Samsung: memory shortage runs to 2028. You're last in line.

Two earnings calls this week told the same story from opposite ends of the supply chain. Samsung, which makes the memory, posted a record quarter and warned the shortage gets worse before it gets better. Apple, which buys the memory, nearly doubled its inventory because it doesn't trust it can get parts later. The memory shortage is no longer a price story you can absorb with a bigger budget line — it's an allocation story, and small buyers sit at the back of the queue.

What actually happened

On its Q2 2026 call, Samsung reported record operating profit of roughly 89 trillion won — about $62 billion — on surging HBM4 and server DRAM demand. The forward guidance is the part worth reading twice: per KED Global, Samsung expects the global memory shortage to deepen in 2027 and persist through 2028, driven by hyperscaler AI spending that keeps demand above available supply. The company said multiyear contracts could cover up to 70% of its planned capacity.

That last number is the one nobody's talking about. Seventy percent committed means the remaining sliver clears at spot, and you're bidding against everyone else who didn't sign a five-year agreement with a memory fab.

Apple, arguably the best supply chain operator on earth, is behaving accordingly. On Apple's fiscal Q3 call — $109.4 billion in revenue, up 16% — Tim Cook described a "hundred-year flood" in memory pricing and said the impact from supply constraints would increase significantly sequentially. TechCrunch reports Apple's inventory sits at $11.1 billion, up from $5.7 billion the prior September. Apple already raised Mac and iPad prices in June.

Why the memory shortage matters for your business

We've written before about DRAM prices as a budget item. This is a different problem: availability with a date attached. When the fabs commit most of their output to contracted customers through 2028, the failure mode for a five-person shop isn't paying 30% more — it's a reseller quoting twelve weeks and delivering in twenty.

Four things to do now. Pull hardware refreshes forward if they land inside the next four quarters, and spec memory high on those orders; you are not going to buy that upgrade cheaper later. Stop putting hardware on the critical path of any project. If a build only ships when a box arrives, it doesn't have a launch date, it has a hope. Re-run the self-hosting math with lead time in it — hosted inference costs more per token and needs zero purchase orders, which is worth real money when the alternative slips a quarter. And write lead-time assumptions into vendor quotes so a slipped delivery is their problem, not a surprise in your schedule.

The uncomfortable read: AI capex isn't just bidding up your token bill. It's quietly consuming the components underneath every laptop, server, and phone you'll buy through 2028.

Sources: KED Global, TechCrunch, Apple Newsroom.