Open Source AI Will Do to Anthropic and OpenAI What the Open Internet Did to AOL

We've seen this movie before. In the mid-1990s, AOL had 30 million subscribers and a stranglehold on how America experienced the internet. It was fast, polished, and carefully curated, a walled garden that felt like the future. Then the open internet arrived and ate it alive. Today, Anthropic and OpenAI are playing AOL's part. And open source AI is the open internet.

The Walled Garden Always Loses

AOL's genius, and its fatal flaw, was the same thing: control. Control over the experience meant control over the business model. That control also meant an inherent ceiling on what the platform could become, because innovation that happened outside AOL's walls couldn't get in.

Anthropic and OpenAI have built the same kind of garden. Claude, GPT-4, and their successors are extraordinary models, but they're black boxes. You can't run them on your own infrastructure. You can't inspect them, fine-tune them to your domain, or guarantee they'll be there tomorrow. That dependency is precisely what enterprises are now waking up to.

The Fable 5 incident made it visceral. Anthropic suspended access to a model, not because of anything their customers did, but because of a political and regulatory decision made in Washington. Enterprises that had built workflows on top of Claude were rug-pulled overnight. That's not a bug in the system. That's a feature of closed AI: someone else holds the off switch.

The Regulation Play

Here's the part that doesn't get talked about enough: the frontier AI labs are not passive victims of government regulation. They're actively lobbying for it.

This is an old playbook. Large incumbents push for regulation they know will create compliance costs that only they can absorb. With enough regulatory barriers, only the largest players can remain. This is why Anthropic's executives are constant fixtures in Washington. Safety is real, but the strategic value of being the "responsible" AI lab that welcomes oversight is also very real.

Open source, by definition, can't be regulated the same way a commercial API can. You can't issue a government directive to a model weight sitting on someone's server in Germany. This is precisely why the incumbents and their regulatory allies treat open source AI as a threat.

The Infrastructure Shift Is Already Happening

Three things are converging right now that will accelerate local and on-premise AI in ways that were unthinkable 18 months ago:

Apple's MLX framework has made it genuinely practical to run capable models on Apple Silicon. A MacBook Pro can now run a 70B parameter model, not blazingly fast, but fast enough for real work. The developer experience is polished. The hardware is already in the hands of millions of engineers.

NVIDIA's RTX Spark laptop line is pushing high-performance AI inference onto edge devices. An RTX 5090 laptop is not a consumer toy; it's a workstation that fits in a bag and runs models that would have required a data center rack three years ago.

And models like Llama 3, Mistral, Qwen, and DeepSeek have closed the quality gap with the frontier labs at an alarming rate. Six months ago you'd accept a meaningful capability trade-off to run local. Today, for most enterprise use cases, you barely notice.

History Rhymes

The computing industry has lived this transition twice already.

The first time was the shift from mainframe to PC. IBM dominated the mainframe era completely. When the PC arrived, the smart bet looked like IBM; they had the enterprise relationships, the distribution, the brand. But the open architecture of the PC ecosystem meant that innovation could happen everywhere at once. IBM never recovered as a hardware company.

The second time was the shift from proprietary Unix servers to Linux. Sun Microsystems, SGI, and HP built extraordinary Unix boxes. They were faster and more reliable than anything else. But Linux, free, open, good enough, and improvable by anyone, ate the market. Sun is gone. HP sold its server business. Linux runs the internet.

The pattern is consistent: a period of legitimate dominance by a closed, polished, well-resourced incumbent, followed by an open alternative that's initially "not quite as good," then by the open alternative improving faster than the incumbent can, culminating in the incumbent's irrelevance.

What This Means If You're Building

If you're a founder or a CTO making infrastructure decisions today, the Fable 5 incident should be a forcing function. Not because you need to abandon hosted AI, but because you should be building with optionality.

Design your AI layer to be model-agnostic. Use abstraction layers that let you swap between hosted and local models. Know which of your use cases can tolerate a disruption in hosted API access and which ones can't. For those who can't, start running local models now while it's still a proactive choice rather than a crisis response.

The enterprises that get this right will have a durable competitive advantage. The ones that build deep dependencies on any single closed model provider are making a bet that the off switch never gets flipped on them.

The Bottom Line

Anthropic and OpenAI are impressive companies building genuinely powerful technology. This isn't a critique of the quality of their work. It's a critique of the relationship's architecture that they're asking the market to enter.

AOL was also impressive. AOL built real things that real people used and loved. AOL lost anyway, not because it was bad at what it did, but because the open alternative was better at being a platform for everyone else's innovation.

Open source AI is that platform now. The question isn't whether it will win. The question is how long you want to wait to start betting on it.