Your favorite AI will be gone soon

For years everyone was scratching their head on why didn't Apple create their own Large Language Model. They had the money. They could easily attract the right talent. And nothing happened.

Google also just released Gemma 4, which blew everyone away. And we'll get to how it all connects.

Apple simply outsourced all that AI stuff, showing very little interest in making their own models. Of course that's not entirely true, as they do have a lot of image processing models, image detection and a lot of smaller things that make the iPhone what it is.

They also now have pretty capable foundational (local) models that you can run on your phone. For a lot of basic tasks, they perform surprisingly well.

My app does ALL of the processing locally on device

All of the Longevity Deck processing happens ON DEVICE, because I want to keep the user data 100% private. It does require guidance of the model and detailed .md files, but it works.

There are no publicly available "Apple Frontier models". It seems like they checked out of the race.

But the truth is far more interesting.

The bubble is almost bursting at the seams now…

AI landscape before the bubble pops

AI is a bubble. I don't mean the technology itself. Think of it this way, when the dotcoms crashed, it didn't mean the internet is over. It just meant overinflated products with little value got cut.

This time it's a little different, because we have both the big frontier model companies (OpenAI, Google, Anthropic, Meta) and millions of AI wrappers.

The logical ending to this story is most wrappers ceasing to exist as the bubble pops, just like with the dotcoms. But this time I think it will be fundamentally different.

This time the crash will take out most of the frontier AI companies as well. And who will be left standing on the empty battlefield?

Apple.

Source: Apple

Apple disrupts

They have a history of disrupting multiple industries already. From music production, graphic design, all the way to the music sales industry, content distribution and more.

With AI we're at an early stage, where a handful of big players try to desperately reach any kind of monetization. The competition is fierce. Every week or two a new model comes out that's 0.01% better than the previously leading one. And achieving that fractional supremacy takes Billions of dollars and enough energy to power a small country for a month.

Apple is nowhere to be seen here…

Apple is also one of the only companies not in the now infamous "AI circular economy" graph. What it means is company A pays company B a billion dollars, and then company B pays company A a billion dollars (back). Even though nothing was really exchanged, the economy now has TWO billion dollars in circulation.

This is exactly how values get inflated.

And it hurts even more when it comes crashing right down.

The energy problem

The frontier models are huge and have extreme hardware requirements.

For a while they've been the only way to experience AI. And with the constant FOMO pushed by AI companies, you simply had to subscribe or get left in the dust by the infamous "people using AI".

ALL monthly AI plans are heavily subsidized

The issue is that practically ALL frontier models are subsidized. Your $100 max plan is actually costing the company around $150–200. They want to hook you to be so dependant on the tech, that eventually you'll pay MUCH more just to keep using it. Maybe even $1000.

It's like a dealer giving you a free first hit to hook you on the thing. (not that this actually ever happened)

But these large AI companies are running out of time. Their plot to hook you up and then raise prices may backfire badly very soon.

Especially, since too many vibe-coded apps are being produced now. That leads to people not caring for them / not downloading them. Which discourages the weekend-builders from making more. And from vibe-coding in general.

There's no interest anymore, which leads to people quitting

This is in a way why they all tried to be service-based companies as well. This is the "health push", the failed "sora" video push, the claude cowork. All the things to hook you up with more than just an LLM.

A service that uses the model is a bit less replacable.

All those models and services run on huge, expensive data centers.

This is how Apple will disrupt it.

AI Hardware currently vs AI hardware future

What if you could shrink a data center?

Take that whole big water-cooled building and shrink it down to a Mac mini. Run a perfectly capable LLM locally. Own the data. Privacy has always been championed by Apple so it makes total sense.

Using AI as a tool, on Apple hardware, instead of a cloud service. It's cheaper, more secure, more private and less risky that way.

Of course this mostly is about the early adopters. The tinkerers. The curious. But you know what?

Only 0.03% of people actually pay more than $20 for frontier AI models. Meaning the exact same people will be the first to switch away and go local.

The majority of the world uses no AI at all, with the second largest group using the free, ad-based models.

You are here, thinking the entire world is passing you by

M5, MLX, it all starts here

Apple's MLX framework is designed to make AI models run faster on Macs by taking advantage of how Apple chips handle memory. Unlike traditional systems where data has to be copied back and forth between different processors, MLX lets the CPU and GPU share the same memory directly, cutting out a major speed bottleneck. The latest version of Ollama, a popular app for running AI locally, now uses MLX and sees roughly twice the speed when generating responses.

The newest M5 chips push this even further with built-in Neural Accelerators. It's dedicated hardware specifically designed to handle AI tasks. Apple's tests show these chips can start responding from a 14-billion-parameter model in under 10 seconds, with image generation seeing nearly 4 x speed improvements.

A modern Mac with enough memory can now run serious AI models locally, with no cloud subscription, no usage limits, and no data leaving your machine.

Sure, they're not frontier-model levels yet. But they're infinitely cheaper. Yours. Private.

The future of AI is not API

For a while we all thought using AI will always happen in the cloud. Or via some API in our web wrapper.

This is slowly shifting towards local hardware, with the LLMs themselves being significantly less important part of the equation. All the major ones are now similar in capabilities. And local models are usually a year behind the frontier ones.

Those initial, huge investment backed AI's were a necessary step to get here. But now they're becoming increasingly irrelevant. Their future is not very bright. Still most of the world's population uses no AI models. Only a small percentage pays for AI.

And the group that pays the most right now? This is the exact same group that will go local as soon as this becomes viable.

With latest MLX advancements by Apple, that moment can be this June.

Mac Mini with 128GB RAM would POP the AI bubble

M5 Mac Mini?

An M5 Mac mini with 128GB of ram, priced aggressively would be the final nail in the coffin of AI companies. Sure you can get 512GB ram Mac Studios, but most people will be fine with 128GB. And in the smaller, cheaper form factor this can totally crash the AI bubble.

All those AI bros that ridiculed apple, would be waiting in line to get one and run Gemma 4, Qwen 4 (when it comes) and others.

The companies that helped AI reach the outputs it can do now, will move towards services, or slowly bleed out cash.

And who will come out victorious at the end?

Apple. Like they (most recently) always do.

Betting on LLMs seems to never have been a great idea.