In full
For most of the last decade, memory was something engineers stopped thinking about. It got cheaper every year, so the rational move was to use more of it: cache aggressively, keep data sets in RAM, over-provision the instance and move on. That assumption quietly broke in 2026. Memory is now the part of the supply chain everyone is talking about, because the AI build-out has turned it into the scarce ingredient.
Three companies, one redirected factory
Almost all of the world's memory comes from three manufacturers: Samsung, SK Hynix and Micron. That concentration normally goes unnoticed. It matters now because all three have pointed the same way at once. The memory that AI accelerators depend on, high-bandwidth memory or HBM, is far more profitable than the ordinary DRAM in a laptop, and demand for it is enormous. So the factories have been retooled towards it. By industry estimates HBM is taking roughly a quarter of total DRAM wafer output in 2026, up from under a fifth the year before, and much of that HBM is already sold out under multi-year deals with the large cloud providers.
Nothing was taken away from the consumer market. The capacity was simply pointed at a more profitable customer, and the rest of us are buying what is left.
The knock-on is mechanical. Capacity diverted to HBM is capacity not making conventional DRAM and NAND flash storage. Supply tightens, and price follows. Through 2026 the forecasts for ordinary memory and storage have been described by analysts as the steepest in living memory, with conventional DRAM contract prices rising by tens of percent quarter on quarter and HBM itself repriced upward for the year. The cause is not a shortage of sand or a broken plant. It is allocation: the most valuable buyer gets served first.
Why it lands on your bill
Memory is not a niche component. It is inside every phone and laptop, and inside every server your application runs on. When the raw part gets more expensive, the cost moves up the chain: into the price of devices, into the price of the servers a cloud provider buys, and eventually into the per-hour cost of the instances you rent. In South Africa that effect arrives twice, because the hardware is imported and priced in dollars, so a rand that moves against the dollar widens the gap again before anything reaches a local invoice.
Local providers have already said so out loud. In March 2026, South African cloud and hosting firms warned publicly that sustained price increases were coming through the year, citing hardware costs rising anywhere from single digits to thirty percent or more, the rising cost of building data centres, electricity and backup power, and the same memory squeeze driving it all. The warning was specific: the effect would reach ordinary customers in the second half of 2026.
What changes for how you build
None of this is a reason to panic, and most of it is outside your control. What is inside your control is the set of assumptions baked into your architecture. For years it was cheaper to throw memory at a problem than to engineer around it. That trade has shifted, and a few habits are worth revisiting.
- Right-size, don't reflexively over-provision. The instance sized "with headroom to be safe" is now a standing cost. Measure what you actually use.
- Mind the data footprint. Caching everything in memory, keeping full data sets resident, and duplicating state across services all cost more than they did. Cache what earns its place.
- Expect cloud prices to drift up. Build the assumption of rising per-unit infrastructure cost into budgets and pricing, rather than being surprised by it mid-year.
- Treat the three-supplier concentration as a real risk. A market served by three factories can move sharply and together. Resilience is partly commercial, not just technical.
The longer point
The memory squeeze is a small, concrete example of a larger pattern: the AI build-out has physical limits, and those limits show up in prices long before they show up in headlines. Power, land, water and memory are all finite, and AI is now the buyer willing to pay most for them. For anyone running software, the practical takeaway is unglamorous. The free lunch on memory is over for a while, and the systems that handle it best will be the ones that knew what they were spending it on.
AI's hunger for high-bandwidth memory has pulled the world's three memory makers towards it, leaving ordinary memory and storage scarce and pushing prices to their sharpest rise in years. It reaches you through dearer devices and a rising cloud bill, and in South Africa the weak rand sharpens the effect. You cannot fix the supply chain, but you can stop treating memory as free. Measure what you use, cache what earns it, and budget for infrastructure costs that are going up rather than down.