Global RAM prices have reached record levels as surging AI training and inference workloads consume a disproportionate share of DDR5 manufacturing capacity.
A 32GB DDR5 memory kit that sold for around $72 in mid-2025 now costs $392 as of August 2026. Same module, same capacity, and the price has multiplied more than fivefold in twelve months. Meanwhile, on one retailer's product pages, automated software checks stock every 6.5 seconds to snatch up every batch the moment it lands. This isn't a market glitch or a passing bubble. It's the moment the AI race stopped being just a data center story and started emptying the shelves of the RAM that ends up in everyone's computer.
The numbers describe a change in scale, not a seasonal adjustment. According to tracking by Tom's Hardware, a 32GB DDR5 kit today starts at roughly $375 at the low end, while high-frequency or RGB-lit versions climb well beyond that. A year earlier, the same capacity cost $80 to $120. Anyone thinking in euros finds the same story. A Kingston Fury Beast 32GB kit cost 114 euros in September 2025 and now runs 523 euros in August 2026, a 360 percent jump.
The pressure isn't coming from retail alone. Samsung raised the list price of its 32GB DDR5 modules from $149 to $239, a 60 percent increase set upstream, before the module ever reaches a shelf. On larger capacities the climb is even steeper. A 64GB kit has hit $1,118, nearly six times what it cost twelve months ago. Overall, depending on capacity and speed, DDR5 list prices have moved between 355 and 933 percent in a year, wiping out two decades in which memory had been the one component that kept getting cheaper. The number to remember: $392 for 32GB works out to almost $12 per gigabyte, a price ordinary buyers haven't seen since the early DDR5 years. Building a mid-range PC today often means spending more on RAM than on the processor.
To understand the price surge, you have to look at what's happening inside the fabs. Samsung, SK Hynix, and Micron alone produce more than 95 percent of the world's DRAM, and all three companies have shifted their lines toward HBM (High Bandwidth Memory, the ultra-fast memory stacked on top of or next to the accelerator chips used to train AI models). HBM pays off far better. Revenue per wafer is estimated at three to five times that of traditional DDR5, so for a manufacturer the economic choice is obvious.
But there's a second, subtler effect that explains why the shortage is so severe. Making HBM consumes an enormous amount of silicon. A single gigabyte of HBM takes up roughly the wafer area of four gigabytes of standard DRAM. Every accelerator that leaves the fab therefore pulls capacity away from a much larger quantity of consumer memory than the raw chip count would suggest. Industry estimates put AI-related demand at nearly 20 percent of the planet's entire DRAM wafer capacity in 2026, with data centers set to consume roughly 70 percent of global memory chip production.
The picture has tightened further with long-term supply deals. In July 2026, Nvidia signed a multiyear supply agreement with SK Hynix worth up to $500 billion, locking in future production capacity long before the consumer market could even compete for it. SK Hynix has said its HBM, DRAM, and NAND capacity is "essentially sold out" through all of 2026. Micron, for its part, has chosen to exit the consumer memory market entirely to focus on enterprise and AI customers. Anyone buying a stick of RAM today is, in practice, competing with hyperscalers that have already booked out the fabs for years. This isn't an isolated case. We saw the same mechanism emerge with the memory crunch that hit iPhone 18 pricing, where Apple found itself paying for the same demand that's now emptying RAM shelves.
On top of the structural shortage, a second siege has piled on, this one digital. Security firm DataDome has documented an automated scalping operation targeting memory pages at several retailers. A single bot checks stock every 6.5 seconds, averaging more than 550 requests per page and over 50,000 requests per hour on the sites it targets. On one of these product pages, malicious bots account for 91 percent of traffic, roughly ten automated hits for every visit from an actual person. In total, the company says it has blocked more than 10 million requests from this one bot alone.
The mechanism is the same one already seen with graphics cards and game consoles, just applied to an even more liquid good. Whoever runs the bots wants real-time price and stock data to buy up batches the instant they appear, then resell them at a markup. The evasion techniques are sophisticated. The bots append random parameters to every request (a practice called cache busting) so each one looks like a fresh visitor and forces the server to serve live data instead of a cached copy, and they learn a site's rate limits to stay just under the threshold that would trigger a block.
"We've observed malicious actors discussing using AI to reverse-engineer anti-bot protections or to automate scripting tasks," said Jérôme Segura, VP of Threat Research at DataDome.
It's this double squeeze that makes this shortage different from those of the past. On one end, AI drains production upstream. On the other, automated tools, some of them AI-driven in turn, sweep up whatever's left downstream. The result, for anyone trying to buy a couple of memory sticks, is an empty shelf even when the stock technically existed.
Not everyone is losing out on this surge. For the three big manufacturers, it's a golden moment. Samsung has warned that the shortage will keep pushing prices up through all of 2026, and HBM margins are so much richer that every wafer diverted away from DDR5 is, for them, a profit-driven choice. An HBM3E module sells for around $60 to $100, against $5 to $10 for an equivalent amount of conventional DDR5, and HBM3E list prices are already up roughly 20 percent year over year.
Inside this scenario, an unexpected opening is emerging, and it's coming from China. CXMT (ChangXin Memory Technologies), China's leading memory maker, has pushed its DDR5 to 8,000 MT/s and passed 90 percent production yield, a figure comparable to the three giants. The company already supplies Lenovo and is aiming for mass production of a 96GB module in 2026. On its own, CXMT held just 4.9 percent of the global DRAM market in 2024, against a combined 93.4 percent for Samsung, SK Hynix, and Micron. That's far too little to flip the market overnight, but it's the first time a serious fourth player has tried to fill the void the three giants left behind while chasing AI. For PC builders, an alternative source of affordable DDR5, even under lesser-known brand names, could become a lifeline over the next two years.
The practical question a lot of people are asking is simple. Should you buy now or wait? The honest answer is that there's no quick return to the old prices. Micron warns that the supply shortage will last through 2028, and meaningful new production capacity isn't expected before late 2027. So if you need the memory to get work done, there's not much point waiting for a price drop that the manufacturers themselves aren't predicting. Anyone who can hold off has a few sensible moves available. Buy only the capacity you actually need, without chasing the huge capacities that are taking the biggest hit right now. Look for previous-generation modules on the secondhand market, where the pressure is less intense, the same way buyers already do when hunting for used GPUs to run AI models locally. And for businesses, factor in that memory costs now weigh on IT budgets in a structural way, a line item worth weighing with the same care used to choose between capital expenditure and operating expenditure. The same goes for individual consumers planning upgrades.