Estimated reading time: 15 minutes
Why memory prices are soaring, what AI data centres have to do with it, and why businesses and consumers may be feeling the effects for years to come.
Something unusual is happening to one of the least glamorous parts of modern technology.
RAM.
A quick note before we start: this is a slightly longer and more detailed read than usual.
Prefer the short version? We’ve also written a plain-English overview of the RAM crisis, with the technical detail stripped out.
Why AI Could Make Your Next Server, Laptop or Phone More Expensive
For years, memory and storage broadly followed a familiar pattern. Technology improved, capacity increased and the cost per gigabyte gradually came down. Businesses replacing servers could reasonably expect more performance for their money. Consumers buying a new PC, games console or phone generally got more memory and storage than the generation before.
Today, that assumption is being challenged.
At IntraLAN, we have seen like-for-like server quotations rise to around three times the level we were seeing approximately 18 months ago. Some server RAM and SSD configurations have increased by around four times.
And this isn’t simply something happening inside our supplier network.
TrendForce reported that conventional DRAM contract prices increased by approximately 93% to 98% in the first quarter of 2026 alone. Enterprise SSD contract prices rose by approximately 80% during the same quarter, as demand significantly exceeded supply.
At the consumer end, a German retail price index reported in August puts DDR5 memory at around 486% of its July 2025 price level, effectively close to five times the price. Internal SSDs were around 2.2 times their pre-crisis level. Germany is only one retail market, so those figures should not be treated as universal pricing, but they illustrate both the direction and severity of the change.
So what happened?
The short answer is AI.
The more interesting answer is that the race to build artificial intelligence has created an infrastructure boom on an extraordinary scale. It requires huge quantities of specialised memory, conventional server RAM and flash storage.
And so far, the companies building it show very little appetite for slowing down.
🧠 First, AI needs an extraordinary amount of memory
When we talk about an everyday business laptop, 16GB of RAM remains a useful point of comparison.
AI systems operate on a very different scale.
Modern AI accelerators use High Bandwidth Memory, normally shortened to HBM. It is a specialised form of DRAM designed to sit close to powerful GPUs and move enormous quantities of data at extremely high speed.
It isn’t the same memory module you would install in an office PC or conventional server.
But it comes from the same wider memory manufacturing industry, and AI needs a lot of it.
Take NVIDIA’s GB200 NVL72, one of the rack-scale systems being deployed in modern AI data centres.
A single rack connects 72 Blackwell GPUs and 36 Grace CPUs. NVIDIA specifies 13.4TB of HBM3E GPU memory, alongside 17TB of LPDDR5X CPU memory.
To put that into everyday terms:
13.4TB of GPU memory alone is roughly equivalent in capacity to the RAM in 840 laptops fitted with 16GB each.
Add the CPU memory and a single rack contains approximately 30.4TB of memory, equivalent in raw capacity to around 1,900 16GB laptops.
That comparison is about memory capacity, not computing performance. A laptop and an AI accelerator are obviously very different machines.
But it gives a sense of scale.
And that is one rack.
NVIDIA’s newer Vera Rubin NVL72 specification goes further again, with 20.7TB of HBM4 GPU memory and 54TB of CPU memory per rack. The GPU memory alone equates to roughly 1,300 16GB laptops. NVIDIA currently labels the Vera Rubin specifications as preliminary and subject to change.
This appetite for specialised memory matters because HBM consumes a disproportionate amount of manufacturing capacity.
TrendForce estimates that HBM will account for around 22% of DRAM wafer input among the three largest memory manufacturers by the end of 2026, while producing only around 9% of their DRAM bits. By the end of 2027, it expects HBM to consume around 30% of DRAM wafer input.
In simple terms, making more AI memory can leave less manufacturing capacity available for other types of memory.
So even if your business has no intention of buying an AI server, it can still feel the effects.
🏭 Now multiply one rack by thousands
Perhaps the easiest way to understand what is happening is to stop thinking about individual AI chips and look at an actual AI data centre.
