You can't git clone a gigawatt
There's always a moat in tech. It just moves. From Windows apps to user feedback, the hard thing to copy changes. My next bet is electricity.
Jack Dorsey says there is no moat, and Elon Musk says Sparta never had one. Meanwhile, Musk’s own company reported running 69 gas turbines in July 2026 to keep Grok’s data centers powered. How long would it take to copy that software moat?
There’s always a moat in tech. It just moves.
How the moat moved
- 1981 · Hardware. Developers wrote software for the IBM PC. Compaq legally reverse-engineered its BIOS and introduced its portable in 1982. The clones followed.
- 1990s · The operating system. Users came for the Windows apps. Developers came for the users. A US court called it the applications barrier to entry.
- 2000s · Search data. Google learned from what people searched and clicked. In 2025 a US court ordered it to share some of that data with rivals.
- 2008 · Distribution. The App Store decided what ran on the iPhone, and took up to 30%.
- 2010s · The interface. A pro’s hands know the Photoshop shortcuts. An agent has no muscle memory. In October, TechRepublic credited Brandon Thomas and collaborators with seven open source Adobe-style apps built with Claude. PhotoCraft calls itself early alpha, not a replacement for daily professional work.
- 2023 · The model. A leaked memo from a Google engineer argued Google had no moat there, and neither did OpenAI. In January 2025 DeepSeek claimed performance comparable to o1 on math, code and reasoning tasks. Nvidia lost $589 billion in market value that January, Bloomberg reported.
- Now · Your users’ data and their AI usage. Memory, chat history, every suggestion they accept or reject. Cursor gives a specific example: its researchers described frequent daily Tab model rollouts in September 2025, trained on accepted and rejected suggestions.
A working agent loop can be copied. The history of how your users interact with your AI agents doesn’t come with it.
How long it takes to copy a software moat
Downloading code, matching a benchmark and connecting a power plant measure different things. The comparison shows different timescales, not how long it takes to win customers.
| What takes time | Reported duration | What the number measures |
|---|---|---|
| Open models reaching earlier closed-model capability | About 4 months | Epoch AI’s May 2026 analysis: average lag since January 2026 on its capability index. |
| Procuring a power transformer | 128 weeks on average; up to 4 years | Wood Mackenzie’s Q2 2025 average and the IEA’s 2025 survey. |
| A generation project’s grid agreement in PJM | 40 months on average | Carbon Direct’s May 2026 analysis. This concerns power generation, not a data center’s connection request. |
| Projects reaching operation in PJM | More than 7 years on average | Data Center Knowledge’s May 2026 report of PJM data, for projects entering service in 2025. Includes work after approval; not a general data-center hookup time. |
Next, my bet: electricity
A November 2025 report quoted Satya Nadella describing “chips sitting in inventory that I can’t plug in.” His problem was getting powered data centers built fast enough.
PJM’s board estimated that data centers would account for roughly 30,000 of 32,000 MW of additional peak demand between 2024 and 2030. That’s about 94% of the projected growth in one region, not national electricity use.
The turbines show that a grid connection isn’t the only route. By late August, 11 had powered down under a plan to replace the temporary turbines with a permanent power plant. That still means building physical infrastructure.
A long wait alone doesn’t prove a lasting moat. And if power becomes one, a team renting compute meets it as a cost to control, not an advantage it owns.
Agents make code cheaper to copy. They don’t make a transformer arrive sooner.
You can git clone code. You can’t git clone a gigawatt.
On this page