Memory Starvation: Why It's Worse Than You Think - Tom's Hardware

The problem of running out of memory is not only about RAM and SSDs, but a much larger problem with unexpected consequences!

As we know, memory prices are skyrocketing, and it's clear that the blame is entirely on artificial intelligence.The fact that it is now well established, but which, in my opinion, should lead us to ask ourselves a specific question: which sectors will be affected by this shortage?

Well, yes, I'll tell you right away that things are going to get worse, and now that we're in this mess, I think it's our duty to try to figure out when we're going to get out of it.Can we assume that it will all be over in a few months, or are we likely to carry these cost increases with us for years?

As you can see, there are quite a few questions, so let me tell you what they are, and at the same time we will try to understand what is happening and especially why!

Why and what did you do?

For the first time in more than a decade, the industry is facing simultaneous and uncontrolled growth driven by three factors: the need for more RAM-dense servers, the aggressive expansion of AI data centers, and market consolidation dominated by multiple manufacturers who can directly affect wafer flow.

Companies that build and operate AI infrastructure have literally bought everything available on the market.The hyperscaler has purchased all the DRAM modules for its servers.This is usually DDR5, but they also buy DDR5 and DDR4 memory for the consumer sector, which they use in all situations where ECC memory is not required, that is, in all systems that support AI: web servers, caching systems, API gateways, recording systems, almost everything around the server. In the meantime, SSD and NAND memory are also replenished.This is because artificial intelligence requires large amounts of memory and quickly accessible storage to be able to process data.

The impact was immediate: PCs and laptops were sold, lines were converted to modules for high-margin professional services, and the price of parts for home computers skyrocketed.

The price increase we are witnessing and experiencing is generally a structural crisis generated by the rise of AI, which upset the traditional balance between server memory and user memory.The defining factor here was the speed with which the AI ​​sector demanded new servers, starting a real race to demand new clusters and new GPUs.An event that clearly does not correspond to the speed at which a factory that produces DRAM and NAND can be expanded, which does not take weeks or months, but years, billions of investments and careful planning;This raises the possibility, as mentioned at the beginning, that this shortage will last not months, but years, and affect more sectors than one can imagine.

How much profit do producers make?

Having understood that this is therefore a real problem, it is worth asking at this point whether manufacturers are facing it.To do this, as always, I would say to start with a price analysis.

Many types of DRAM saw price increases of 50% to 60% in wholesale prices and 200% to 300% in off-the-shelf prices.At the heart of the problem is the promise of server memory.

Between October and November 2025, major cloud providers in the United States and China agreed to a 50% increase in RDIMMs for data centers, instead of only 70% of what was ordered.In other words, they are paying half the price for less goods than before, which speaks volumes for the bargaining power of manufacturers.

According to Reuters, the benchmark price of a 32GB DDR5 module sold by Samsung rose from $149 in September to about $239 in November 2025, an increase of more than 60% in two months.Other capacities, such as 16GB and 128GB, have increased by 40% to 50%.

On the consumer side, we have cases where 64GB DDR5 devices have gone from around $150 to over $500 in a matter of months.In other words: a single pack of RAM can cost more than a PlayStation 5. At the same time, manufacturers such as Adata have clearly warned that the prices of RAM and DDR4/DDR5 SSDs will rise unevenly, while some American chains such as Micro Center have begun to remove fixed prices from the shelves, replacing them with general price indicative checks that appear on the market.

To understand why manufacturers have chosen to raise prices so much, we must also take into account the market in recent years, and not only the issue of artificial intelligence.After two years of low prices, with margins squeezed to a minimum and production lines remaining below capacity, Samsung, SK Hynix and the other memory giants were faced with a sudden explosive demand, driven by hyper-scalers like Google, Microsoft, Amazon and below, who were willing to pay anything to get enough DRAM to run AI-supplied data centers.

In a market where memory manufacturers are few and far between, it's clear they've been wringing their hands rather than trying to control said prices.Reports suggest the contract is being renegotiated to raise prices by up to 50% in the fourth quarter of 2025 or lose shipping priority, while Samsung itself has decided to increase the prices of some of its most popular DDR5 chips by up to 60% from September, making the message clear.That is: as long as demand for AI is strong enough, manufacturers will not push prices back to previous levels.Instead, they will make the most of this window by focusing on years of low-margin investments in wafers and higher-margin technology innovations such as server memory, HBM, and high-performance DRAM.

In this situation, it is now clear that selling memory to the consumer market is no longer a priority, with ramifications for the availability and pricing of the whole thing.So to answer the original question, I would say "yes, manufacturers use it".

How do we get out of this crisis?

For the timing of this crisis, let's say right away that we think the first signs of a return to more stable prices will be visible only in 2027. This means that in 2026 no major solution to this shortage is expected, and producers will continue to collect billions of dollars in profits.

But apart from that, let's try to understand how supply and demand will change, and what will affect one and the other: on the supply side, the only factor to consider is the increase in production capacity, so the construction of new factories, the updating of the existing lines and the introduction of more advanced processes.

