DRAM Now Worth More Than Gold as AI Feast Turns Into Global Shortage

The Memory Surcharge

Korea shipped DRAM at an average export price of $92,183 per kilogram in mid-August. That figure — confirmed by data from the Korean Trade Statistics Promotion Institute — is roughly equal to the value of 620 grams of 24-karat gold. Prices have climbed 401% year over year and sit 12.5 times above the level recorded in January 2023, when artificial intelligence was still a novelty rather than an industrial obligation.

This is not a normal cycle. Memory markets have historically rolled over sharply after expansion spurts. The last downturn, beginning in 2022, crushed chipmakers into losses so deep that Samsung posted its first annual loss in two decades and SK hynix nearly bled dry. Buyers laughed — until they needed GPUs to be useful.

AI training workloads turned DRAM from a commodity into an infrastructure input with no substitute. HBM — high-bandwidth memory — became the chokepoint that made or broke every data-center build. What started as a specialized shortage in advanced packaging cascaded downward: standard DRAM lines were cannibalized to feed HBM yield, capacity allocations tightened, and every buyer downstream discovered that memory scarcity does not respect product categories.

Who Pays, Who Profits

The winners are obvious and already rewarding their workers. SK hynix and Samsung have both paid out generous employee bonuses as margins recovered from near-zero to levels unseen since before the 2022 crash. Reports indicate Micron’s Taiwan-based workforce is pressing for similar payouts. In an industry where headcounts are trimmed ruthlessly during downcycles, bonus culture returning this aggressively signals management confidence that the upcycle has structural legs rather than being a temporary price spike.

The losers are anyone whose product contains memory and cannot fully pass costs through. Apple raised retail prices this year and is simultaneously lobbying the U.S. government to allow procurement from China’s CXMT — a move that would bypass Korean and American suppliers entirely. The significance of that pivot is easy to miss: Apple is treating Chinese memory as a strategic hedge against supplier concentration, a signal that even companies with enormous purchasing power are losing leverage over memory access.

MSI has responded by mixing DDR4 and DDR5 chips in its Katana 15 HX C14W gaming laptop, a pragmatic compromise that acknowledges DDR4 inventory remains available at more reasonable prices while DDR5 commands a premium. This kind of hybrid procurement will become more common, not less.

NVIDIA may face the sharpest margin pressure. Servers built around its upcoming Rubin GPUs could see prices rise 15% to 17% purely from memory costs, according to recent reports. For a company whose valuation already prices in relentless execution, input-cost escalations that hit the bill of materials this directly force a choice: absorb the hit, raise prices, or redesign around different memory configurations. None of those options are free.

The Real Bottleneck Is Not Capacity — It Is Time

The three-firm oligopoly governing global DRAM — SK hynix, Samsung, and Micron — controls the vast majority of production. Micron is currently expanding capacity in the United States, a move driven as much by geopolitical risk mitigation as by demand forecasting. The CHIPS Act and allied incentive structures make American fabrication economically attractive even if the geography is suboptimal for proximity to the assembly ecosystems in East Asia.

But new fabs do not produce revenue on day one. Construction timelines for advanced memory facilities run two to four years minimum, and yield ramp adds additional friction. The current price explosion reflects not just strong demand but constrained supply that cannot respond quickly enough. Even if all three producers committed every dollar of free cash flow to expansion today, the supply curve would remain inelastic through 2027.

This timing mismatch is the essential story. AI data-center construction is accelerating now. Memory supply will not catch up for at least 18 to 24 months. The 401% price increase is the market communicating exactly that gap in real time.

Second-Order Effects Are Already Erupting

The memory crunch is no longer confined to server rooms. Automotive manufacturers buying DRAM for infotainment and driver-assistance systems are encountering the same allocation pressure as cloud providers. Consumer electronics face component redesign costs that compound across product cycles. Even energy-intensive industries that rely on memory for control systems and monitoring are feeling indirect pressure as data-center expansion consumes electricity that might otherwise serve factories.

Gold’s price has also climbed in 2026, weakened-dollar driven and amplified by energy-cost inflation stemming from the conflict in Iran. The comparison between DRAM and gold prices is therefore doubly striking: memory is appreciating both against its own history and against a traditional store of value that has benefited from the same macro uncertainties making infrastructure investment more expensive worldwide.

CXMT’s record-setting IPO in China underscores the geopolitical dimension. A Chinese memory producer with meaningful capacity would change the oligopoly calculus. The U.S. lobbying effort on Apple’s behalf signals that Washington recognizes the strategic risk of total dependence on two Korean firms, even as it simultaneously restricts technology transfer that could accelerate CXMT’s capabilities. That contradiction will define policy debates for years.

What Comes Next

Expect memory prices to remain elevated through at least the end of 2027. New capacity coming online from Micron’s U.S. expansion and potential Korean builds will eventually ease physical shortages, but the price floor has shifted upward. The post-2022 crash taught every producer a brutal lesson about the cost of underinvestment, and capital discipline in the memory sector will likely persist even as demand softens in individual quarters.

Buyers who can secure long-term supply agreements at locked-in prices will gain advantage over those negotiating spot-market access. Companies already holding inventory will enjoy cost arbitrage. Those designing new products around memory-intensive architectures should budget for input costs that are materially higher than 2023 baselines.

The 401% surge is not merely a pricing anomaly. It is the market pricing in the reality that AI infrastructure requires memory at a scale the industry did not plan for, and that recovering from the last downturn took longer than anyone expected. Gold may be the metaphor, but the substance is simpler: the chips inside your devices, your car, and every data center running a generative-AI model now carry a scarcity premium that is reshaping cost structures across the entire technology economy.