What Everyone Is Missing About the M5 MacBook

While reviewers argue over benchmarks, Apple’s eight-year bet on the Neural Engine reveals exactly where the computing industry is headed.

A new MacBook Pro shipped. And the reviews are full of what you’d expect — Geekbench scores, render times, frame rates. The numbers people know how to compare.

But there’s a 16-core neural engine inside that machine that barely gets mentioned. Apple has been making it bigger every generation since 2017. And they’re not alone. Qualcomm, Intel — every major chip company on Earth is quietly dedicating more and more transistors to the same type of component.

When companies that aren’t aligned on much else all start making the same investment at the same time, that’s not a product feature. That’s a signal. And if you know how to read it, you can see exactly where the industry thinks it’s going.

Photo by Ales Nesetril on Unsplash

The Vertical Stack No One Else Has

Apple left Intel in 2020 with the M1. At the time, people treated it like a gamble. Could they really design a laptop chip as good as Intel’s?

Six years later, the answer is obvious. But the how is more interesting than the whether.

Most chip companies design a processor, license it out, and let manufacturers build around it. Intel designs the chip. Dell puts it in a laptop. HP puts it in a different laptop. Separate companies, separate decisions, separate optimizations.

Apple does it differently. They design the chip. They design the memory architecture. They write the operating system. They build the developer frameworks. One company, one stack, from the transistor all the way to the API.

Take unified memory. On a traditional PC, the CPU has its own memory and the GPU has its own memory. When data needs to move between them, it gets copied. That takes time and power. On the M5, the CPU and GPU share one pool. No copying. The data is just there. Apple could do this because the team designing the chip is the same team designing the operating system’s memory scheduler.

That’s not a technical advantage you can buy. It’s a structural one that takes years to build.

A Signal Written in Silicon

In 2017, Apple put the first neural engine in the A11 chip inside the iPhone X. Two cores. 600 billion operations per second. All it did was power Face ID and Animoji. Developers couldn’t even access it — Apple kept it locked behind their own software.

One year later, the A12 jumped to eight cores. 5 trillion operations per second — nine times faster at a tenth of the power. This time, Apple opened it up with a framework called Core ML, letting any developer run machine learning models on the neural engine.

2020: 16 cores, 11 trillion operations per second. Same year, the M1 shipped with an identical neural engine, bringing it to the Mac for the first time.

M4: 38 trillion operations per second.

M5: A faster 16-core neural engine, plus new neural accelerators embedded inside the GPU cores for the first time.

Follow that arc. Two cores to sixteen. 600 billion to 38 trillion. Every single generation, Apple chose to dedicate more transistor budget to this component. And transistors are finite — every square millimeter given to the neural engine is taken from the GPU or CPU.

They kept making that trade-off for eight years. Not because Face ID needed it. Because they believed the future was going to demand it.

Local vs. Cloud — The Debate That’s Already Decided

There are two schools of thought about where AI should run.

One says the real work happens in data centers. You write a prompt, it travels to a server farm full of GPUs, the model runs, and the answer comes back. The models are too massive to run anywhere else. This is the architecture behind every ChatGPT, Claude, and Gemini API call.

The other says meaningful AI should run locally — on the device in front of you, private by default. No round trip to a server. No API cost per call. No dependency on someone else’s infrastructure. Your data never leaves the machine.

Neither side is going to win outright. The future is a mix — some workloads will run locally, some in the cloud, and developers will decide which is which. But Apple is building the strongest case anyone has ever built for the local side. A dedicated neural engine. Neural accelerators in the GPU. Unified memory so models don’t waste time copying data. A mature developer framework that routes workloads across all of it automatically.

From a privacy perspective, this matters. Cloud-based AI requires trust — trust that your data isn’t being logged, trained on, or leaked. Local AI doesn’t ask for that trust. It just doesn’t send the data.

That distinction matters more as these capabilities grow.

The Gap Is Where the Opportunity Lives

Right now, most neural engines spend most of their time waiting for something to do. Most developers haven’t shipped their first Core ML model yet. Most of the applications that will justify this hardware investment haven’t been fully built.

That gap — between what the silicon can do and what the software has been written to do — is where the opportunity lives.

Marketing can say anything. Earnings calls can promise anything. But the chip is the actual commitment. You can’t fake transistors.

You can’t fake what a company chose to dedicate its most expensive resource to.

Apple has been investing in on-device AI silicon for eight years. The M5 is generation five of the M-series approach, generation eight of the neural engine thesis. And every generation, the ceiling gets higher.

The story the chip is telling is simple: the industry believes AI compute belongs on your device, not just in the cloud. The only question is whether the software will catch up before the next hardware generation makes the question irrelevant again.

If this post resonated with you, buy me a coffee ☕ — it helps me continue sharing stories, ideas, and reflections.