Transcription
You know, everyone's talking about Nvidia and cloud AI data centers or token factories, assuming that's the model that we're going to be using in the future. In fact, the CEO of Microsoft was at Davos telling governments that AI would eventually become a public utility that was connected to the grid and the communications.
A long-term scalable solution is to have, uh, you know, all of these token factories part of the real economy connected to the grid. He's effectively promising that AI will always be delivered as a cloud service, which is, of course, conveniently integrated with Microsoft's cloud infrastructure. But I think he's wrong, and that won't be the model that ends up winning. In my view, the real AI race is being won by Apple, and almost nobody is paying attention.
Let's forget the hype about models for a second. AI is mostly just mathematics. Matrix multiplication, vector operations, linear algebra, billions of calculations per second. CPUs are terrible at this. They're built for logic, not repetition. And most computers today ship without dedicated GPUs. That's a decision that Intel made years ago when they dominated the PC industry. They bet on integrated graphics, which is a half-ass approach that they thought would be good enough for most users. They let the discrete graphics market become someone else's opportunity. And that someone became Nvidia.
For years, Nvidia made GPUs for gamers. It was a niche market that Intel didn't care about. But then deep learning arrived, and suddenly those gaming chips turned out to be exactly what AI needed. Thousands of tiny cores doing the same calculation simultaneously. Nvidia's side business became the backbone of the entire AI industry, and today they're the world's most valuable company. Meanwhile, Intel is fading away like IBM and other giants of that era.
Now, GPUs are great, but the problem with them is that programming them is a nightmare. It's not just the hardware, it's the drivers and the language. Everything about dealing with GPUs was difficult. So, Apple said, "Let's scrap all that. We've got a better idea."
About a decade ago, Apple made a decision that at the time seemed crazy. They decided to abandon Intel, the giant of the chip industry, and go their own way. It was audacious because Intel had dominated the semiconductor industry for decades, and they were the undisputed leader in chip design and manufacturing. But by 2020, the first Apple M1 chips arrived, and they were an instant hit. Laptops ran significantly faster and had dramatically better battery life, and everyone was stunned.
The secret was a completely different chip architecture. You see, Apple put the CPU, GPU, and neural engine all on one chip and shared the same unified memory pool. There was no copying data between processes and no bottlenecks moving around information, which meant they could send GPU-intensive jobs to the GPU, and send CPU-intensive jobs to the CPU, and then use the neural engine to pick up the slack for matrix multiplication. So when large language models exploded, the Apple ecosystem was already ready. Developers like me could just start coding directly into the GPU.
And now AI tools like Claude have made GPU programming dramatically easier. I remember a few years ago when I started developing the Kira algorithm, just how hard it was to get GPU programming right. And that's because it works in a completely different way to what most people are familiar with. Instead of feeding it instructions one at a time, with GPU programming, you need to load all the data into a buffer, prepare a command, and then execute it all at once, and then decode the output. This makes it very hard to get right, and it would often take me days just to debug a simple crash. But with tools like Claude, if you know what you're doing and you precisely describe what you want, you can iterate really quickly. Now, it's still not perfect, but AI coding makes GPU processing so much easier.
So, this means that the Apple ecosystem has a huge structural advantage for insanely fast, insanely cheap matrix multiplication. And guess what? Matrix multiplication is at the very heart of AI. So when Sachin Nadella talks about these giant publicly funded AI token factories that he wants the governments to build, I think he's dead wrong because the unit economics simply doesn't stack up. If you can run matrix operations effectively locally on your device, you don't need the cloud. You don't need to be paying for these massive GPU clusters that burn cash and electricity and move data backwards and forwards all the time. You use the AI right there on the device that you're using. Whether it's a phone, a watch, a tablet, or a laptop, it's about as efficient as you possibly get because the cost and latency are almost zero.
Microsoft has gone in on a model that, to be honest, doesn't make sense. macOS is winning in AI because Apple removed the friction between CPU, GPU, and the neural engine. And for developers like me, building an application with AI baked in has become remarkably straightforward. Apple's core machine learning framework is natively integrated into the hardware and the operating system, and this dramatically cuts development and testing time. Meanwhile, Windows is still betting for a world where GPUs live in data centers, and they're stuck in that model because of the relationship they've had with Intel all these years. And that's why Intel and Microsoft are in such big trouble.
But of course, if you listen to social media, you'd think that Apple has lost the AI race. And to be fair, Siri is a huge disappointment. And Apple Intelligence, well, I'm not even sure what to make of that. But I still believe that Apple will ultimately be the real winner of AI. Not because of the models, but because of the chips.
You know, there's nothing magical about AI. If you strip away the hype, it's just an advanced programming technique that allows developers to solve old problems in new ways. Now, large language models get all the attention, but these models are just software, and software relies on hardware to run. The closer the hardware is to the user, the faster, more private, and secure the experience becomes. Apple understood this long before anyone else. So, the M-series chips turned every Mac into a serious AI machine before most people even realized that it mattered. And while big tech was focused on who had the biggest data center, Apple quietly put an AI engine in every laptop and every device. There are now hundreds of millions of devices all around the world capable of running AI locally. And that's how Apple wins at AI. Apple wins not by depriving communities of electricity and water, but by turning the devices that you already hold into powerful matrix multiplication engines.
If you enjoy these videos, please consider subscribing to my channel for a very contrarian view of the tech.