For most of the past decade, memory became steadily better value. Capacities increased, prices fell, and specifications once associated with workstations found their way into ordinary laptops. Sixteen gigabytes became a sensible baseline, while 32GB stopped feeling excessive.
That progress has now gone into reverse. Conventional DRAM prices have risen sharply, manufacturers are reconsidering the specifications of cheaper products, and some of the world’s largest technology companies have begun passing the increase directly to customers. Apple has raised prices across parts of its Mac and iPad ranges, Microsoft will soon charge $499 for an Xbox Series S that launched at $299, and PC manufacturers are being forced to rethink what they can include without wiping out the margins on cheaper machines.
AI is the main reason. Hyperscalers are not buying the same desktop modules sold to consumers, but their spending is reshaping the market much earlier in the supply chain. Demand for AI infrastructure has made advanced memory far more valuable, pulling manufacturing capacity and investment towards the customers willing to pay the most. That matters because memory production is concentrated among a few companies and cannot be expanded quickly. Once suppliers move capacity towards high-bandwidth memory and large server contracts, companies buying conventional DRAM are left with less choice.
AI’s Memory Problem
The AI hardware boom is usually discussed in terms of processors. Nvidia’s GPUs dominate the market, while Google, Amazon, Microsoft and Meta are spending heavily on their own accelerators. Those processors are only useful if data can reach them quickly enough. Large models repeatedly access huge numbers of parameters while they are trained or used, and as processor performance has increased, memory bandwidth has become a larger constraint. An accelerator may be capable of extraordinary computational throughput, but some of that power is wasted whenever it has to wait for data.
High-bandwidth memory, or HBM, is designed to reduce that delay. Instead of arranging conventional memory chips around a circuit board, manufacturers stack several layers of DRAM vertically and place them beside the accelerator. This creates a much wider connection between processor and memory, allowing far more data to move at once.
That performance comes with a much harder manufacturing process. Individual dies must be fabricated, tested, thinned and bonded into stacks before the completed unit is packaged beside the processor using specialised equipment. A defect in one layer can compromise the entire stack, while a packaging failure may waste both the memory and the expensive accelerator beside it.
HBM also consumes far more manufacturing capacity per usable gigabyte than conventional DRAM. Demand from AI data centres therefore affects the memory inside consumer computers even though the products are different. Samsung, SK hynix and Micron are dedicating more wafers and packaging capacity to HBM because it offers better returns. Those same factories could otherwise have produced much larger quantities of ordinary memory.
Why Memory Companies Prefer AI Customers
Samsung, SK hynix and Micron dominate the global DRAM market. Few companies can compete because advanced memory fabrication requires enormous investment and years of process development. Even after a factory has been built and equipped, manufacturers still need to improve yields before it can produce chips profitably. The established suppliers therefore have a great deal of influence over where new capacity goes.
AI customers are attractive because they place very large orders and are willing to reserve production years in advance. Micron says its HBM output for 2026 is already allocated and has disclosed around $22 billion in customer commitments. Some of those agreements include deposits and minimum pricing, giving the company more certainty than it would receive from the consumer PC market. In an industry that has repeatedly suffered from sudden downturns, that kind of commitment is valuable.
The memory business has always been cyclical. Suppliers tend to expand when prices are high, but new factories take years to arrive. By the time they begin producing, demand may have weakened, leaving the market with more memory than it needs. Since DRAM from one major supplier can often substitute for DRAM from another, excess supply pushes prices down quickly. The factories are expensive to run but also expensive to leave idle, so companies keep producing as the market worsens, deepening the downturn.
The collapse in 2023 showed how severe this can become. Micron’s annual revenue fell from $30.8 billion to $15.5 billion, and the company recorded a net loss of $5.8 billion. SK hynix suffered a major operating loss, while Samsung’s semiconductor division lost trillions of won across several quarters. Those losses explain why suppliers are reluctant to respond to the current shortage by expanding conventional DRAM output as quickly as buyers would like. A factory approved during the AI boom may not produce meaningful volumes until the market has already changed.
HBM gives memory companies a way to reduce some of that uncertainty. It is more profitable, and its customers are prepared to make longer commitments. AI servers also require large amounts of conventional DRAM for their CPUs and supporting systems, so demand is rising beyond HBM itself. Much of the new investment, however, is still being directed towards the parts of the market offering the highest returns.
