On a Friday at 5:21 p.m., the US government sent Anthropic a letter.
By Friday night, Claude Fable 5, the most powerful AI model on the planet, was gone. Disabled for everyone. No warning, no appeal.
Greg Isenberg had his whole weekend planned around building with it. Instead he spent the weekend ripping frontier models out of his workflow entirely and running everything on his own hardware. Then he turned it into an episode about local ai business ideas: five businesses that only exist because intelligence now runs on your desk for free.
I felt this one personally. I publish three articles a day with AI agents. One vendor decision, one outage, one letter, and my whole content machine stops. So I watched the full episode and pulled out everything worth stealing.
What actually happened (and why it still matters)
The short version, as Greg tells it: government letter Friday afternoon, model dark by Friday night.
Fable 5 is back online as I write this. Doesn't matter. The lesson stuck, and it's bigger than one model.
Every business I know (mine included) is built on models that live on someone else's servers, under someone else's terms. Your agents, your content pipeline, your customer support, your product features. All of it one policy change away from a blank screen.
Greg's line for it: most of the time you're happy being on the electricity grid. It's cheaper, someone else maintains it. But the people who are truly resilient have a generator in the garage. When the hurricane hits and the lights go out, they keep running.
→ Local models are the generator. You don't switch to them full time. You own them so nobody can take them away.
And here's the part most founders have wrong: local models stopped being garbage about six months ago. Greg's estimate is that a model running on a gaming GPU or a decent Mac now handles around 80% of what most people use ChatGPT or Claude for. The gap between free-and-local and expensive-and-cloud closed faster than almost everyone expected.
Local models, explained like you're busy
A local model is an AI model that runs entirely on your own computer. You download the file once, and from that point on it's yours. You don't need internet, an API key, or a per-token budget. It runs on your machine the way a video game does.
That buys you three things cloud models can't sell you:
1. Privacy. Your data never leaves your machine. Nice for you personally, sure. The bigger deal is that it unlocks entire industries (healthcare, legal, finance) that legally cannot send data to a third-party API.
2. Zero marginal cost. Hardware is the only real spend. After that, every query is free. Run a model 24 hours a day for a month and your bill is the electricity. That changes the math on a whole category of products.
3. Nobody can turn it off. The model on your drive works whether or not the company that made it still exists. Whether a government likes it or not. On a plane. In a bunker. It just works.
The stack Greg actually recommends
Most people research the perfect model for weeks before they've run a single one. Wrong order. Download the runtime first:
→ Ollama if you're technical. Command line, one command per model, the developer favorite.
→ LM Studio if you're not. Real interface, point and click, where Greg starts non-technical people.
Then match model size to your hardware, because this is where everyone overbuys or gives up:
→ A 4B model runs on basically anything, including an 8GB laptop and a lot of phones.
→ A 12B model is the sweet spot for a 16GB RAM machine. This is where most people should live.
→ A 27B to 35B model wants a good Mac with 30GB+ of memory or a dedicated GPU. This is where it starts feeling genuinely capable.
→ 70B and up needs serious hardware: a maxed-out Mac Studio, or a dedicated box like Nvidia's DGX Spark with 128GB of unified memory sitting on your desk like a mini data center.
On which model, he keeps it to four families. Qwen 3 (and the new 3.6 series) is the best all-around pick, strong at coding and outperforming previous-generation models four times its size. DeepSeek is the one for hard reasoning and coding problems (expect 10 to 30 seconds of thinking before answers, that's normal). Gemma is Google's small one, fits in 16GB and even on a phone, with surprisingly clean writing. Llama has the biggest ecosystem, so when in doubt, there's a fine-tune for your situation.
Two pro moves separate the builders from the tourists. First: locally, the context window is your real constraint. Cloud hands you a giant window for free, local makes you pay for it in RAM, so keep sessions tight. Second: give your local model tools. A small model with web search, file access, and the ability to run code beats a giant model with none.
And the real unlock is pointing an agent at it. Greg uses Hermes, which is built to run locally and never stop: it remembers everything, writes its own skills, and you message it from Telegram while the heavy work runs on the box at your desk. I broke down his full Hermes agent desktop episode if you want that setup.