The original phase of xAI’s Colossus supercomputer in Memphis used 100,000 NVIDIA Hopper GPUs. Supermicro’s case study describes compute racks containing eight servers, each with eight H100 GPUs, giving 64 GPUs per compute rack.
At that density, 100,000 GPUs are equivalent to roughly 1,563 fully populated GPU compute racks, before considering additional networking, CPU, storage and supporting infrastructure.
Each H100 carries 80GB of high-bandwidth memory.
That means the original 100,000-GPU deployment contained around 8 petabytes of GPU memory.
Put another way:
The GPU memory in that first Colossus deployment was equivalent in raw capacity to the RAM in around 500,000 laptops fitted with 16GB each.
And xAI says it subsequently doubled Colossus to around 200,000 GPUs, with a roadmap towards one million GPUs.
xAI is only one participant in the race.
💾 AI isn’t just consuming RAM. It is consuming SSDs too
Training an AI model requires data.
Running AI services for millions of users requires enormous amounts of data to be stored, retrieved and moved quickly.
Increasingly, that means enterprise SSDs.
The NAND flash used in an SSD is different from DRAM, but the demand story is becoming remarkably similar.
Counterpoint Research estimates that enterprise SSDs accounted for 48% of all NAND bits shipped globally in the second quarter of 2026, up from only 26% a year earlier. It expects server enterprise SSDs to absorb more than half of global NAND shipments by the end of the year.
TrendForce describes an exceptionally severe supply-demand imbalance in enterprise SSDs during the first quarter of 2026, with supplier inventories falling to historic lows and contract prices increasing by approximately 80% in a single quarter.
This is particularly important as AI shifts from simply training enormous models towards running inference continuously and supporting increasingly autonomous AI agents.
Those systems need fast access to huge datasets, cached information, model data and context.
Storage is no longer just somewhere to save the files afterwards.
It is becoming part of the AI computing architecture itself.
🌍 And there are few signs of the infrastructure race slowing
If today’s AI data centres represented the end of the build-out, memory manufacturers could model the demand, expand production and gradually restore balance.
The problem is that the next wave is already being planned.
When OpenAI announced Stargate in January 2025, it committed to securing 10GW of US AI infrastructure by 2029.
In April 2026, OpenAI said it had already surpassed that milestone in secured infrastructure capacity, with more than 3GW added during the preceding 90 days. Its flagship Stargate facility at Abilene, Texas, operates NVIDIA GB200 systems.
The distinction between secured capacity and capacity already operating is important. This does not mean more than 10GW is online today.
It means the pipeline has already moved beyond a target that originally stretched to 2029.
Meta’s expanding Richland Parish data centre in Louisiana will meanwhile be home to Hyperion, its most powerful AI training cluster. Meta describes it as its largest data centre development and says the expansion represents more than $50 billion of total investment.
Then there is Utah.
The proposed Stratos Project Area in Box Elder County covers approximately 40,000 acres and combines hyperscale AI and cloud infrastructure with dedicated on-site energy generation.
Official project material describes 3GW as the committed first phase, with potential generation of up to 9GW as part of a wider multi-site strategy.
Nine gigawatts is a difficult number to visualise.
US Energy Information Administration figures show that Utah sold around 2.87 million megawatt-hours of electricity during May 2026. Averaged across the month, that is roughly 3.9GW of power consumption across the entire state.
The theoretical 9GW Stratos figure is therefore more than twice that recent statewide average.
That doesn’t mean a 9GW data centre is about to appear fully formed in Utah. Stratos is a phased development and 9GW is a potential full build-out figure.
But that almost misses the point.
A few years ago, large data centres were commonly discussed in megawatts.
We are now seriously discussing AI infrastructure in gigawatts.
💰 The world’s biggest technology companies appear more worried about falling behind than overspending
There is another reason to question whether demand will disappear quickly.
Look at how much the companies competing in AI are investing.
Alphabet generated $39.1 billion of operating cash flow in the second quarter of 2026, but spent $44.9 billion in capital expenditure. The vast majority of that capex went into technical infrastructure supporting AI investment.
