AI is making your RAM expensive — and relief isn't coming until 2028
If you've been putting off a PC upgrade and noticed RAM prices look nothing like they did two years ago, that's not your imagination. DDR5 memory has surged 3–4× since September 2025 lows, according to PC Gamer, with a Crucial Pro 32 GB kit jumping from roughly $80–100 to $364 in the space of a year. DDR4 has tracked the same curve. Motherboard sales dropped 40–50% year-over-year as a result — people simply aren't building PCs when memory alone costs $400–750.
The HBM squeeze
The culprit is a specific type of memory called HBM (High Bandwidth Memory), which AI accelerators like Nvidia's GPUs require in enormous quantities. Samsung, SK Hynix, and Micron — three companies that control more than 95% of global DRAM production — have steadily reallocated wafer capacity away from consumer DDR5 and toward the far more profitable HBM. OpenAI's Stargate project alone signed deals covering 900,000 wafers per month from Samsung and SK Hynix, soaking up roughly 40% of global DRAM output. When fab lines are finite, something has to give — and that something is the memory stick you want for your desktop.
On a historical basis, computer memory has been falling at an exponential rate for decades. But we just undid about 20 years of progress.
— Daniel Lemire (@lemire) August 5, 2026
RAM on a per unit basis is about as expensive as it was in 2007.
To my knowledge, it is an historical anomaly. I cannot recall a similar… pic.twitter.com/UH6LgZ1fc4
The ripple effects go beyond PC builders. Nvidia itself is now testing stripped-down versions of its next-generation Rubin Ultra chip with just 192–256 GB of HBM, compared to the 1 TB originally planned, because supply constraints make the full configuration impossible to produce at scale. CNBC reported that Dell's COO acknowledged the company is already changing product configurations to offset memory costs — changes that will "likely affect retail prices for devices."
No quick fix
The structural problem is that building new fabs takes years, not months. Micron announced a $9.3 billion facility in July 2026, but it won't produce a single chip until Q3 2028. Analysts at TrendForce don't forecast meaningful price relief until late 2027 at the earliest, and SK Hynix has warned the HBM shortage could extend past 2030 as AI demand continues to outpace new capacity.
This breaks a 15-year assumption that memory would always get cheaper. Developer Daniel Lemire's price analysis shows that the cost per gigabyte of DRAM has effectively stagnated for over a decade — a 64 GB server module today costs roughly what equivalent capacity cost in the late 2000s, adjusted for purchasing power. SSD storage keeps falling; RAM does not.
What it means now
For anyone shopping now, there is no good time to wait for a deal that isn't coming soon. If you need more memory for a build or upgrade, budget for today's prices. Software developers face a parallel problem: the era of throwing RAM at poorly optimized code is over for the foreseeable future.
