I have to admit something.
I was one of the peanut gallery saying Apple missed the AI bus.
When generative AI exploded over the last couple of years, the story seemed pretty obvious. OpenAI had ChatGPT. Microsoft wired it into everything. Google scrambled to catch up. Meta opened the gates with Llama. NVIDIA became the arms dealer to the entire AI economy. The industry narrative formed quickly: Whoever controlled the GPUs and the data centers would control the future of AI.
Apple didn’t appear to be part of that story. They weren’t announcing giant model training clusters. They weren’t talking about trillion-parameter models. They weren’t spending tens of billions on data centers. If anything, it looked like Apple was late to the party and would eventually have to do what many expected—use their enormous cash pile to acquire their way into the game.
Well, I may have been wrong.
Again.
Because once again, Apple might be either the luckiest tech giant in the world or they are playing three-dimensional chess while much of the rest of the industry is playing checkers. (Well, maybe not Jensen Huang. When you reach a $5 trillion market cap, you’ve clearly figured a few things out.)
A chart that has been circulating on X this week captures the moment nicely. Investor Josh Kale posted a graphic showing standardized quarterly capital expenditures across the major tech companies. The picture is striking.
Amazon’s quarterly capex is approaching $40 billion. Microsoft and Alphabet are both around the $30 billion mark. Meta has climbed past $20 billion. Collectively, these companies are pouring well over $100 billion per quarter into infrastructure, much of it aimed at building the massive AI data centers required to train the next generation of models.
Apple, on the other hand, is moving in the opposite direction. According to the chart, Apple’s capex spending is actually down roughly 19%.
In other words, while the hyperscalers are engaged in what looks like an AI infrastructure arms race, Apple is largely sitting it out.
At least on the surface.
Here’s the original post for context.
What makes the situation fascinating is what’s happening outside those data centers.
Developers are increasingly running AI models locally on Apple hardware. Mac minis have been selling out in some channels. Mac Studios are showing multi-week backlogs. The new Apple Silicon machines with large unified memory pools are capable of running surprisingly large models directly on the device.
This isn’t just internet chatter. We’re seeing it ourselves.
We ordered several Mac minis this week alone. The reason is simple: They make excellent small servers for running agents. They’re quiet, energy efficient, relatively inexpensive and surprisingly capable when paired with modern local inference frameworks. The problem we’re running into now is availability. Mac minis aren’t always easy to get quickly, and Mac Studios are backlogged in some configurations. As a result, we’ve started scavenging older machines. Any M1 Mac or better can serve as a perfectly capable AI node.
That’s a strange sentence to write in 2026.
The reason this works comes down to Apple Silicon’s architecture. Apple’s unified memory model allows the CPU and GPU to share the same memory pool. That means large models can use system memory rather than relying solely on dedicated GPU VRAM. A Mac with 128GB of unified memory can run workloads that would require much larger and more expensive GPU setups in traditional configurations.
Frameworks like Apple’s MLX and the broader ecosystem forming around local inference are making these machines increasingly attractive for developers experimenting with agents, private models and device-level AI.
When you step back, the irony becomes hard to ignore. The company spending the least on AI infrastructure may have accidentally backed into becoming part of the AI infrastructure.
Instead of building giant centralized data centers, Apple spent the last decade building hundreds of millions of powerful edge computers. Macs, iPhones and iPads now ship with custom silicon optimized for performance per watt and machine learning workloads. Those devices sit on desks, in backpacks and in pockets across the world.
Now the industry is discovering that many AI workloads don’t actually need a hyperscale data center. They need a capable local machine.
To be clear, the hyperscalers are not wrong. Training frontier models absolutely requires enormous infrastructure. The compute required to build systems like GPT-class models or next-generation multimodal AI is staggering. Those workloads will continue to demand vast clusters of GPUs and massive data center footprints.
But the operational side of AI—the part where models are run, agents execute tasks and private workloads stay close to the user—may look very different.
That’s where Apple’s device ecosystem starts to look unusually well-positioned.
Apple has followed a pattern like this before. The company largely avoided the enterprise server race that defined much of the early cloud era. It never attempted to compete directly with Amazon or Google in building hyperscale infrastructure. Apple has always insisted its primary focus was consumer devices.
Yet walk into most developer organizations and you will still see Macs everywhere.
Apple’s approach has often been to skip the first phase of a technology wave and focus instead on the part that turns it into a product people actually use every day. They were not first to smartphones, tablets or smartwatches. But once they entered those markets, they redefined them.
AI may end up following a similar arc.
Of course, this could still go the other way. The companies building massive AI infrastructure today may ultimately dominate the entire ecosystem. If the future of AI is defined primarily by giant centralized models running in hyperscale environments, Apple risks remaining a secondary player.
But if the future includes millions of smaller models operating closer to the user—inside applications, inside devices and inside organizations—the picture starts to look different.
In that scenario, Apple’s strategy doesn’t look slow. It looks patient.
Whether this outcome was intentional or accidental is almost beside the point. The reality is that while the rest of the industry is spending enormous sums trying to build the AI future, Apple may have already distributed a large portion of the hardware that will run it.
Crazy like a fox.
And somewhere Steve Jobs is smiling.
