Memory Went From a Line Item to a Rationed Resource, and That Breaks a Founding Assumption

What Happened

Apple raised MacBook and iPad prices with no new product to show for it. The reason was memory. A 512GB MacBook Air that cost $1,099 two weeks earlier now runs $1,299. The 1TB MacBook Pro jumped from $1,699 to $1,999. The entry MacBook Neo went from $599 to $699, and the 128GB iPad Air from $599 to $749. Apple said it had “never seen a component price increase this much, this quickly.” On the day of the announcement, the stock fell about 6%, erasing roughly $275 billion in market value.

The driver is AI data centers. They now consume an estimated 70% of the memory produced worldwide. Samsung, SK Hynix, and Micron, which together make more than 95% of global DRAM, have redirected large blocks of capacity toward the high-bandwidth memory (HBM) that sits on Nvidia GPUs. That left the ordinary DRAM and NAND that goes into laptops and phones critically short. Spot DRAM prices rose roughly 98% in the first quarter of 2026, with some forecasts calling for another 58% to 63% this quarter. Micron booked $41.46 billion in fiscal Q3 revenue, up about 345% year over year, and crossed a $1 trillion market cap in May. The suppliers are now outrunning the companies that build the finished devices.

What It Means for Founders

The real shift is not that prices went up. It is that the whole way you treat memory has changed. For twenty years founders built on a free assumption: silicon gets cheaper over time. Hardware bills of materials and cloud inference prices were both penciled in on top of that falling curve. The curve just bent. Memory is no longer a cost that drops each year; it is a rationed input, allocated to hyperscalers first, with smaller teams buying whatever is left at a premium.

The change hits two kinds of founders in opposite ways. Hardware-dependent teams, think robotics, edge devices, on-prem AI boxes, watch their entire BOM wobble. If memory doubles, the question is not margin but whether the product survives. SaaS teams that priced their plans on the belief that inference keeps getting cheaper are exposed the other way, as their unit cost drifts up instead of down. At the same time this rationed economy opens new markets. Anything that helps customers use less memory grows in value: efficiency tooling, small and quantized models, on-device inference, and refurbished or secondhand hardware channels. Scarcity always puts a price on thrift and reuse.

What You Can Do Now

If you sell hardware, rerun your BOM under a memory-doubles scenario and confirm the business still works there. The risk is not just price but securing volume at all, so locking supply contracts early beats buying on the spot market later. If you build AI software, measure your real per-request memory and inference cost today, then stress-test whether your pricing survives that cost rising. The earliest credible relief is late 2027, and some projections push it into 2028. Rather than wait out the cycle, the team that first designs a business that pencils out on expensive memory takes the next round.