5 local AI business ideas that nobody can switch off
This is the part that made the episode for me. These aren't "AI wrapper" ideas. Each one exists specifically because the model runs on hardware you control.
1. On-device AI for regulated industries
Healthcare, legal, finance. They have money, they have problems AI can solve, and they legally cannot send their data to a cloud API.
Build the product where the model runs entirely on the customer's device and the data never leaves the building. That privacy constraint is your moat, because cloud-based competitors can't enter the market at all.
2. The "your data never leaves" version of existing tools
Pick any popular cloud AI product: note takers, meeting summarizers, document analyzers. Build the local version.
Same product, different pitch: nothing you give us touches the internet. Sell it to lawyers, doctors, therapists, anyone handling sensitive documents. Greg's take is that this single sentence on a landing page closes deals the incumbents can't.
3. Air-gapped agents for sensitive operations
Some businesses can't be online at all. Defense contractors, certain financial operations, anyone paranoid about leaks.
An agent setup that runs fully offline on local hardware serves them, and they'll pay accordingly. The nuance: the client doesn't need to be in a sensitive industry. They just need one sensitive operation. That's the niche.
4. Offline AI for places with no internet
Ships. Planes. Rural clinics. Field operations. Disaster zones.
Useful agents that work with zero connectivity are a product the entire cloud industry structurally cannot serve. You're not competing with OpenAI for these customers. Nobody serves them at all.
5. Resilience as a service
This is the meta-play, and my favorite.
After that weekend, every serious company is asking the same question: what happens to our AI workflows if our provider gets cut off? You sell the answer. A fallback layer that kicks in when cloud models disappear.
You're literally selling insurance against exactly what happened with the Fable 5 ban. And like all insurance, it sells best right after the disaster.
My take: the barbell, not the bunker
Greg is explicit that the lesson isn't "cloud bad, local good." Frontier models are still the smartest tools available, and I'll keep using them for the work that needs them.
The lesson is: don't build your entire business on something that can disappear with a single letter. Own a part of your stack.
For me that looks like a barbell. Cloud models for the 20% of work that needs frontier intelligence, a local model on a 16GB machine for the routine 80%: drafts, summaries, cleanup, classification. It costs nothing per query, and nobody can ban it.
It's the same pattern as the tiny AI arbitrage businesses I covered from an earlier Greg episode: the founders who win aren't the ones with the fanciest stack. They're the ones who spot a structural shift early and pick one boring, specific wedge into it.
Don't just nod along either. Greg's homework is the right homework: download Ollama or LM Studio tonight, pull Qwen 3, and force yourself to do one real task fully local. That's when it clicks.
FAQ
What are the best local AI business ideas right now?
The five from Greg Isenberg's episode: on-device AI for regulated industries like healthcare and legal, "your data never leaves" versions of existing AI tools, air-gapped agents for sensitive operations, offline AI for places with no internet (ships, clinics, field operations), and resilience-as-a-service fallback layers for companies afraid of losing cloud access. All five exist because the model runs on hardware the customer controls.
Are local AI models actually good enough to build a business on?
For most routine tasks, yes. The switch happened roughly six months ago. A model running on a gaming GPU or a 16GB Mac now covers around 80% of everyday use cases like drafting, summarizing, and analysis. For hard reasoning you still want frontier models, which is why the smart setup is both: cloud for the hardest 20%, local for everything else.
What hardware do I need to run a local AI model?
Less than you think. A 4B model runs on an 8GB laptop. A 12B model, the sweet spot for most people, wants 16GB of RAM. The 27B to 35B tier needs a 30GB+ Mac or a dedicated GPU, and 70B+ models need a maxed-out Mac Studio or a dedicated box like Nvidia's DGX Spark with 128GB of unified memory.
Was Claude actually banned?
Per Greg's telling: the US government sent Anthropic a letter on a Friday at 5:21 p.m., and by that night Fable 5 was disabled for everyone, with no warning or appeal. The model later came back online. The episode is about the structural lesson, not the news cycle: any cloud model can vanish overnight, so own a fallback.
Steal playbooks like this every week
I break down stories like this on the Profitable Founder Podcast: bootstrapped founders sharing the real numbers behind how they grew, including the parts that went wrong.
If you're building toward $100K MRR, it's the closest thing to sitting in on the conversations I wish I'd had at $15K a month.
