Jensen Huang’s dismissal of fears about a memory glut in 2029 rests on a straightforward observation: “You can always use more.” For an industry accustomed to treating memory as a cyclical commodity, that might sound almost too simple. But the Nvidia chief’s argument deserves serious consideration because artificial intelligence is changing what computing systems need memory to do. Memory is becoming a more significant constraint on the scale, responsiveness and usefulness of AI. Forecasts of a future surplus therefore have to account for a moving target: applications that keep finding productive uses for greater capacity.
The strongest case supporting Huang is that AI demand extends well beyond training ever larger models. Running those models creates its own expanding requirements. Each additional user, longer conversation and more demanding task adds information that systems must process or retain. Applications that work with documents, images, audio and video broaden that requirement further. An assistant answering a short question represents one level of demand. An assistant examining a company’s records, comparing drawings, reviewing hours of footage or maintaining context across a lengthy assignment represents another. As applications become more useful, the amount of information users expect them to handle grows. There is no obvious reason to assume today’s workloads mark the practical ceiling.
Memory also influences how effectively expensive processors can work. Computing capacity delivers limited value when a system cannot supply data quickly enough or accommodate the workload users want to run. That gives customers an economic reason to invest in memory alongside processing power. The relevant calculation is the cost of completing useful work across the entire system. Additional memory can be worthwhile when it helps improve utilization, supports a more capable application or avoids repeatedly moving and retrieving information. Seen from that perspective, Huang’s confidence reflects a practical engineering consideration: customers have reasons to want more memory that extend beyond simply buying a larger specification.
The familiar assumption that efficiency will eventually suppress demand is also open to challenge. More efficient AI makes existing workloads cheaper, but it can also make previously uneconomic applications affordable. A business might begin with occasional document summaries and expand into continuous analysis once the cost becomes manageable. A developer might move from a short exchange with a coding assistant to asking it to examine an entire project. Lower costs can widen the market and increase usage per customer at the same time. Efficiency matters, of course, but treating every improvement as a permanent reduction in aggregate memory demand overlooks what customers may choose to do with the savings.
Another reason to take Huang seriously is his position in the industry. Nvidia operates at the intersection of AI hardware, software and large infrastructure purchasing decisions. That does not give its chief executive certainty about 2029, but it gives his assessment a useful vantage point. His confidence deserves to be weighed against forecasts that may extrapolate future supply more readily than future applications. Announced manufacturing capacity is relatively easy to put into a spreadsheet. Demand from services that are still being developed is much harder to measure. The danger is that the forecast captures the factories while underestimating what their output will enable.
There is also a plausible path by which greater availability supports demand rather than simply depressing prices. Customers facing expensive or constrained memory have to limit configurations, postpone deployments or narrow the tasks their systems can perform. More supply can loosen those restrictions. Better system economics can bring additional buyers into the market and allow existing buyers to expand. That does not mean every product will remain scarce or every supplier will enjoy uninterrupted pricing power. It does mean that a successful expansion of production should not automatically be interpreted as the beginning of a bust. Some of that expansion may be necessary to serve demand that current constraints are holding back.
The case supporting Huang is ultimately a case for a larger and more varied computing market. AI systems are being asked to remember more, work across more information and serve more users. Those requirements give memory a growing role in the value customers receive from their infrastructure. A confident prediction of a 2029 glut assumes manufacturers will eventually outrun that appetite. Huang is questioning the assumption, and with good reason. The industry may discover that the additional capacity now worrying investors is precisely what the next generation of applications needs.