The result was negative free cash flow of $5.9 billion for the quarter.
Alphabet has also increased its 2026 capital expenditure guidance to $195 billion to $205 billion, and says capex is expected to increase significantly again in 2027.
Meta generated $31.86 billion of operating cash flow in Q2 2026. Capital expenditure, including finance lease principal payments, reached $31.08 billion, leaving free cash flow of just $784 million for the quarter. It expects 2026 capital expenditure of $130 billion to $145 billion.
Amazon generated $161.4 billion of operating cash flow over the twelve months to June 2026, but its free cash flow moved to an outflow of $7.6 billion. Amazon says the deterioration primarily reflects increased investment in artificial intelligence.
Microsoft generated $182.9 billion of cash from operations during its 2026 financial year and added $115.9 billion of property and equipment. In its final quarter alone, capital expenditure reached $41 billion, with Microsoft saying roughly two thirds related to shorter-lived assets, primarily CPUs and GPUs used across both AI and non-AI infrastructure.
These businesses are not facing imminent financial failure.
Far from it.
They remain some of the largest, most profitable and financially powerful companies on Earth.
But the scale of the expenditure tells us something important.
The world’s biggest technology companies currently appear to regard the competitive risk of falling behind in AI as greater than the risk of spending too much to stay in the race.
That is an interpretation rather than something we can read from a balance sheet, but the behaviour is difficult to ignore.
And it matters to everybody else trying to buy memory.
A small business replacing a server is obviously not bidding directly against Google or Microsoft for a box of RAM.
But somewhere upstream, both are competing for access to finite semiconductor manufacturing capacity, advanced packaging, servers, storage and components.
One buyer is purchasing a replacement server.
The other may be committing tens of billions of dollars to infrastructure.
🏭 So why don’t Samsung, SK hynix and Micron simply make more?
This seems like the obvious solution.
Prices rise.
Manufacturers build more factories.
Supply catches demand.
Prices fall.
Unfortunately, semiconductor manufacturing does not work quickly.
The DRAM industry is also highly concentrated.
Counterpoint Research estimates that in the second quarter of 2026 Samsung accounted for approximately 39% of global DRAM revenue, SK hynix 26% and Micron 25%.
Together, those three companies represented roughly 90% of global DRAM revenue.
They are investing heavily.
SK hynix approved approximately 54 trillion Korean won of additional investment in its Yongin Y2 and Cheongju M17 facilities in August 2026.
Yet the first cleanroom at M17 is scheduled to open in December 2028, while the Y2 cleanroom is planned for June 2029. Even then, SK hynix says equipment installation and actual production capacity will be expanded sequentially in line with customer demand.
Micron has increased its planned US investment to more than $250 billion through 2035. Its first new Idaho fab is expected to produce its first wafers in mid-2027, with a second following in late 2028.
And Micron’s own assessment of the wider market is striking.
The company says memory demand continues to significantly exceed supply and expects tight conditions to persist beyond 2027. Even though it expects industry supply to improve gradually in 2028, Micron says it currently has no clear line of sight to when memory supply will catch increasing demand.
It points to the scale and complexity of new fabs, construction lead times, skilled-worker shortages, permitting and the need for additional energy infrastructure as constraints on expansion.
There is another difficulty. The technology driving much of today’s demand barely existed in its current form four years ago. In March 2022, NVIDIA had only just announced its H100 generation of AI accelerators. HBM3E did not enter volume production until 2024, and today’s 72-GPU Blackwell rack-scale systems were only announced that same year.
That means memory manufacturers are being asked to make multi-billion-dollar decisions about factories that may not reach full production until 2028 or 2029, based partly on demand for a category of hardware that has transformed almost beyond recognition in less time than it takes to build the factory.
That is the fundamental problem.
AI demand can accelerate in months.
A semiconductor factory can take years.
🎈 But what if the AI boom turns out to be a bubble?
This is where the manufacturers face an uncomfortable decision.
The current shortage tells Samsung, SK hynix, Micron and others that the world needs more capacity.
But a modern semiconductor fab costs billions to construct, equip and bring into production.
