So far we've talked about the unsustainable economics, circular money, and the AGI gamble.
There's another threat to the mega models that nobody's talking about. It's the threat from below.
While OpenAI, Anthropic, and Google burn billions building massive general-purpose models, something sneaky is happening.
Small, cheap, local models are starting to eat the market from the bottom.
Here's A Real Example
Austin, Texas is using an on-premise AI system for residential permits.
They aren't using GPT-5 or Claude. They're using a local model running on their own servers.
It's smaller. It's less capable overall. But it does the thing they care about very well.
And it's financially sustainable.
All they're paying is compute costs. No API fees. No subscriptions. No dependency on a company that might 10x their prices next year.
The 95% Failure Rate
Remember that MIT study? 95% of AI pilot projects fail.
Part of that failure is cost. Companies are trying to integrate these mega models into their workflows, and even at today's artificially low prices, the math doesn't work.
But what if those same projects used cheap local models instead?
Maybe some of those 95% would have succeeded.
That's the mega-model dilemma.
The Impossible Position
OpenAI and others are stuck.
Their models are expensive to build. Expensive to run. They're subsidizing usage now, but eventually prices have to go up.
When that happens, companies will start asking: Do we really need GPT-5 for this? Or can a smaller model do the job?
For a lot of use cases, the answer is going to be the smaller model.
I'm Seeing This With Clients
I've built automation systems for clients using both approaches.
One client needed AI to process customer inquiries. We tested GPT-4. It worked great. It was also expensive at scale.
We tested a smaller local model fine-tuned for their specific use case. It worked almost as well. Cost was a fraction.
They went with the local model.
That's happening more and more. The mega models are impressive. But impressive doesn't always mean practical.
The Market Is Splitting
Mega models are at the top. Chasing AGI. Burning cash. Hoping for a breakthrough.
Cheap focused models are at the bottom. Solving real problems at sustainable costs.
And the bottom is growing faster.
Every month, new local models get released. Llama. Mistral. Smaller versions that run on consumer hardware.
They're not as capable as GPT-4 or Claude. But they're good enough for specific tasks. And good enough at 1/10th the cost is a compelling value proposition.
The Threat To OpenAI
The threat to OpenAI isn't only running out of money.
It's becoming irrelevant for most practical use cases while they chase a moving target that might never arrive.
If AGI doesn't come, and local models keep getting better at specific tasks, what's the value proposition for expensive general-purpose models?
You're paying premium prices for capabilities you don't need.
What This Looks Like In Practice
Small businesses don't need GPT-5. They need something that automates their invoicing or handles basic customer service.
Local governments don't need AGI. They need systems that process permits or answer citizen questions.
Most companies don't need a model that can do everything. They need a model that does their specific thing really well.
That's what local models are becoming. Specialized tools for specific jobs at sustainable prices.
The Inconvenient Truth
OpenAI's entire business model depends on general-purpose models being so much better than specialized models that the price premium is worth it.
But as local models improve, that gap narrows.
And when the gap is small enough, most companies will choose the cheaper option.
Not because it's better. Because it's good enough.
What I'm Doing
I'm building with both.
For complex, variable tasks that require real flexibility, I use Claude or GPT-4.
For repetitive, well-defined tasks that don't change much, I'm increasingly using local models.
That's the practical approach. Use the right tool for the job, not the most expensive tool for everything.
And honestly? Most business automation falls into the second category. Repetitive, well-defined tasks.
That's local model territory. And it's getting bigger.
The Bottom Line
The mega models aren't going away. They're useful for certain things.
But the assumption that everyone needs them for everything? That's looking shakier every month.
The threat from below is real. And it's growing while the mega models burn billions chasing a breakthrough that might not come.
Next up in Part 6: What happens when the bubble pops. The loud pop versus the slow deflation. Neither are pretty.