On June 21, 2020, I published Apple Chips == AR Glasses. I thought Apple's rumored move away from Intel was bigger than a CPU swap. If Apple controlled the CPU, GPU, Neural Engine, cameras, and operating system, it could build the next computing platform without waiting for anybody else.

The timing still makes me smile. Apple announced its Mac transition to Apple Silicon the next day.

Six years later, the interesting part is not simply whether the prediction was right. Apple did replace Intel processors and AMD graphics across the Mac line. Apple also built Vision Pro, which is almost exactly the device I described as something between a VR headset and HoloLens. The interesting part is where I placed the center of gravity. I thought the next platform would be on our faces. Instead, the largest change happened in data centers, where NVIDIA chips train and run the models that are changing how we use every other computer.

What the Prediction Got Right

The central bet was vertical integration. A company that designs the chip, software, and device can make trade-offs that are difficult when those parts come from separate vendors. The M-series Mac proved that point. CPU, GPU, memory, and Neural Engine became one system rather than a collection of replaceable parts.

That integration did not make the Mac less powerful, as I worried it might. It made the Mac more interesting. Unified memory became useful not only for graphics but also for running models locally. The Neural Engine stopped looking like a feature for a few phone tasks and started looking like an early signal of where all personal computers were going.

The sensors were also a signal. Multiple cameras and LiDAR did become part of Apple's spatial-computing system. Vision Pro uses cameras, eye tracking, hand tracking, custom silicon, and software together. I got the shape of the machine mostly right.

I got the use wrong. Vision Pro did not arrive as the Nintendo competitor I imagined. Apple positioned it as a spatial computer, and its size and price made it a very different product from a mass-market game console. AR glasses may still become an everyday computer, but the difficult parts are now obvious: battery, heat, weight, display quality, social acceptance, and a reason to wear the device all day. A headset demonstration and a durable platform are not the same thing.

NVIDIA Moved the Center of Gravity

In 2020 I was looking at chips from the device inward. The AI boom forced me to look from the data center outward.

NVIDIA did not win this position with a GPU alone. CUDA gave developers a programming model. Libraries made common operations fast. Systems such as NVLink and DGX connected accelerators into larger machines. Frameworks, cloud providers, researchers, and companies built around that stack for years. By the time large language models created enormous demand for parallel computation, NVIDIA was selling a working system, not an isolated component.

The financial scale shows how quickly that system became infrastructure. For its fiscal 2026, NVIDIA reported $215.9 billion in annual revenue. Its fourth-quarter data center revenue alone was $62.3 billion. Those numbers are not just a story about chip sales. They show where the industry is spending to create the next layer of computing.

Apple's integration is optimized around a person carrying or wearing a device. NVIDIA's integration is optimized around racks of accelerators behaving like one computer. Both strategies connect hardware and software tightly, but they operate at different scales. Apple asks how much intelligence can fit within a power and privacy budget on the device. NVIDIA asks how much computation can fit within the networking, cooling, and power budget of a data center.

NVIDIA is not alone. Google has TPUs. Amazon has Trainium and Inferentia. Microsoft has Maia. These companies are repeating the same lesson from Apple Silicon: when a workload becomes important enough, designing the hardware and software together becomes a strategic advantage. GPUs remain valuable because models and methods are still changing quickly. Custom accelerators become valuable when a company runs a large, predictable workload and can optimize the whole path.

The competition is therefore larger than NVIDIA versus AMD. It is CUDA and a broad ecosystem versus vertically integrated cloud systems, each with its own chips, models, networks, and customers. NVIDIA has the strongest general platform today. The cloud companies have scale, internal demand, and a reason to reduce their dependence on any one supplier. Both can be true at once.

The Computer Now Includes the Power Plant

The phrase "state of AI" can make this sound like a competition between models. The models matter, but they sit inside a physical system. A model needs chips. The chips need high-bandwidth memory and networking. The racks need cooling. The data center needs land, transformers, and electricity.

The International Energy Agency estimates that global electricity demand from data centers will more than double by 2030 to about 945 terawatt-hours. That is a reminder that AI is not weightless software. The constraint can move from model architecture to chip supply, then to memory, networking, cooling, or the local electrical grid.

This also changes what "on-device AI" means. Local models can reduce latency, protect private data, work without a network, and avoid paying a cloud inference bill for every interaction. Cloud models can be much larger and can draw on specialized infrastructure. The future is probably not one replacing the other. It is a negotiated boundary. Devices will do the work that benefits from being close to us, while data centers will do the work that benefits from enormous shared computation.

That brings me back to glasses. The useful version of AR glasses may depend less on putting a complete supercomputer on somebody's face and more on dividing the work among sensors, a phone, local models, and cloud models. The glasses become an interface to intelligence that lives across several computers. In that sense, Apple chips may still lead to AR glasses, but NVIDIA and the rest of the AI infrastructure world are part of the path too.

The 2020 prediction taught me that the visible product is often downstream of a less visible capability. I saw Apple's control of silicon and correctly expected a new device, but I underestimated how much the next platform would depend on data-center systems, software ecosystems, and power. The quest is no longer only to put a computer on our faces. It is to decide what belongs on the device, what belongs in the data center, and who controls the system connecting them.