Korva is a scroll of the internet’s best ideas.
Breaking a memory contract can sink you
Why memory dominates
Memory makers are trading short-term upside for long-term agreements, where the customer prepays and the price has a floor and a ceiling. Memory is the axis that dominates all others, because the more memory paired with each unit of compute, the more tokens come out. That is also why demand has not responded negatively to the lower costs.
The cost of breaking a deal
Only four buyers matter at scale: Amazon, Google, AMD, and Nvidia, which is much bigger than the rest combined. Picture 2028, with oversupply and falling prices, and Google breaking its agreements to get a lower price. Memory makers cut capacity in a glut, and oversupply is followed by undersupply. Google's allocation would be the first thing they remember.
Apple used to be able to do whatever it wanted, because its volume was so big that even if it squeezed Hynix, Micron would take the business. With four large buyers and a crowd of startups, a memory maker can answer a broken price agreement by breaking the volume agreement and giving that volume to a competitor. Market share for the next several years will be determined by supply chain allocations and what each company has pre-purchased.
Why do memory makers want prepayment?
A modern memory fab costs billions of dollars and takes years to build. Prepayment lets the maker finance construction without taking all the risk of a price collapse. The memory industry has a long history of boom and bust, with prices swinging sharply as supply catches up with demand. Over the decades the field consolidated from dozens of producers in the 1980s to three that make the large majority of the world's DRAM: Samsung, SK Hynix, and Micron.
Why does oversupply follow a shortage?
When prices spike, every maker adds capacity at the same time. A new fab takes about two years to build, so the added supply arrives together, long after the shortage that triggered it. Prices then fall, makers cut investment, and the cuts set up the next shortage. This repeating pattern is why DRAM is described as a cyclical industry.
Why does memory limit how many tokens come out?
Generating each token requires reading the model's weights and its working memory, called the key-value cache, from memory. For most inference work, the speed of that reading is the bottleneck rather than the arithmetic. This is why high-bandwidth memory is stacked in layers and mounted right beside the processor. A chip with more memory bandwidth per unit of compute can serve more tokens each second.
More from Invest Like the Best with Patrick O'Shaughnessy
