AI Code and Capital

Every day brings another keynote, Substack post, or breathless venture-capital thread declaring that human engineers are fundamentally obsolete. We are told that generative AI is writing all the software now, finally delivering management from the expensive burden of employing our uncooperative, union-curious tribe. But I spent some time looking at the actual legal physics underlying this massive wave of data-center capex, and we are collectively ignoring a catastrophic bug in the business model. The entire financial engine of this ecosystem hinges on who gets to own the output; and according to federal law, nobody does.

1 · The Hypocrisy. Just look at the glaring contradiction embedded in recent industry dispatches. On Wall Street earnings calls, executives like Larry Ellison boast that AI is taking over vast swaths of Oracle's coding duties. Executive Vice President Mike Sicilia doubles down, enthusiastically telling investors that AI is churning out their new code at unprecedented velocities. Meanwhile, back in the actual engineering trenches, Oracle explicitly bans AI-generated code from OpenJDK. That stark policy shift recently triggered a massive discussion on Hacker News, and it tells you everything you need to know about the gap between executive marketing and enterprise risk management. View this disconnect as a corporate immune response. Ellison and Sicilia are out there selling the automated dream to Wall Street; the general counsel's office is frantically patching the legal attack surface to protect the crown jewels.

2 · The Oracle Irony. Oracle actually serves as the perfect controlling device here. Remember, they fought Google all the way to the Supreme Court over a handful of Java API declarations. They understand down to their very bones that software capitalism is, at its core, intellectual property. The financial engine of the tech sector relies exclusively on the proprietary ownership of code. When a corporation pours billions into a new technology platform, they absolutely require the ability to lock up the results, patent the algorithms, and extract rent for the next twenty years. Without an ironclad copyright monopoly, the entire Silicon Valley ecosystem ceases to function. It just stops.

3 · The IP Vacuum. Now apply that rent-seeking model to a generative LLM. I am not a lawyer, obviously, but I can read a government mandate. According to explicit guidance from the U.S. Copyright Office, if the traditional elements of authorship are generated by a machine, the work lacks human authorship and simply cannot be copyrighted. The Copyright Office's AI policy draws a ruthless boundary right through the C-suite's business plan. Follow the fatal logic: if you build your proprietary commercial product using an LLM, you do not actually own the resulting code. If you lack ownership, you possess no defensible moat against a competitor who simply copies it. Lacking a moat, you cannot extract rent.

4 · Slop and the Maintainer Toll. Shift your gaze from financial fantasies down to the ground-level reality of active open-source repositories. When an LLM writes code, it invariably generates vast volumes of subtly broken boilerplate slop. Human maintainers are forced to untangle it, bearing an enormous uncompensated labor burden as they review massive, machine-generated pull requests. The AI hype machine claims this automation lowers the cost of software development, which is frankly hilarious. In reality, these tools are simply dumping the expensive verification work onto open-source volunteers who are already burning out. And now every engineering group gets to stare down a relentless firehose of unverified, machine-generated patches flooding the mainline branch. Dealing with that endless deluge of synthetic technical debt is a miserable new reality; someone has to clean up the mess.

5 · The Historical Cycle. Anyone who has been observing technology hype cycles for a few decades recognizes this exact pattern. The underlying corporate desire to fire the programmers is eternal, and the promised silver bullets all look exactly the same. Compare this codeless-programming delusion to the 1990s era of computer-aided software engineering tools, 4GLs, and klunky drag-and-drop UML generators. Those tools perpetually failed because translating ambiguous business logic into exact execution is exactly what programming actually is. You can radically change the syntax. You can wrap it in a slick chatbot interface. But you cannot wish away the fundamental friction of specifying exact state logic to a literal-minded computer.

6 · Follow the Money. Let's look at the underlying math. The server and energy capex required to train and run these generative AI models is astronomical. We are talking about nation-state levels of investment, requiring an eventual monopoly payout to justify the initial burn rate. And yet, the resulting software output legally belongs to the public domain. The entire financial structure collapses under its own weight when the product you are paying billions to produce cannot be legally hoarded. [Note to the accounting department: Capex requires a moat. -Ed.]

So here we are. The tech giants are burning through capital at a rate that would make a sovereign wealth fund sweat, all to produce code that anyone can legally copy and paste the second it hits GitHub. You can't run a software empire without owning the software. The legal physics are absolute. We live in a non-ideal world fueled by proprietary software. In such a world, if you can’t copyright the code, the capital dries up. Do the math.