The DDR5 Crash: How Google’s ‘TurboQuant’ Secretly Tanked Global RAM Prices

If you had visited any major electronics store in early March 2026, you would have noticed that the memory section was remarkably empty. DDR5 kits that were previously priced at $200 were now selling for $490. At their peak, Corsair’s flagship Vengeance 32GB modules cost $500, which a year and a half ago would have seemed absurd. DRAM shortages could last until 2028, according to a public warning from Micron. Even with their factories operating at full capacity, Samsung and SK Hynix were unable to keep up. The narrative was resolved: RAM was being consumed by AI at rates the industry had never predicted, prices were rising, and consumers would just have to accept it. Next, a paper was released by Google.

On March 24, 2026, Google disclosed information about TurboQuant, a KV cache compression algorithm for large language model inference that can reduce AI workloads‘ memory requirements by up to six times while increasing inference speed by eight times.

For anyone outside the machine learning research community, the paper itself is difficult to read due to its complexity and technical nature. However, anyone watching the memory market could immediately understand its implications. The entire demand forecast that had been driving prices upward for eighteen months was suddenly called into question if AI data centers, the main cause of the global DRAM demand surge, could run their models on one-sixth of the memory they currently consume. The market reacted to this information as quickly as it typically does to negative earnings calls and geopolitical shocks.

DDR5 prices fell by more than 30% in China within days of the announcement. In a single day, some retailers experienced drops of more than 100 yuan. Over the course of the next month, RAM prices on Amazon in the US dropped by more than 20%, with Corsair’s flagship 32GB kit going from $490 to $379.99. The reaction was swift and harsh throughout the semiconductor supply chain. In less than a week, the combined market capitalization of Micron, Samsung, SK Hynix, and associated memory suppliers lost about $450 billion. As investors attempted to recalculate where demand was actually going, memory stocks that had been riding the AI demand wave for the majority of 2025 saw a sharp decline.

There was more to the TurboQuant shock than just one. It came at a time when the market was already dealing with another unsettling development: OpenAI’s discreet withdrawal from its most ambitious hardware plans. Early in 2026, there were rumors that Sam Altman had signed letters of intent to buy about 40% of the world’s RAM output for OpenAI’s infrastructure expansion.

The implied demand that had been factored into memory market projections started to fade as OpenAI shifted toward leaner, more efficient model architectures and reduced some of its most ambitious capacity forecasts. These letters were non-binding. AI data centers purchasing less than anticipated on the supply side and customers refusing to pay peak prices at retail on the demand side resulted in what Tom’s Guide called a “pincer maneuver” as a result of TurboQuant and OpenAI’s retreat.

All of this raises the question of whether the price relief is genuine and long-lasting or if it is merely a short-term confidence shock that will pass as the industry adapts. Jevons Paradox, a 19th-century economic observation that efficiency gains tend to increase rather than decrease resource consumption because lower costs make previously unprofitable uses feasible, is being invoked by some seasoned market observers.

If TurboQuant makes it six times less expensive per parameter to run a large model, AI companies should build much larger models and run more workloads concurrently rather than purchasing the same amount of hardware at a lower cost. In the history of semiconductors, the same pattern occurred: as chip efficiency increased, demand grew to match and eventually surpass the savings. The industry’s more seasoned analysts are advising caution before declaring the shortage to be over because memory manufacturers are aware of this.

It’s difficult to ignore the similarities to what transpired in the natural gas market following the introduction of shale fracking technology: the price plummeted, then recovered, and then more as low-cost extraction created previously untapped markets. In one version of the TurboQuant story, the algorithm democratizes access to large AI models for data centers in emerging markets and smaller businesses that previously couldn’t afford the memory costs, ultimately increasing demand rather than decreasing it.

Over a period of three to five years, that version might be the right one. However, in the short term, the perception of abundance—even ambiguous, algorithmically generated abundance—was sufficient to cause the kind of inventory sell-off that causes DDR5 prices to drop by 30% over the course of a weekend. This is the first real window of opportunity since the crisis started for the gamer who has been waiting six months to upgrade their computer. It’s still genuinely unclear if it remains open long enough to matter.