TrendForce reported that conventional DRAM contract prices rose by roughly 90 to 95 per cent in the first quarter of 2026 and forecast another increase of 58 to 63 per cent in the second. These figures describe the prices paid by major manufacturers rather than the cost of an individual RAM kit, which is why the consumer effect takes time to appear. Hardware companies can keep using components purchased under older contracts until that inventory runs out, but once it does, the higher cost has to show up somewhere.
What This Means for Hardware
Hardware manufacturers have a limited set of responses. They can raise the price of the final product, include less memory than they originally planned, or accept a smaller margin. Premium companies may be able to absorb the increase temporarily, but that becomes much harder at the cheaper end of the market. An extra $50 in component costs is inconvenient on a $3,000 workstation. On a $400 laptop, it can remove most of the profit.
The crisis will not always appear as a dramatic price increase. Some laptops will remain at 8GB for another generation when they might otherwise have moved to 16GB. Memory upgrades may become unusually expensive, while configurations that no longer make financial sense may quietly disappear. In other cases, companies are simply charging more for existing products.
Apple’s June price increases made the pressure much harder to ignore. The 512GB MacBook Air rose from $1,099 to $1,299, while the 1TB MacBook Pro increased from $1,699 to $1,999. The MacBook Neo, introduced only months earlier at $599, now starts at $699. These changes were not attached to a redesign or a new processor generation. Apple raised the prices of existing machines after saying it could no longer fully shield customers from higher memory costs.
Apple is better placed than almost any other hardware company to manage a component shortage. It buys at enormous scale, secures supply far in advance and operates with margins that most PC manufacturers would envy. The MacBook Neo shows how much the increase changes the product. At $599, it had a clear role in education and among buyers considering a Chromebook or an inexpensive Windows laptop. At $699, it moves into a more competitive part of the market and loses some of the appeal that justified its introduction.
The iPhone has not been included in the current round of increases, although the timing probably explains that. Apple is only a few months away from its usual September launch. Increasing the price of the current range in June would create an awkward mid-cycle change shortly before those phones are replaced. Any adjustment can instead be introduced with the next generation and discussed alongside the usual improvements. The unchanged price of the current iPhone range therefore tells us little about what Apple may do at its next event.
The $499 Xbox Series S
Microsoft’s Xbox increases are the clearest consumer example because the company has linked them directly to rising memory and storage-component costs. The Xbox Series S launched in November 2020 for $299. It was designed as the affordable console of the generation, offering less performance than the Series X and removing the disc drive in exchange for a much lower price. The 512GB model later rose to $399. From August 2026, Microsoft will add another $100, bringing it to $499. The Series X, which launched at $499, is set to reach $799.
A six-year-old budget console will therefore cost 67 per cent more than it did at launch and exactly as much as the original flagship. This is the reverse of the normal console cycle. Hardware usually becomes cheaper to manufacture as a generation progresses. Production matures, chip yields improve and component costs tend to fall. Even when the official price stays the same, inflation makes the machine cheaper in real terms.
The current Xbox generation has moved in the opposite direction. The hardware is ageing, yet the price is climbing sharply. Microsoft says the cost of memory and storage components has multiplied within months and expects the pressure to continue into 2027. Its August changes add $100 to 512GB models and $150 to 1TB models. The company is also withdrawing the 2TB version.
The Series S was built around affordability, and at $499 it no longer fulfils that role convincingly. The removal of the 2TB model also shows that the cost increase is influencing which configurations Microsoft believes are worth keeping in the range. This is not a small adjustment intended to preserve margins. The economics of the product line have changed.
Allegations of Collusion
The structure of the memory market has attracted legal attention. Samsung, SK hynix and Micron collectively control close to 90 per cent of global DRAM production, and a proposed US class-action lawsuit accuses them of coordinating reductions in conventional DRAM output. The complaint alleges that the companies used the shift towards HBM to restrict supply and raise prices across the wider market.
Those claims have not been proven. There are clear commercial reasons for each supplier to prioritise HBM independently. AI demand is strong, HBM requires more capacity, and hyperscalers are prepared to pay more than consumer-hardware manufacturers. A concentrated market and rapidly rising prices do not by themselves prove collusion.
The allegations still matter because the memory industry has faced previous antitrust cases involving price fixing. Its concentration also means that decisions made by three companies can influence almost the entire market. Each supplier may independently decide that expanding conventional DRAM too quickly would lower prices and hurt margins. If all three reach the same conclusion, supply remains tight even without an agreement between them.
The legal question is whether the companies simply responded to the same incentives or whether they coordinated their decisions. Until the case develops further, the claims should remain clearly identified as allegations, but the history of the industry makes them difficult to ignore.
What Happens If the AI Boom Slows?