What happens if they invest based on today’s demand forecasts and the economics of AI change before those factories are running at full capacity?
Perhaps AI models become significantly more memory efficient.
Perhaps new chip designs alter the mix of components required.
Perhaps businesses decide some AI use cases don’t generate sufficient return.
Or perhaps today’s extraordinary investment cycle eventually cools.
Manufacturers know that building too little capacity means missing revenue today.
Building too much could leave them with expensive factories and oversupply tomorrow.
That helps explain why investment is enormous but still measured. SK hynix, for example, is explicitly planning to phase actual production capacity in line with customer requirements rather than simply filling every new cleanroom with equipment immediately.
The memory industry has been through boom-and-bust cycles before.
Nobody wants to solve today’s shortage by creating tomorrow’s glut.
⏳ How long could the memory shortage last?
That may be the hardest question in the whole story to answer.
For much of the current cycle, the expectation has been that additional semiconductor capacity and improvements in manufacturing efficiency should start bringing some relief during 2027 and 2028.
TrendForce currently expects NAND supply growth to begin outpacing demand during 2027, potentially easing some of the pressure in the second half of the year. It also notes that more substantial output from new manufacturing capacity is unlikely to arrive until 2028.
Micron offers a similarly cautious outlook across the wider memory market. It expects industry supply to improve gradually during 2028, but says it still has no clear line of sight to the point at which supply will fully catch increasing demand.
More recent warnings suggest even that timetable could prove optimistic.
Phison CEO Pua Khein-Seng has reportedly been told by NAND manufacturers that moving from a decision to invest in additional production to meaningful mass production could take around four years. If substantial new investment decisions are being made now, that potentially pushes some additional supply towards 2030.
That does not mean SSD shortages will definitely continue until 2030, and there is clearly disagreement over when the market will begin to rebalance.
But the range of forecasts tells us something important.
The industry is not talking about a problem that can necessarily be fixed over the next few quarters. Even the more optimistic scenarios involve years rather than months, while the more pessimistic ones extend towards the end of the decade.
And while those factories are being planned and built, AI demand is not standing still.
That is what makes the timescale so difficult to predict.
🛒 Why should someone who doesn’t run a data centre care?
Because memory is everywhere.
The same underlying technologies and manufacturing ecosystem increasingly sit behind the devices businesses and consumers buy every day.
Servers and business storage
This is where we are already seeing some of the clearest effects. Higher DRAM and enterprise SSD prices feed directly into server configurations, storage arrays, virtualisation hosts, databases and backup infrastructure.
A server refresh budget written 18 months ago can now look wildly optimistic.
Desktop PCs and laptops
TrendForce estimates that under normal conditions DRAM and SSDs account for roughly 15% of the bill of materials of a mainstream notebook.
Following the recent price increases, it estimates that share could exceed 30%.
In one model based on a $900 notebook, TrendForce calculated that memory inflation alone could require a retail price increase of more than 30% if manufacturers and distributors maintained existing margins.
That doesn’t mean every £800 laptop suddenly becomes £1,050.
It means manufacturers face choices.
Increase the price.
Absorb some of the cost.
Reduce margins.
Change specifications.
Or delay product transitions.
Smartphones and tablets
The situation is similar.
TrendForce estimates that contract prices for the memory in a mainstream 8GB RAM and 256GB storage smartphone configuration were nearly 200% higher year-on-year in the first quarter of 2026.
Memory that historically represented around 10% to 15% of a smartphone’s bill of materials could now represent 30% to 40% in some configurations.
At the same time, on-device AI is itself increasing memory requirements. TrendForce estimates that some edge AI models require 40GB to 60GB of system storage simply for local AI processing and caching.
So AI is applying pressure from both directions.
It is consuming memory in the data centre and encouraging devices at the edge to use more of it too.
Games consoles
An Xbox Series X contains 16GB of GDDR6 memory and a custom NVMe SSD. Sony’s current PS5 Digital Edition is sold with an 825GB SSD.
Those are entertainment devices, but their components still exist inside the wider memory and storage supply chain.
