Conference Presentation, Panel
AI Has a Memory Problem | Gelsinger, SK hynix & More | RAISE Summit 2026
- AI demand is described as insatiable with indefinite growth potential, driven by expanding context windows and model sizes that necessitate increasingly larger and faster memory solutions to achieve low first-token latency.
- HBM remains the dominant technology for satisfying near-term AI demands and is projected to stay the best option for processing AI for several more years, despite supply limitations that will persist irrespective of capital investment due to three-to-four-year innovation-to-volume cycles.
- New stacked memory architectures from entities like D-Matrix and Next Silicon are expected to emerge before the end of the decade, offering bandwidth improvements of 10 to 50 times over current shoreline-limited HBM, though these solutions will not materialize until late this decade.
- A fundamental memory shortage is expected to continue for several years as demand grows exponentially while supply increases only linearly, creating a scenario where years of memory plenty remain distant and shortages will likely persist before the supply-demand dynamic shifts.
- Production capacity is constrained by a four-to-one trade-off favoring HBM, which reduces the availability of PC, server, and handheld memory, while manufacturing yields face compounding losses from stacking multiple layers.
- The industry has transitioned from commodity pricing to contract-based long-term agreements of five years or more, reflecting a view of memory as critical AI infrastructure investment where the focus has shifted from cost per bit to bandwidth per watt.
- Short-term capacity additions are expected to be modest over the next couple of years, reflecting capital decisions made three years ago, while bringing a major factory online takes about four years with profitability likely not arriving until year six.
- Market deficits are currently running at 7% for HBM and similar levels for DRAM, with price increases likely to continue and potentially impair company top lines if memory intensity per token falls dramatically due to the Jevons paradox.
- Risks include a bullwhip effect where excessive CapEx based on optimistic projections could lead to a bear case, alongside potential demand restrictions in price-sensitive smartphone and PC markets if prices become too high, causing consumers to postpone purchases.
- External shocks such as data center construction unrest in the US or geopolitical issues in Taiwan pose significant supply chain risks, with Taiwan identified as lacking sufficient energy supply to prevent industry brownouts if blocked for three weeks.
- Building resilient supply chains is viewed as a task requiring government industrial policy rather than venture capital, and there is uncertainty regarding whether demand durability will hold for the very long, expensive capital cycles associated with new factory construction.
- Long-term forecasts are difficult to make for the current tight market, which is characterized by a 7% deficit and the reality that HBM does not solve the inherent "memory wall" problem of AI, even as production shifts toward more thermally and bitwise efficient architectures in the future.