Anyone shopping for a new computer or server right now can feel it in the price tag: memory is more expensive than it has been in years. According to a Bloomberg report, Nvidia is raising prices for its AI flagship chips, Vera Rubin and Grace Blackwell, by roughly 15 percent, effective for deliveries starting in early 2027. The cause is not Nvidia itself but one step further up the supply chain, with the three memory makers Samsung, SK Hynix, and Micron, which together control more than 95 percent of global DRAM production.
Key takeaways
- Nvidia is raising prices for AI servers built with Vera Rubin and Grace Blackwell chips by about 15 percent for deliveries starting in 2027.
- The trigger is a structural memory chip shortage: DDR5 server RAM prices have doubled within a year, and in some cases more, according to market data.
- Samsung, SK Hynix, and Micron are shifting manufacturing capacity from conventional DRAM to high bandwidth memory for AI accelerators, which brings three to five times the revenue per wafer.
- HBM’s share of global DRAM wafer output rose from 8 percent in 2024 to 23 percent in 2026.
- Analysts do not expect meaningful relief before late 2027, when new fab capacity in the US and Asia comes online.
How Nvidia’s price hike came about
Nvidia has not publicly commented on the increases, according to the Bloomberg report. The surcharges first hit its contract manufacturers, then get passed down to cloud providers such as Amazon, Microsoft, Google, and Meta, as well as AI labs like OpenAI and Anthropic. These companies are already pouring billions into AI infrastructure, and the higher prices indirectly finance the growing market power of their most important supplier. A single AI server uses eight to ten times as much DRAM as a conventional server and more than three times as much NAND storage, which makes hyperscaler demand the decisive price driver. Companies like Waymo, which is now building entire data centers on wheels with custom AI chips, feel the squeeze twice over: once on compute, and once on the memory that feeds it.
Why HBM is reshaping the memory market
The real reason behind the shortage is a deliberate choice by the three major manufacturers. High bandwidth memory, or HBM, is the type of memory AI accelerators need to handle their massive data loads, and industry estimates put its revenue per wafer at three to five times that of conventional DDR5. Manufacturing is a zero-sum game: every wafer that becomes HBM is one wafer less for ordinary server and consumer RAM. HBM’s share of global DRAM wafer production has climbed from 8 percent in 2024 to 23 percent in 2026, according to market data, and Gartner expects overall DRAM prices to rise roughly 47 percent this year. Samsung itself warned publicly that in 2026 there will be issues around semiconductor supplies that affect everyone, not just Samsung. Concretely, that means a 32-gigabyte DDR5 module that cost around $149 in the spring was going for $239 by September, a 60 percent jump within a few months. The jump looks even steeper at the wholesale level, the contract prices server makers pay directly to the memory fabs: there, per-unit pricing for DDR5 more than doubled, from roughly $7 to about $19.50. Market researcher Gartner expects industrywide DRAM prices to climb roughly 47 percent this year, on top of an already sharp 50 percent increase in 2025. Even older DDR4 modules, long treated as a fading product line, are being pulled up in price, simply because production lines are tight and many older servers still depend on them.
What this means for buyers outside the AI industry
The fallout does not stay confined to data centers. Market observers report that the cost of a server system with 256 gigabytes of DDR5 memory has risen from roughly $800 to $1,200 in 2025 to somewhere between $2,000 and $3,500 now. Small and mid-sized companies face an additional disadvantage, since they are structurally outcompeted for available capacity by hyperscalers with long-term supply contracts. The squeeze is also showing up in the consumer market for PCs, laptops, and graphics cards with dedicated memory, because they draw on the same production lines. Market watchers give buyers planning a major hardware purchase advice that sounds a lot like what data center operators are hearing: buy generous memory capacity upfront rather than counting on cheaper upgrades later, and lock in long-term supply commitments wherever possible.
What this means
What looks like an ordinary supplier price hike is actually a structural shift of the entire memory industry toward AI. The three manufacturers that control the market have little incentive to quickly return to the old capacity split as long as AI customers keep paying multiples for HBM. Analysts expect no meaningful relief before late 2027, when new fabs built specifically for HBM come online in the US and Asia. Until then, memory prices are set to become a quiet but powerful cost driver for practically every technology purchase, from Nvidia’s next chip generation to a new laptop at the electronics store. The bill for the AI boom, it turns out, is not paid only by those training the models, but by anyone buying memory over the next two years.