The current shortage depends heavily on continued investment in AI infrastructure. The largest data-centre operators are expected to spend hundreds of billions of dollars in 2026, and estimates for the wider build-out over the next several years run into the trillions. There are good reasons for that spending. Demand for AI services is growing, cloud providers are trying to secure enough computing capacity, and the technology is already producing real revenue.
The question is whether that revenue will justify the scale of the investment. Hyperscalers are building because they expect demand to rise, but they are also acting defensively. None wants to underinvest while a competitor gains an advantage. That can produce too much capacity even when each company believes it is making a sensible decision.
The internet provides a useful comparison. It transformed the economy, but the telecommunications boom of the late 1990s still led to excessive investment and large losses. The technology was real; the assumptions about how quickly every project would become profitable were not. AI could follow a similar pattern. It can remain economically important while some of the infrastructure built around it proves excessive.
Memory companies would be highly exposed to a correction. New factories are being approved while demand is unusually strong. If data-centre projects are delayed just as additional capacity becomes available, the market could shift quickly from shortage to surplus. Long-term contracts would soften the first stage of that decline because some customers are committed to paying for reserved capacity, but those agreements cannot remove the cycle entirely. Contracts expire, future orders can be cut, and projects can be postponed.
Improvements in efficiency could also affect demand. Smaller models and better inference techniques may allow companies to perform the same work with less memory. Greater efficiency can encourage more use, so the overall effect is uncertain, but suppliers are still investing on the assumption that memory requirements will continue rising at an exceptional rate.
My view is that AI itself is not a bubble in the sense that the technology will disappear. The infrastructure boom may still have gone too far. Some data centres will produce strong returns, while others may struggle to justify their cost once the race to build capacity slows.
Why Supply Will Take Time to Recover
The memory industry entered the AI boom after a severe downturn. During the pandemic, demand for computers surged and customers overordered. When sales weakened, excess inventory pushed prices down, leading manufacturers to cut output and postpone investment. Those decisions stabilised the market but left less spare capacity when AI demand accelerated.
New fabrication plants take years to build and qualify. HBM also depends on advanced packaging, so increasing wafer production alone does not remove the bottleneck. Samsung, SK hynix and Micron are investing heavily, but much of the added output will not arrive until 2027 or later. A large share of that investment is also directed towards HBM and server memory rather than inexpensive consumer DRAM.
Manufacturers have little incentive to expand faster than necessary. They remember what happened in 2023 and know that a sudden slowdown in AI spending could leave them with too much capacity. The cautious approach protects their margins, although it keeps prices high for longer.
Why the New Prices Might Remain
Lower memory costs would not automatically return MacBooks or Xbox consoles to their previous prices. Companies set prices according to demand and competition rather than by adding a fixed margin to the cost of each component.
If buyers continue purchasing the MacBook Air at $1,299, Apple may see little reason to restore the old $1,099 price. It could instead increase the standard memory allocation or offer more frequent discounts. Microsoft could revise the Series S and retain the $499 starting point, using additional capacity or bundled services to make the package more attractive.
Once a higher price has been accepted, future products are compared with it. The next MacBook will be judged against the current $1,299 model rather than the version that cost $1,099. The next Xbox will launch into a market where Microsoft has already established $499 as an entry price.
Commodity RAM is more exposed to competition because buyers can compare equivalent modules from several brands. Complete devices work differently. Apple controls macOS hardware, while console customers may have years of purchases tied to one platform. That gives manufacturers more freedom to preserve higher prices even if component costs later fall.
Who Pays for the Boom?
Samsung, SK hynix and Micron are enjoying one of the strongest periods the memory industry has seen in years. HBM has improved their margins and given them customers willing to reserve future production. The largest AI companies also benefit because their scale allows them to secure supply before smaller buyers can.
The cost falls across the rest of the market. Consumers pay more for existing products. PC manufacturers are forced to compromise on specifications. Smaller AI companies and research institutions face higher infrastructure costs while competing with hyperscalers that can commit billions to future supply.
The Xbox Series S shows how far the effects have travelled. It launched at $299 as an affordable entry point and will soon cost $499. Microsoft attributes the increase directly to memory and storage-component costs that have risen rapidly.
Memory manufacturers are following the strongest source of demand. Right now, that demand comes from AI data centres, and consumer hardware is becoming more expensive as a result. What happens next depends on whether the AI build-out continues at its current pace. Memory companies have spent decades moving between shortages and gluts. The present boom has made them exceptionally profitable, but it may also be setting up another downturn if demand slows before the new capacity arrives.