Ray Wang on How AI Is Causing DRAM Prices to Surge

Episode Summary

Executive Summary: The episode examines the AI-driven surge in memory chip demand, especially DRAM and HBM, and how constrained wafer capacity, clean-room limits, and technology migration have created a major supply squeeze. Guest Ray Wong argues this is more than a normal cycle: AI has become a new structural demand driver, crowding out consumer electronics while forcing chipmakers to prioritize HBM and server DRAM. The result is rising prices, margin pressure for device makers, and a likely shortage persisting into 2027.

Main Topics: AI as the new demand engine for memory chips (Priority: 5/5): Ray explains that AI training and inference dramatically increase demand for both HBM and conventional DRAM, especially as model context windows, token usage, and agentic workloads grow. Why DRAM is commodity-like and cyclical (Priority: 5/5): The discussion frames DRAM as a commoditized product with limited differentiation and prices driven largely by supply-demand balance, explaining why the industry has historically experienced sharp booms and busts. HBM’s technical complexity and supply crowd-out (Priority: 5/5): HBM uses stacked DRAM dies and requires advanced front-end and back-end manufacturing, making it more profitable but also more wafer-intensive, which diverts capacity away from commodity DRAM. Demand destruction in consumer electronics (Priority: 4/5): Rising memory costs are already affecting PCs, smartphones, cameras, and gaming consoles, causing higher prices, product downgrades, and delayed launches across consumer device makers. Supply-side constraints and chipmaker strategy (Priority: 5/5): Ray outlines limits from clean-room capacity, slow node migration, and the need for more advanced manufacturing, which restricts how quickly supply can respond even as producers invest more. Longer and more structural super cycle than past memory booms (Priority: 4/5): The guests compare this cycle to earlier memory super cycles but argue AI makes it different because the demand shock is prolonged and partially self-reinforcing through hyperscaler and lab competition. Allocation, pricing power, and who gets memory first (Priority: 4/5): In a shortage, top-tier customers and server/HBM buyers get priority, while smaller consumer electronics firms face higher prices and possible inability to source enough chips.

Key Arguments: AI demand is not concentrated in one segment; both HBM and traditional DRAM are needed across training, inference, and server workloads, making memory demand broad-based. HBM is wafer-intensive: on the same wafer basis, it yields far fewer bits than commodity DRAM, so the industry’s shift toward HBM directly reduces supply available for PCs, phones, and consoles. DRAM remains a commodity because standards are shared and products are similar, so pricing power is limited and competition centers on volume and cost per bit. Consumer electronics are already seeing demand destruction, including price hikes, lower sales outlooks, and delayed launches, because memory is becoming too expensive or too scarce. The shortage is likely to continue because capacity expansion takes time, clean-room space is tight, and advanced-node migration cannot be executed instantly. Even if commodity DRAM margins temporarily exceed HBM margins due to spot prices, suppliers still view HBM as a strategic long-term growth area and will keep investing in it. Hyperscalers and AI labs are effectively bidding up memory supply because securing hardware capacity is foundational to competing in AI. The current cycle is longer than historical DRAM cycles because AI demand is still rising while supply remains constrained, with shortages likely extending into 2027.

Data Points: PC memory amount in early consumer PCs: less than 10 MB - The hosts recall how little memory early PCs had in the mid-1990s compared with today. Typical modern PC memory: 16 GB - Used as a comparison to show how much memory needs have grown over time. HBM vs commodity DRAM wafer output: 3x more bits for commodity DRAM vs 1x for HBM - Ray explains the relative wafer efficiency of standard DRAM compared with HBM. HBM future scaling: 8, 12, and eventually 16 DRAM dies stacked - Description of HBM architecture and why it boosts bandwidth and capacity. AI user scale mentioned: 800 million users for ChatGPT - Ray cites this as evidence of massive growth in AI-driven memory demand. Chinese smartphone outlook cut: 12% to 15% reduction in 2026 outlook - Ray references a media/mobile forecast cut attributed to memory constraints and market weakness. Chinese market share of global DRAM demand: about 25% - Used to explain where Chinese memory suppliers compete most directly. Chinese supplier revenue concentration: 90% to 95% from China and Hong Kong - Describing a Chinese memory maker’s revenue base and domestic focus. Memory cycle duration historically: 15 months typical, 18 months longest - Ray contrasts past DRAM cycles with the current AI-driven cycle. Current cycle duration estimate: about 2.5 to 3 years so far - Measured from AI demand acceleration beginning around late 2023 to the podcast time. Potential cycle horizon: through second half of 2027 - Ray’s view on when shortages may still persist. Impact on mobile demand: 10% to 15% less demand in a referenced forecast - A cited example of demand destruction in smartphones due to memory costs.

Pivotal Quotes: "I think this is the key difference that we are seeing for this, right?" — Ray Wong: He explains why the current AI-driven memory cycle differs from past cycles: HBM demand is also crowding out commodity DRAM supply. "Because if you couldn't secure volume, you couldn't even make any money" — Ray Wong: On why pricing matters less than securing chips in the current shortage for consumer electronics makers. "It's not only constrained demand, this thing is also constrained to supply" — Ray Wong: Summarizing why he sees the AI memory boom as a true super cycle rather than a normal upswing.

Implications: Expect continued memory shortages, higher device prices, and investment shifts toward HBM/server DRAM. The AI buildout is becoming a real-world supply shock that may reshape consumer electronics and chipmaker strategy through at least 2027.