The Improbable Economics of Scarcity, Surplus, and Silicon You know how sometimes you wait in a line outside a popular restaurant, convinced that if the queue is that long, the food must be amazing? Then, two years later, you discover you can just walk in without a wait, and you start to wonder if the food is still as good? Welcome to the world of Graphics Processing Units (GPUs)—the latest victim of economic surrealism where scarcity and hype run a Möbius strip around actual utility. ## **The Great GPU Famine** We're in a GPU shortage. This isn't your run-of-the-mill, "sorry, we're out of pumpkin spice lattes" shortage. This is more like "you’ll trade your firstborn child for an Nvidia RTX 3090" level of scarcity. A whole economy has sprung up around GPUs, and it's got the chaos of a Wild West gold rush mixed with the precision of high-frequency trading algorithms. Miners—of cryptocurrencies, not coal—are buying GPUs faster than you can say "Blockchain." Researchers bemoan the sky-high prices that inhibit the progress of AI research. And gamers? They’ve been reduced to swapping war stories about the "good old days" when GPUs were merely overpriced, rather than mythically unobtainable. ## **An Ominous Surplus** But there’s a plot twist. This shortage will turn into a surplus. How? Two words: efficiency and Apple. But before you rejoice, remember: the restaurant that no longer has a queue isn't necessarily offering the same menu. ### **A Bite of Apple** First off, Apple is in the picture. When Apple decides it's time to jump into a market, it's akin to Oprah giving a book her seal of approval—the landscape shifts. Their custom Apple Silicon is expected to extend its tentacles into AI training. You can imagine the pitch now: "What if the chip that made your MacBook fan obsolete could also train your neural networks in silence?" Apple has the ecosystem, the brand, and—let's not forget—the money to bring some serious innovation to the table. If Apple’s chips can deliver on performance, Nvidia's and AMD’s command of the market might face an existential question. ### **Efficiency or the Art of Doing More with Less** It's not just Apple that could burst the GPU bubble. The very process of AI training is expected to get more efficient. Algorithmic improvements, new hardware-software co-optimizations, and innovative architectures are all ushering us into an era where you might not need an army of GPUs to train Skynet. And if fewer GPUs can do the job, then prepare for a situation where GPUs are as abundant as Netflix sequels. ## **The Plummeting Cost of Thinking Machines** Another angle that doesn't get enough airtime is the cost of inference—the process by which trained models make predictions or decisions. Inference is like the final exam after a semester of training. While the focus is often on the high cost and computational needs of training, inference costs are also dropping. Again, efficiency is the hero we need but didn't ask for. Lower inference costs translate into cheaper, faster, and more accessible AI applications. It democratizes the technology, taking it from the ivory towers of tech giants to the average developer tinkering in their garage. ## **The Uncanny Valley of Surplus** But let's play this out. What happens when the shortage turns into a surplus? Do we trust the restaurant with no line? Will a flood of GPUs—or their more efficient counterparts—devalue the technology? Will innovation slow down if the urgency dissipates? ## **The Economics of Irony** So, here we are. We’re transitioning from a period where GPUs are scarcer than moral integrity in a political drama, to one where we might have so many that we don't know what to do with them. And all along the way, Apple looms large, AI gets democratized, and every CUDA core is having an existential crisis. It's like waiting in line for hours for a roller coaster ride, only to find out there's a second, empty roller coaster that’s faster and more exciting right next to it. The initial shortage created the market, the impending surplus might just mature it, and Apple—like a rogue roller coaster operator—might flip the switch that changes the game altogether. As for me, I’m excited to see how this all pans out. In an economy of ironies, the only thing you can reliably predict is unpredictability. And that, dear reader, is the improbable economic ballet of GPUs, Apple, and the ever-elusive promise of efficient computation. Stay tuned, or rather, stay queued. Thanks for reading Matt’s Substack! Subscribe for free to receive new posts and support my work.