AI is killing the cheap smartphone

One of the most remarkable things about the last few decades is how cheap computers have gotten.In 1985, if you were a reasonably affluent American, the best computer that you could afford was the IBM PC AT. The PC AT would cost you about $6,000—$19,400 in 2026 dollars—and thus represented about a quarter of the median American’s annual income; and it ran on an Intel 80286 processor, capable of something like 900,000 instructions per second. Today, if you find yourself in a market stall in Nairobi or Lagos, you’ll be able to find a cheap smartphone—like the Tecno Spark Go, manufactured by China’s Transsion—for somewhere between $30 and $120. That phone will run on a processor capable of billions of calculations per second.In other words: you can buy a computer thousands of times more powerful than the best consumer device from 40 years ago, for something like 0.3 percent of the price. No other good in history has experienced a decline in cost on that scale: poor people can now carry around in their pockets computers many orders of magnitude more powerful than what the richest slice of the world’s population could afford a few decades ago. And that great cheapening of consumer electronics has enabled a diffusion of computing power to the world’s population that is nothing short of miraculous. Hundreds of millions of the world’s poorest people are able to access the internet because of cheap smartphones like the Tecno Spark Go.That era is now coming to an end.In 2026, the International Data Corporation, which tracks the smartphone market, predicted that worldwide smartphone shipments would fall 13 percent, their largest single-year decline ever. The crash would be most intense in Africa and the Middle East, where smartphone shipments would fall by more than 20 percent, and would be concentrated in the cheapest end of the smartphone industry. This shock represented not a temporary blip but indeed “a structural reset of the entire market”: a huge share of the world’s population is getting priced out of smartphone ownership.So the trend of the last few decades, of consumer electronics getting better and cheaper every year, faces a sharp reversal: the poor world is now entering a smartphone crisis.

This is happening for a simple reason.Smartphones, like other computers, use memory: and the global supply of memory is remarkably inelastic, because memory is really hard to produce. For a long time, most memory went to smartphones and laptops; but in the last few years, AI has emerged as an enormous and hugely profitable consumer of memory. This has resulted in a huge reallocation of memory away from consumer electronics and toward AI. The inevitable result is that smartphones are much more expensive to make now than they were a few years ago. In the short term, this means that the cheap smartphone, which spread computing and internet access to the poorest parts of the world, is dead.But at the rate that things are going, it seems like the poor world will only be the first to get hit. If AI consumption continues to grow at current rates—or if it accelerates, as seems manifestly possible—it won’t be long before the smartphone crisis spreads to the rich world. Consumer electronics are about to get much more expensive.Smartphones are computers. They’re very small computers, and also have things like touchscreens and radios. But in terms of their internal architecture, smartphones are basically the same as what you’d get with a laptop or a server. They have a processor that performs calculations and runs the logic that makes the device do what you tell it to do. They have memory that holds the data that the processor is currently working on. They have storage that retains data when the device is turned off. And they have a circuit board that connects all these different things together.The big story of computing over the last few decades is the processor. You can think of the processor as a huge array of transistors—tiny switches that flip ON and OFF to perform logical operations. We’ve done a good job—a very good job—of figuring out ways to make transistors smaller and more efficient, which means that processors have improved at an exponential rate over the last few decades. This is Moore’s Law.But processors can only process the data that they have access to: and the data that they have access to is what they get from memory: specifically, in modern computers, from DRAM, “dynamic random access memory.”

Here the story is very different. DRAM has gotten better; but it hasn’t gotten better at anything like the rate that processors have: in the 1980s and ‘90s, processor speeds improved at 60 percent per year, while DRAM speeds improved at just 7 percent per year.And that means that for the last few decades, the main bottleneck for computer performance has been memory. Computer scientists call this the “memory wall.” A huge amount of the work in computer architecture over the last few decades has been finding various ways around the mismatch between processors and DRAM.So why hasn’t DRAM improved as fast as processors?Simply put: it’s just a really hard problem. Just like a processor is a huge array of transistors, a memory chip is basically a huge array of memory cells: and each memory cell has both a transistor and a storage unit called the capacitor, which holds the electrical charge corresponding to an individual bit of data. We know how to shrink the transistor. But shrinking the capacitor is a lot harder. As the capacitor gets smaller, it becomes harder for it to reliably store its electrical charge: the charge might leak out, or disappear, or be altered by interference from its neighbors. So if you want to make DRAM more efficient, you need to resort to all sorts of increasingly exotic architectures.And that’s exactly what’s happened. DRAM needs to get more efficient, in order to keep up with the improvements in processors. So modern DRAM manufacturing is an extraordinarily complex and expensive process. Building a single state-of-the-art DRAM fabrication facility, a “fab,” will cost you about $15 to $20 billion; acquiring all the necessary equipment, like lithography tools and etching machines, will cost you another few billion; and then it’ll take you a few years of producing substandard and defective memory chips before your yields start to look competitive.Which leads us to the peculiar economics of the companies that manufacture DRAM: the “memory makers.”The most important thing to know about memory, beyond the fact that it’s expensive and difficult to make, is that it’s fungible. Processors are bespoke: you can’t swap an Intel chip for an Apple chip. But memory chips are not bespoke. DRAM chips all conform to the same industry-wide standards, so a chip from one memory maker will slot into the same device as a chip from any other.

