I have this old Mac mini that just runs one thing: a script that pings a server in Taiwan. It’s not for uptime. The script measures the round-trip time, logs it, and does nothing else. It’s my poor man’s seismograph for the semiconductor industry. When that number gets jittery, it means something is happening in the world that will eventually find its way to my desk.
Last week, it got jittery.
The Silicon Heartbeat
Most people think of software as pure abstraction. Code is just text, and it runs on some magical cloud that lives everywhere and nowhere at once. For a long time, that was mostly true. You could build a whole company without ever thinking about the physical box your code was running on, let alone the factory that made the chips inside it.
That era is over. The AI boom isn't just a software revolution; it's a hardware-guzzling machine. The demand for specialized chips—the GPUs and TPUs that power these large models—has become so intense that it’s warping the global supply chain. It's creating a clear divide between those who have access to high-end compute and those who don’t.
For a small studio like mine, this isn't some abstract geopolitical headline. It shows up in very real ways. Cloud providers that used to have a bottomless well of powerful instances now have waiting lists. The cost to rent the beefiest GPUs, if you can even get them, has skyrocketed. Suddenly, the physical location of a data center isn't just about latency for your users; it's about whether you can get the hardware you need at a price that doesn't bankrupt you.
We’re all being forced to think like hardware people now, even if we only write software. The heartbeat of the industry isn't just the pace of new frameworks; it's the output of a handful of foundries in very specific parts of the world.
Designing for the Metal
When you can’t get your hands on unlimited power, you have to get smarter about how you use the power you have. For me, that’s meant a return to a core principle: local-first design. Instead of assuming every hard problem needs to be solved on a server farm, I'm asking what can be done right on the device in my user’s hand.
Apple has been quietly building the perfect platform for this. Their A-series and M-series chips, with their powerful Neural Engines, are little AI powerhouses. They've been putting silicon specifically designed for machine learning tasks into every iPhone and Mac for years. Most of that power sits idle most of the time.
So, instead of building a feature that requires a round-trip to a massive model in the cloud, I’m looking for ways to use smaller, specialized models that can run directly on the device. It’s not just a cost-saving measure, though it is certainly that. It's a better user experience. It's faster. It works offline. And it’s inherently more private, because the user's data never has to leave their phone.
This forces a different kind of design discipline. You can't just throw a giant, general-purpose model at a problem. You have to be more deliberate. You have to think about performance per watt. You have to optimize, compress, and quantize your models until they’re lean enough to run efficiently on a device that fits in a pocket. It’s a constraint, but it’s a healthy one. It makes you a better engineer.
The Real Moat
For years, the startup mantra was to move fast and break things, enabled by infinitely scalable cloud infrastructure. The new reality is that the most valuable resource—specialized compute—is finite and fiercely contested. It’s not just startups fighting for it; it's nation-states and global tech giants.
This changes the game. It used to be that the company with the most venture capital could buy the biggest servers and win. Now, the winners might be the ones who are the most clever about the hardware they have. The real competitive advantage isn't just a smarter algorithm; it's an algorithm that runs ten times more efficiently on a chip that you can actually get your hands on.
I see founders talking about their "AI strategy," and it's usually just a plan to pay a cloud provider for API calls. That's not a strategy; that's just being a customer. A real strategy involves thinking about the entire stack, down to the metal. It means understanding the trade-offs between different hardware architectures. It means building relationships with suppliers, whether they're cloud providers or chip designers.
It also means building software that is resilient to supply chain shocks. If your entire product depends on a single type of GPU that is only made in one factory, you don’t have a business; you have a gamble. Building for on-device processing and diversifying how and where you run your backend isn't just good engineering; it's good risk management.
That little Mac mini in the corner isn’t just a network monitor. It's a reminder that software doesn't float in the ether. It runs on real silicon, made in real factories, and shipped on real boats and planes. The whole magnificent, fragile system depends on it.