Manufacturers such as Samsung, SK Hynix and Micron have already announced massive investments, with new product lines for advanced DRAM, new generation NAND and, above all, the most profitable memories such as HBM and high-density server modules.However, it can take 18 months to three years between capital allocation and the launch of the first products off the assembly line.This means that by the end of 2024, when the increase in demand becomes clear, the additional capacity we are targeting will actually not be achieved until 2027.

Meanwhile, manufacturers could try to rebalance internally by moving some wafers from server DRAM and HBM to consumer DDR4 and DDR5 modules.But how will this strategy actually benefit him?After all, why would they undercut an AI that is currently making a lot of money?

The other part of the equation is demand, and here the scenario is much more unpredictable: the main driver of the last two years has been the hype around AI, which has led all kinds of companies to launch experimental projects, build private clusters and pre-book servers, creating a demand that is currently largely on paper.If for some reason AI enthusiasm wanes, we could see a natural slowdown in memory orders, especially in areas where purchasing is more impulsive.The same will be true for operators who have accumulated excessive stock out of fear of not being able to find available modules: once internal needs are met, many of these customers will reduce orders in subsequent quarters, providing relief to the market.

Another possible way to reduce demand relates to software optimization.Efforts to improve AI models by reducing the memory required for instruction and training have grown rapidly in recent months.If these improvements combine, the market may see a reduction in demand, not because AI will stop development, but because it will grow more efficiently.

Finally, one cannot ignore the possibility that many companies will scale back or deploy non-core AI projects if hardware costs are high.However, it is impossible to predict the extent to which this may actually occur.Not often, because these things happen, but I doubt a year is enough for it to happen.

Anyway, let's define possible scenarios for the next few years.

The first and most likely scenario is that prices continue to rise (yes, it’s not over yet), then stabilize at extreme levels and start to decline from late 2026, early 2027. That said, memory is unlikely to return to pre-boom levels, so we'll have to get used to the new baseline.

The second, more promising scenario is that demand stabilizes faster than expected and producers can increase production faster than expected.If that happens (though I think it's unlikely), we could see some adjustment and fewer price cuts in the middle of next year.

The third scenario, on the contrary, is pessimistic, at least for the industry: if producers continue to invest heavily and ramp up production quickly, and at the same time the hype about artificial intelligence dies down, we will find ourselves in a situation of overcapacity, that is, supply exceeds demand.In this case, producers will be forced to cut prices brutally to empty warehouses.In any case, the most likely future is one in which prices will not drop before 2027.

What will cost more?

It is clear that RAM and SSD memory modules are the most expensive products, at this point we can ask ourselves: "What other products will this shortage affect?"

The truth is that the whole (more or less visible) ecosystem of products based on these chips will increase in price: smartphones, laptops, gaming PCs, graphics cards, game consoles, even cars and medical devices. The dynamic is simple: memory and storage are one of the main cost factors in the industry for every modern electronic device.onboard, or squeeze margins. For manufacturers, the most likely solution is a combination of the three.

Apart from the aforementioned models and SSDs, the first area where we see clear results is smartphones.The most affected are low- and mid-range models, where every euro counts and margins are tight.At the high end of the price spectrum, they are more likely to try to hide price increases by maintaining an official price list and offering fewer discounts and promotions compared to previous years.In this sense, it is reasonable to think that next year's iPhone will also be affected.Maybe that's what's happening.

The other main front is laptops and PCs in general.Companies such as Dell and HP are already on the fence, warning of possible price increases.Business and professional models, which often offer more RAM and larger SSDs, risk a significant increase in price.For consumer laptops and gaming laptops, the most common change, at least in the short term, will be a specific low-cost factory: the base configuration with less RAM or a smaller SSD.

The crazy deals we saw in the years of near-segmentation of memory will become fewer and fewer, and many low-end PCs will be stuck with 8GB of RAM and 256-512GB SSDs to keep psychological prices below a certain threshold.Of course, graphics cards are not immune to that, which is a shame because they have started to decline.

At the moment, for example, AMD has already informed its partners that it will increase the price of GPUs by at least 10% in 2026, clearly stating that the cost of VRAM memory is the main reason, and some sources believe that NVIDIA is also evaluating similar steps to increase internal margins.

The discussion naturally extends to gaming consoles, both today and next-gen.There are rumors of ongoing Xbox cost increases, while the PlayStation, given a larger pool of memory, isn't expected to see any further increases, at least in the short term.Looking a little further, to the hypothetical PlayStation 6 or the next Xbox, the conclusion is almost inevitable: if the memory remains structurally unchanged until 2027 and beyond., manufacturers will have only two options.Either you launch new machines at a higher starting price than the current generation, or you sacrifice something in terms of RAM or storage.

Finally, there are less well-known industries such as automotive, medical equipment, networking equipment, and IoT, where memory is at risk of price increases and release delays.Unfortunately for us, we have to be prepared to spend more, offer less, and even offer something with less memory, which obviously won't please anyone.