Future consoles, graphics cards and gaming PCs cannot escape the economics of that supply chain.
Cars
Modern vehicles increasingly depend on memory and storage for infotainment, driver-assistance systems, connectivity and software-defined features.
The importance of securing future supply is illustrated by Micron signing long-term strategic memory and storage agreements with both General Motors and Ford during July 2026.
So the consumer buying their next laptop, phone, games console or car may never knowingly buy a DRAM chip or NAND package.
They can still end up paying for the shortage.
🔮 What happens next?
There are two powerful forces pulling in opposite directions.
On one side, memory manufacturers are investing tens of billions in additional capacity.
New processes should improve output.
Higher prices will suppress some consumer and business demand.
New fabs will eventually come online.
All of those things should help.
On the other side, the AI infrastructure build-out remains extraordinary.
OpenAI says its secured Stargate capacity has already passed a 10GW target originally set for 2029. Meta is investing more than $50 billion in a single Louisiana data-centre development. xAI is talking about a roadmap towards a million GPUs. A 40,000-acre development in Utah contemplates energy infrastructure potentially reaching 9GW.
At the same time, the next generations of AI hardware are carrying increasing quantities of HBM, while inference and agentic AI are creating rapidly growing demand for high-capacity enterprise SSDs.
That makes predicting the point at which supply finally catches demand extremely difficult.
What we can say with more confidence is that businesses probably shouldn’t build future technology budgets around the assumption that memory prices will simply return to their 2024 or early-2025 levels in the near future.
For a CIO or CTO, that means hardware-refresh planning needs to account for component volatility, potentially shorter quote-validity periods and greater budget uncertainty.
For a CISO, indefinitely extending the life of ageing infrastructure simply to avoid today’s prices creates a different set of risks around resilience, warranty, vendor support and security.
For a business owner, a project that was economically straightforward two years ago may now require a very different budget or approach.
And for consumers, the consequences could show up as higher prices, smaller base specifications, more expensive upgrades or manufacturers keeping existing product generations on sale for longer.
🧭 Don’t panic buy. But do plan earlier.
None of this means businesses should rush out and buy hardware because someone predicts prices might rise next month.
Technology purchasing should still start with what the business actually needs.
But the market has changed enough that timing deserves more attention than it did before.
If you know a server, storage platform or significant fleet of PCs is likely to need replacing over the next 12 to 24 months, it makes sense to start that conversation earlier.
Understand what your current hardware can realistically support.
Look at warranty and end-of-support dates.
Work out which upgrades are genuinely required.
Consider whether cloud, on-premises or hybrid infrastructure changes the economics.
And most importantly, don’t assume that delaying a necessary project automatically means the hardware will be cheaper when you eventually buy it.
The bigger story isn’t really RAM
The extraordinary thing about the current memory crisis is that RAM itself is almost incidental.
What we are really watching is the physical reality of the AI boom.
AI is often discussed as software. Chatbots. Copilots. Agents. Automation.
Behind those services are enormous physical systems containing GPUs, CPUs, networking equipment, RAM and increasingly vast amounts of high-performance storage.
They require land.
They require factories.
They require billions of dollars.
They require gigawatts of electricity.
And they require memory on a scale that is beginning to affect what everyone else pays for it.
And increasingly, infrastructure measured not just in megawatts but in gigawatts.
Whether today’s level of AI investment ultimately proves justified, excessive or somewhere in between remains to be seen.
The technology companies themselves cannot know that with certainty.
The memory manufacturers cannot know it either.
That uncertainty is part of the reason new supply cannot simply appear overnight.
For now, however, the world’s largest technology companies are continuing to spend, new AI data centres are being planned on an extraordinary scale, and meaningful new semiconductor manufacturing capacity takes years to deliver.
That is why the RAM crisis matters far beyond the IT department.
If your business is planning a server, storage or PC refresh over the next 12 to 24 months, it may be worth revisiting the assumptions behind the budget and timing.
Market figures, company financial information and infrastructure announcements checked against published sources as at 19 August 2026.
