A short while ago, I published a piece arguing that the sudden, performative hesitation by frontier AI companies wasn’t driven by altruistic safety concerns, but by a cold, physical reality: the data wall.

I laid out the math. The public internet had been scraped dry. Citing projections from research groups like Epoch AI on the limits of LLM scaling, human-generated text was on track for exhaustion. The web was drowning in recursive synthetic slop, and model collapse was threatening the brute-force scaling paradigm. I argued that the labs were hitting the brakes because their engines were simply starving for clean fuel.

I was right about the physics. But I was fundamentally wrong about the strategy.

I gave these executives too much credit for engineering humility.

The labs aren’t tapping the brakes because they fear their models will collapse on bad data, nor because they fear AGI will break out of a datacenter and end civilization.

They are tapping the brakes to lock the door behind them and choke open-source AI to death.

The Leaked Memo They Couldn’t Ignore

In early 2023, an internal Google document leaked to the public via SemiAnalysis with the headline: "We Have No Moat, And Neither Does OpenAI."

The author warned that while tech giants were spending hundreds of millions of dollars training monolithic, proprietary models behind expensive API paywalls, a ragtag global army of open-source developers was solving the exact same problems for pennies. Through efficient fine-tuning (LoRA), aggressive quantization, and architectures like Meta’s Llama, the open-source community was delivering 90% of the capability at 1% of the infrastructure cost, running locally, privately, and for free.

For closed-source foundation labs, this wasn't just a technical challenge; it was an existential economic catastrophe.

Think about the math. If you are burning tens of billions of dollars on compute clusters, locked into non-cancelable cloud contracts with hyperscalers, and paying top-tier researchers millions a year, you must charge premium, toll-booth API prices to survive.

As Mark Zuckerberg bluntly argued in his manifesto, Open Source AI Is the Path Forward, concentrating model access inside a handful of closed cloud vendors creates an anti-competitive ecosystem that inflates costs for everyone else.

How do you charge a developer $20 a month or $0.03 per thousand tokens when an open-weight model downloaded from Hugging Face and running entirely offline on their private server does the job just as well, without rate limits, without telemetry, and without corporate censorship?

You can’t. In an open, competitive market, proprietary API margins trend toward zero.

Unless, of course, you make releasing open weights practically illegal.

The Anatomy of the Regulatory Guillotine

You don't crush open-source software by saying you hate open-source software. You crush it by expressing deep, profound, theatrical concern for the safety of humanity.

Instead of lobbying to break up the hardware monopolies or hold cloud providers accountable for massive energy footprints, closed labs have thrown their weight behind specific legal mechanisms that dismantle open competition:

1. The Gated Club of Compute Thresholds

Frontier policy proposals, like the high-profile framework pioneered in California’s SB 1047, established arbitrary compute triggers as the legal boundary for strict compliance.

If training a frontier model requires federal security clearances, mandatory government registration, and millions in recurring third-party audits, only corporations with multi-billion-dollar balance sheets and dedicated regulatory teams can participate. It turns frontier AI from a permissionless software discipline into an aerospace-style defense cartel.

2. Downstream Strict Liability: The Poison Pill

This is the silver bullet aimed directly at open source. When legislation tries to hold the original model developer legally liable for anything a third party does with the model downstream, it creates an asymmetrical trap.

If you host a proprietary API behind a closed server, downstream liability is easy to manage: you monitor every prompt in real time, filter inputs, and ban accounts that violate your terms.

An open-source creator (whether a university, a non-profit, or an independent developer) has no way to monitor or police how a downloaded set of mathematical weights is used. Under downstream strict liability, distributing open weights becomes an uninsurable legal catastrophe.

3. Mandatory Kill Switches

Legislative mandates demanding that models possess a centralized kill switch capable of executing an emergency, remote shutdown if dangerous capabilities surface sound reasonable to non-technical politicians.

In software terms, however, a model file running offline on a developer's local workstation cannot have a remote off-switch. Mandating a remote kill switch by law is, by definition, a backdoor ban on decentralized, open-source software.

When Safety Masks a Balance Sheet Crisis

To understand why this regulatory capture is happening right now, look at the financial statements.

The AI buildout has an ugly secret: the only players printing sustainable, defensible profits right now are the hardware providers selling silicon (Nvidia, TSMC) and the cloud hyperscalers taking a 35% to 40% compute toll on every dollar spent (Microsoft, Amazon, Google).

The foundation labs, meanwhile, are hemorrhaging cash. They are trapped in a brutal cycle of capital dilution, subsidized inference, and diminishing returns on the training run.

When your model improvements begin to plateau, your worst nightmare is that the open-source community catches up. When you have a massive technological lead, you don't care about regulation; you run as fast as you can.

You only call for the referee to blow the whistle when you feel the runners behind you breathing down your neck.

The Linux Battle With Higher Stakes?

We have seen this movie before. In the late 1990s and early 2000s, proprietary software giants tried desperately to convince governments and enterprise buyers that Linux and open-source software were dangerous, insecure, and unfit for serious commerce. They failed, and open-source infrastructure became the foundational bedrock of the modern internet.

The closed-source AI lobby learned that lesson. They realized that market forces alone wouldn't kill open source; open source is simply too resilient, too cheap, and too adaptable.

So they turned to the state.

I thought they were hitting the brakes because the fuel tank was empty. I was wrong. The fuel tank is running low, but they are slamming on the brakes so they can jump out, weld the toll-gate shut, and charge the world rent for the road they didn't finish paving.

It was never about data. It was never about safety. It was about monopoly.