The AI Boom Has Locked Up the World's Memory Chips

SK Hynix is spending to double its memory production capacity over the next five years. Its chairman, Chey Tae-won, warns the shortage will persist until 2030 anyway.

Until 2030 there’s still some shortage. — Chey Tae-won

The reason is not that demand is growing faster than anyone can build. It is that the new output is already sold. What looks from the outside like a chip shortage is a supply-lock. Prices are spiking, lead times are stretching, consumer devices are getting more expensive. But the mechanism underneath is different. The three companies that control roughly 90% of global DRAM production are signing binding contracts with AI hyperscalers that physically remove memory from the spot market consumer electronics makers have always relied on. The fabs are expanding. The shelves are bare before the concrete is poured. The mechanism is now quantified. Micron has signed 16 take-or-pay Strategic Customer Agreements worth a minimum of $100 billion over three to five years ^[1]. In a single month, its remaining performance obligations jumped from $5 billion to $100 billion. Sandisk signed five multi-year agreements with $42 billion in minimum revenue commitments. These are not purchase orders a buyer can walk away from; they are binding obligations that commit physical output to specific customers years in advance.

provide us greater confidence in our capex (capital expenditures) and R&D (research and development) investments. — Micron Technology

On December 3, 2025, Micron exited the consumer memory market entirely to prioritize AI hardware production ^[2]. A major DRAM and NAND maker simply withdrew from the aisle consumer electronics shops in. The same pattern is unfolding in capacitors. Samsung Electro-Mechanics is signing binding long-term contracts with AI firms for MLCCs, the tiny capacitors that regulate power on every circuit board ^[3].

We are securing forward volume through binding mid- to long-term contracts. — Lee Tae-gon

Nvidia's Vera Rubin platform requires $4,320 in MLCC content per rack, a 182% increase over the prior generation, and Goldman Sachs estimates MLCC demand will quadruple by 2030 while production capacity grows roughly 10% annually. The pattern repeats across the bill of materials: memory, capacitors, and the substrates and connectors that sit between them. The price hikes now cascade through every tier of electronics. Raspberry Pi raised prices three times in six months: the 16GB Pi 5 went from $120 to $299, and CEO Eben Upton cited a sevenfold increase in LPDDR4 RAM costs ^[4]. Framework's 128GB desktop model jumped from $1,999 to $2,459 ^[5]. Its CEO was blunt about the trajectory.

The memory outlook as we enter 2026 continues to get worse. — Nirav Patel

Apple raised prices 15 to 33% across its product line, with the Mac Studio jumping $1,300. Tim Cook described the situation in terms no consumer electronics executive has used before.

Price increases are unavoidable. — Tim Cook

Microsoft reported that console memory costs have increased more than 2.5 times and expects another doubling by fall 2027 ^[6]. Intel and AMD raised CPU prices 10 to 15%, with PC lead times stretching from two weeks to as long as 26 weeks ^[7]. Samsung raised US prices across Galaxy phones and tablets in April, with tablets up $50 to $280, and offered no official explanation ^[8]. The litany is not a cyclical spike. Gartner now forecasts the steepest contraction in device shipments in over a decade: PCs down 10.4% and smartphones down 8.4% in 2026, driven by a 130% surge in DRAM and SSD prices ^[9]. The sub-$500 PC segment will disappear entirely by 2028.

Higher prices will narrow the range of devices available, prompting buyers to hold on to devices for longer, fundamentally altering upgrade cycles. — Ranjit Atwal

The counter-evidence is real. Tariffs are a co-driver of the price increases: HP's CFO explicitly identified both AI memory demand and the Trump administration's 10 to 15% import tariffs when forecasting a double-digit decline in PC shipments ^[10]. Investors remain skeptical that the memory market has broken its cyclical pattern: Micron and Sandisk shares fell 24% and 38% respectively in July despite the $100 billion in take-or-pay contracts ^[1]. Open-weights AI models from China, including Z.ai's GLM-5.2 and DeepSeek-V4, now provide frontier-class capabilities at a fraction of the cost of American closed models, raising the possibility that AI compute demand could eventually moderate from the software side ^[11]. None of this undoes the supply-lock mechanism. Tariffs amplify the price hikes but do not cause the physical removal of memory from the spot market. Investor skepticism about cyclicality does not void contracts that commit output years in advance. And even if open-weights models reduce per-query compute costs, the hyperscalers are not slowing down. Meta's 2026 AI data center capital expenditure is $125 to $145 billion, increased by $10 billion due to DRAM inflation alone, and the company has taken a $90 billion position in AMD as an economic hedge against memory price inflation ^[12]. Nvidia projects global data center capex growing from $600 billion in 2025 to $3 to $4 trillion by 2030 ^[13]. The supply-lock resolves on one condition, and it is not the arrival of new fabs. South Korea is backing a $575 billion initiative for Samsung and SK Hynix to build four new fabrication plants, with combined spending potentially reaching $1.3 trillion over the next decade ^[1]. But the output of those fabs is being pre-sold before ground is broken. The shortage eases only when consumer electronics makers can pay AI-grade prices. The AI buildout has installed a permanent inflationary floor under the cost of every electronic device, and the contracts have already locked it in.