source
- **Link/URL:** No direct link supplied
- **Recommended by:** Gavin Baker. The episode introduces Baker, asks him to address the energy and data-center supply needed to scale AI, and he answers that the best way to understand the people building data centers is to watch the series. [^1]
- **Key takeaway:** Baker’s point is “atoms, not bits”: the work involves the oil patch, 110° heat, and coordinating thousands of people in remote locations. [^1]
- **Why it matters:** This is the most concrete recommendation for understanding AI infrastructure as an operations problem rather than only a model or chip story. [^1]

## A model for the terrestrial-to-orbital constraint

- **Title:** *AI Compute “Keystone” Model* and its linked public brief
- **Content type:** X post and research brief
- **Author/creator:** Vlad Saigau; the post is associated with Mach33 Financial Group. [^2]
- **Link/URL:** [X post](https://x.com/VladSaigau/status/2087955507650724344) · [public brief](https://research.33fg.com/analysis/brief-ai-compute-grows-180x-while-prices-fall-96) [^2]
- **Recommended by:** This is a recommendation chain rather than a single direct endorsement: Elon Musk linked Owen Lewis’s post while arguing that orbital compute may be the only way to scale AI around 2029; Lewis called the underlying orbital-compute point “literally the most important part” and linked Saigau’s work. [^3][^4][^5]
- **Key takeaway:** Saigau’s model says orbital data-center compute could become necessary for AI scaling in the 2030s. Its own projections put terrestrial capacity at a 111 GW peak in 2031 and roughly 438 GW of a 488 GW 2040 fleet in orbit, while the author stresses that no one—including the authors—has a narrow projection and that the framework is built to evolve. [^2]
- **Why it matters:** Read this as an assumption-rich infrastructure framework, not as a forecast to accept wholesale. Its value is that it makes terrestrial capacity, demand, price deflation, and uncertainty explicit. [^2]

## A concrete pattern for self-improving research

- **Title:** Copacetic’s “auto research” X post (no formal title supplied)
- **Content type:** X post
- **Author/creator:** Copacetic
- **Link/URL:** No URL supplied in the source
- **Recommended by:** Tom Blomfield, who calls it a “great tweet.” [^6]
- **Key takeaway:** The post applies an iterative loop to machine learning: have AI generate research ideas, test them overnight, and hill-climb toward a measurable outcome—discarding moves downhill and retaining moves uphill. [^6]
- **Why it matters:** It is a compact design pattern for self-improving ML, where evaluation and repeated experimentation are part of the system rather than a one-off prompt. [^6]

## A collateral-risk reading for compute finance

- **Title:** John Geanakoplos’s academic literature on collateral (the specific paper or title is not named)
- **Content type:** Academic literature / research
- **Author/creator:** John Geanakoplos
- **Link/URL:** No link supplied in the source
- **Recommended by:** Josh Wolfe, one of the panelists identified in the discussion. In the conversation about securitizing AI compute, the transcript directs readers to Geanakoplos’s literature. [^7]
- **Key takeaway:** The housing crisis is framed as a collateral problem—not merely a bubble or a debt problem—with collateral becoming worthless and the debt following it. [^7]
- **Why it matters:** The recommendation is a useful historical check on AI-compute financing: the same discussion flags unknown GPU depreciation and duration mismatch, so Geanakoplos offers a way to pressure-test whether compute can support long-lived financing assumptions. [^7]

---

### Sources

[^1]: [Anthropic's $2T IPO, Zuck's AI Manifesto, Nvidia's $500B AI Bet, Grok's Comeback](https://www.youtube.com/watch?v=kVzYGVJ8zUk)
[^2]: [Vlad Saigau \(@VladSaigau\) on X](https://x.com/VladSaigau/status/2087955507650724344)
[^3]: [𝕏 post by @elonmusk](https://x.com/elonmusk/status/2088306926442430578)
[^4]: [Owen Lewis \(@is_OwenLewis\) on X](https://x.com/is_owenlewis/status/2088071043487502350)
[^5]: [𝕏 post by @is_OwenLewis](https://x.com/is_OwenLewis/status/2088071043487502350)
[^6]: [Building And Structuring An AI Native Company](https://www.youtube.com/watch?v=Z3JyAqh4ixg)
[^7]: [Special Edition MOL with Josh Wolfe, Rachel Holt, Scott Belsky, Scott Stanford, and Peter Deng.](https://www.youtube.com/watch?v=XySpdJ6UPkQ)