On October 28, SK hynix posted financial results so extraordinary they barely seemed to come from a chip company: third-quarter revenue of 104.6 trillion won, or about $78.9 billion, up 328% year over year; operating profit of 83.3 trillion won, with an operating margin of 80%. In the same week, SanDisk also reported an operating margin of 80%, with revenue of $11.03 billion, up 378% year over year. Both companies delivered the same message on their earnings calls: there isn’t enough memory to sell—and this isn’t a shortage that will last just a few months. SK hynix has extended its shortage forecast all the way to 2030. More than half of its capacity for 2027 and 2028 has already been locked up in advance by 14 major customers through long-term agreements, and high-bandwidth memory (a type of high-speed memory stacked alongside AI chips to feed AI accelerators) is essentially sold out through 2029.
Price increases have made their way from earnings reports to store shelves. TrendForce forecasts that contract prices for conventional memory will rise another 10% to 15% quarter over quarter in the fourth quarter, while flash memory prices will climb 15% to 20%. Retail prices are even more extreme: a pair of 16GB DDR5 memory sticks cost less than $100 at the start of 2025, but now go for more than $400; a 32GB kit has jumped from under $250 to over $1,200; and a 128GB kit has soared from $329 to $3,399. The price of an entire PC from ten years ago now buys just two memory sticks. Server memory hasn’t escaped either: 64GB DDR5 server memory sticks have risen from $380 to $820–$890.
The chain of cause and effect is clear: AI data centers need high-bandwidth memory to feed accelerators from Nvidia and Advanced Micro Devices, and high-bandwidth memory competes with ordinary memory for the same production lines. Producing 1GB of high-bandwidth memory takes up about four times as much wafer area as producing ordinary memory. Samsung Electronics, SK hynix, and Micron Technology control more than 90% of global memory production capacity, and they’re doing the math: each wafer of high-bandwidth memory generates three to five times the revenue of ordinary memory, so capacity naturally goes to AI first. PC and phone makers have to bid against one another for what’s left. Microsoft has already raised the price of its Surface PCs by as much as $500 this year, while Apple raised Mac and iPad prices in June, explicitly citing memory costs. Intel is in the worst position: its PC chip business fell 5% to $8.1 billion in the third quarter. Its CFO said the PC slowdown would last at least through the first half of 2027, while its CEO expects the PC chip market to shrink by another 12% in 2027.
This isn’t an ordinary price-hike cycle. Past memory shortages were caused by supply-and-demand imbalances and could be eased by building factories over a couple of years. This time, the shift is structural: capacity has been permanently diverted to high-bandwidth memory. As long as AI keeps burning through cash, ordinary memory will be stuck at the back of the queue. This time, the shovel sellers aren’t just selling shovels—they’ve also taken control of the price of the shovels, and an 80% profit margin is proof. Before 2027, memory prices have nowhere to go but up. High memory prices will choke off the PC replacement cycle, and building a PC, replacing a computer, or buying a phone will only get more expensive over time.
But don’t treat “shortages through 2030” as a perpetual-motion machine. The three major manufacturers will spend more than $80 billion on memory capital expenditures in 2026, with new capacity set to come online mainly in 2027 and 2028. If AI capital spending growth slows, or high prices completely crush demand for PCs and phones, the flip side of an 80% profit margin will be a tombstone marking the top of the cycle. The real turning point is 2028: before capacity comes online, price increases are a certainty; after it does, only those still holding long-term contracts will have reason to smile. The memory battle is now about who can sell their 2029 supply first—the ones who’ve sold it won’t fear falling prices; the ones with no inventory will.
#人工智能 #MemoryPrices
Price increases have made their way from earnings reports to store shelves. TrendForce forecasts that contract prices for conventional memory will rise another 10% to 15% quarter over quarter in the fourth quarter, while flash memory prices will climb 15% to 20%. Retail prices are even more extreme: a pair of 16GB DDR5 memory sticks cost less than $100 at the start of 2025, but now go for more than $400; a 32GB kit has jumped from under $250 to over $1,200; and a 128GB kit has soared from $329 to $3,399. The price of an entire PC from ten years ago now buys just two memory sticks. Server memory hasn’t escaped either: 64GB DDR5 server memory sticks have risen from $380 to $820–$890.
The chain of cause and effect is clear: AI data centers need high-bandwidth memory to feed accelerators from Nvidia and Advanced Micro Devices, and high-bandwidth memory competes with ordinary memory for the same production lines. Producing 1GB of high-bandwidth memory takes up about four times as much wafer area as producing ordinary memory. Samsung Electronics, SK hynix, and Micron Technology control more than 90% of global memory production capacity, and they’re doing the math: each wafer of high-bandwidth memory generates three to five times the revenue of ordinary memory, so capacity naturally goes to AI first. PC and phone makers have to bid against one another for what’s left. Microsoft has already raised the price of its Surface PCs by as much as $500 this year, while Apple raised Mac and iPad prices in June, explicitly citing memory costs. Intel is in the worst position: its PC chip business fell 5% to $8.1 billion in the third quarter. Its CFO said the PC slowdown would last at least through the first half of 2027, while its CEO expects the PC chip market to shrink by another 12% in 2027.
This isn’t an ordinary price-hike cycle. Past memory shortages were caused by supply-and-demand imbalances and could be eased by building factories over a couple of years. This time, the shift is structural: capacity has been permanently diverted to high-bandwidth memory. As long as AI keeps burning through cash, ordinary memory will be stuck at the back of the queue. This time, the shovel sellers aren’t just selling shovels—they’ve also taken control of the price of the shovels, and an 80% profit margin is proof. Before 2027, memory prices have nowhere to go but up. High memory prices will choke off the PC replacement cycle, and building a PC, replacing a computer, or buying a phone will only get more expensive over time.
But don’t treat “shortages through 2030” as a perpetual-motion machine. The three major manufacturers will spend more than $80 billion on memory capital expenditures in 2026, with new capacity set to come online mainly in 2027 and 2028. If AI capital spending growth slows, or high prices completely crush demand for PCs and phones, the flip side of an 80% profit margin will be a tombstone marking the top of the cycle. The real turning point is 2028: before capacity comes online, price increases are a certainty; after it does, only those still holding long-term contracts will have reason to smile. The memory battle is now about who can sell their 2029 supply first—the ones who’ve sold it won’t fear falling prices; the ones with no inventory will.
#人工智能 #MemoryPrices