AI threatens to devour the memory of the market and Dell already puts a figure that scares
Artificial intelligence has been raising expectations about GPUs, servers and data centers for some time, but there is one piece that is beginning to emerge with increasingly uncomfortable force: memory. The latest forecast attributed to Michael Dell sums up the problem pretty well.. According to that estimate, the total demand for memory linked to the AI market in 2028 could be 625 times greater than that of 2022. It is a figure so large that it invites immediate skepticism, but it also perfectly portrays the magnitude of the challenge that is forming around the sector’s infrastructure.
The logic of this calculation is based on two simultaneous movements. On the one hand, the memory per accelerator grows radically. In 2022, an H100 was running with 80 GB of HBM3. By 2028, the maximum jump that is being projected for new generation platforms will be figures on the order of terabyte per GPU in certain scenarios, while more immediate configurations for the Rubin family already raise the bar a lot compared to what was at the beginning of the cycle.
On the other hand, not only will there be more memory per chip– There will also be many more drives deployed in data centers. If both factors are multiplied at the same time, the result stops sounding exaggerated and begins to look like a fairly serious industrial threat.
The bottleneck is no longer just the GPU
For months, much of the conversation revolved around the ability to make accelerators. But the real problem is much broader. An AI system does not live off the GPU alone. You need HBM, spare memory, fast storage, interconnection, and entire platforms ready to move huge amounts of data. When that demand accelerates at once on all fronts, memory goes from being a critical component to becoming one of the potential bottlenecks in the market.
This is especially clear in the case of the HBM.. Only a handful of companies are able to manufacture it at the level needed by the current market, and even then, supply remains limited. The consequence is obvious: more pressure on prices, tighter supply commitments and a growing battle to secure capacity before the competitor. In such a context, the issue is not just how much an AI server costs, but how much memory each client can actually get and when.
Climbing goes far beyond model training
We often think of this infrastructure rush only from the training of large models, but the memory pressure does not stop there. It also affects inference, entire systems deployed by hyperscalers, and architectures where the amount of data in motion is increasing. The more memory needed per accelerator and the more machines go into production, the more stress is placed on the entire supply chain.
Furthermore, memory is not an isolated element within the cost. If its weight grows so strongly, it changes the economic balance of the entire data center. It is no longer just about buying cutting-edge GPUs, but about assuming that a much larger part of the investment is going to go into DRAM, HBM and fast storage. This forces us to redo profitability calculations, industrial priorities and supply agreements for several years to come.
The market may enter a phase of sustained pressure
At a Bank of America event, Michael Dell stated that “as memory per accelerator and system scale simultaneously expand across AI infrastructure, a structure is forming where total memory demand increases by approximately 625 times.”
The big question is whether the industry will be able to respond in time. The expansion of memory capacity is not improvised. It requires years, factories, equipment and enormous investment commitments. Meanwhile, demand does not wait. If AI deployment forecasts continue to accelerate, the market may enter a stage in which memory stops being an expensive companion and becomes one of the factors that decide who scales first and who is held back.
That explains why this estimate has attracted so much attention. Whether the final multiplier is exactly 625, as Dell’s CEO says, or ends up being lower, the underlying message is clear: The AI infrastructure will not only consume more chips, but much more memory per machine and much more total memory. And that completely changes the situation for manufacturers, cloud providers and companies that want to set up their own capacity.
The final reading is simple. AI is no longer just a race to have the best accelerator. Now it is also a war to ensure sufficient, fast and available memory in volume. Whoever controls that part better will have an advantage in costs, deadlines and deployment capacity. Anyone who arrives late may find themselves with a very expensive problem that is very difficult to correct.
That’s why this news matters so much. Not because the market is going to multiply exactly by a closed figure, but because it gives name to a reality that is already being noticed: memory is becoming one of the most strategic resources in the entire AI economy.