DRAM, in other words, is a commodity.And that combination—capital-intensive manufacturing plus fungibility—is a punishing combination. Because memory is fungible, the industry is intensely cyclical: the entire history of the DRAM industry is a history of boom-and-bust supercycles. First, strong demand from one sector or another—like Windows PC adoption in the 1990s—drives surging prices and a wave of investment from every player; cumulative overinvestment in an undifferentiated good produces oversupply; and then oversupply leads to collapsing prices.And because production is so expensive, those down-cycles turn out to be existential: the memory industry is marked by constant wreckage. Intel dominated the memory game in the early 1970s but left in the 1980s, opting to focus on processors. Texas Instruments and IBM, also once major players, left in the 1990s. Germany’s Qimonda collapsed in 2009; Japan’s Elpida, once the world’s third-largest DRAM manufacturer, declared bankruptcy in 2012.And decades of collapse and consolidation left only a few players standing. In the 1990s, there were perhaps 20 meaningful producers of DRAM around the world; today there are three that account for more than 90 percent of global production. South Korea has two, SK Hynix and Samsung; and the United States has one, Micron.And these memory makers have learned a very particular lesson from the unforgiving history of their industry: always leave demand unmet. The only way to survive in a capital-intensive and cyclical industry was to demonstrate an almost superhuman degree of capital discipline. Demand might rise now, but it would always fall. So it was better to let prices spike and see the marginal memory consumer priced out than to expand production and risk destruction when demand inevitably softened.And this, it turns out, is a brutal calculus for smartphone customers.Earlier, I said that memory is “fungible.” That requires a qualification. Memory is fungible between manufacturers: a chip from Samsung will slot into the same device as a chip from SK Hynix. But that doesn’t mean all computers use memory in the same way.

The MacBook Pro on which I’m writing this piece needs memory that can keep up with a powerful processor running many programs at once: so it uses a standard called DDR, “double data rate,” which runs at a reasonably high voltage and offers high bandwidth. The processor on my iPhone is less powerful, so it needs less data at any given moment; but voltage matters enormously, since every milliwatt allocated to memory is drained from the battery. So smartphones use LPDDR, “low-power double data rate,” a variant of DDR engineered to operate at lower voltages. And in the data centers where Claude and ChatGPT are run, an entirely different standard is used: HBM, “high-bandwidth memory,” which I’ll get back to shortly.All three of these are made the same way, from the same starting material. Memory makers receive thin silicon discs called wafers; over several months, they etch billions of memory cells onto them; and then they cut wafers into individual chips and ship them.The key question facing a memory maker, then, is how to allocate its wafers between DDR, LPDDR, and HBM. Some percentage of wafer allocation is locked in through long-term agreements with major purchasers, like Apple or Dell; and some is sold on the spot market, to buyers who want flexibility or lack the scale for long-term agreement. So every quarter, the wafer allocation teams at Samsung, SK Hynix, and Micron decide—based on prices, contracts, and their best guesses about the direction of future demand—how to distribute their wafers across the three categories.For most of the history of the industry, this allocation was straightforward. In the late 2010s, margins were broadly similar for DDR, LPDDR, and HBM; what interested the memory makers most was volume, and wafer allocation basically tracked end-market demand. Phones were the single largest market for memory, so LPDDR got most of the wafers. DDR took most of the rest. And HBM was a niche product for high-performance computing customers, so it got only a small sliver.That changed dramatically with AI.Training and running AI models is extraordinarily computationally intensive. Even simple queries require billions of matrix multiplications, done in sequence and in parallel, over and over again.