There is a causal chain that begins in an artificial-intelligence data centre and ends in the pocket of someone who may never have used a chatbot. That chain is the theme of this analysis, and it is worth tracing in full because it illustrates one of the quietest gaps in today’s digital economy: the one separating those who benefit from AI from those who pay, unknowingly, for part of its physical infrastructure.
The starting point is a figure the consultancy Gartner published on 8 April 2026 and that the sector has already christened with a name of its own: memflation. Memory —the chips every electronic device needs to process and store data— has become abruptly and persistently more expensive. According to Gartner’s estimates, annual DRAM memory prices will rise 125% in 2026, and NAND flash memory prices 234%, with no significant relief expected until late 2027. It is not a passing spike: it is, in analysts’ words, a multiyear structural shift in the memory market, driven by AI demand.
The question that organises this analysis is not only why prices are rising, but who ends up paying them. And the answer, as we will see, is counterintuitive.
The mechanics: why AI eats the memory
To understand memflation, you have to understand a substitution happening in the chip factories. Three companies —Samsung, SK Hynix and Micron— control more than 95% of global DRAM production. Those same factories can produce, with the same silicon-wafer capacity, two different things: the conventional memory that goes into computers and phones, or high-bandwidth memory (HBM), a stacked and far more profitable type of chip used in AI accelerators.
Faced with the voracious demand of data centres, manufacturers have systematically reallocated their capacity toward HBM, leaving consumer memory in critical shortage. The result is that, according to sector analyses, data centres today consume around 70% of all memory chips produced in the world. The consequence is arithmetic: if most of the capacity goes to AI, what is left for everything else is less and more expensive.
The most revealing figure of this inversion came from Counterpoint Research: the spot price of DDR4 —a conventional, supposedly “cheap” memory— came to exceed the contract price of HBM3e, the elite memory for AI. When the cheap costs more than the expensive, the market is signalling a deep distortion.
Who pays: the chain down to the consumer
Here the central paradox appears. The big AI buyers —hyperscalers like Google, Meta, Microsoft and Amazon— saw the shortage coming and secured their supply in advance. HBM is fully allocated by contract through 2027; North American hyperscalers have been negotiating long-term supply agreements since early 2026. That is: the actors with the most resources and the greatest need for memory already have theirs guaranteed, at agreed prices.
The one who did not negotiate a long-term contract is the consumer. And the adjustment falls on them. Gartner estimates the memory price rise will make personal computers 17% more expensive and smartphones 13% more expensive by the end of 2026 compared with 2025. But the effect does not stop at price: it also hits the quantity and quality of what is available.
| Indicator | 2026 forecast (vs 2025) | Source |
|---|---|---|
| Annual DRAM price | +125% | Gartner |
| Annual NAND flash price | +234% | Gartner |
| Final price of a PC | +17% | Gartner |
| Final price of a phone | +13% | Gartner |
| Worldwide PC shipments | −10.4% | Gartner |
| Worldwide phone shipments | −8.4% | Gartner |
That fall in shipments —the steepest contraction in over a decade, according to Gartner— tells a story behind the price: facing pricier chips, manufacturers not only raise tariffs, but delay launches, cut catalogues and downgrade specifications. Counterpoint analysts warn that low-end phones will be the most affected, because the type of memory they use (LPDDR4) is disappearing faster than expected as factories prioritise newer, more profitable standards. The social consequence is regressive: the buyer of a cheap device —typically whoever can least afford it— is the one who most feels the cut.
The arithmetic of the bottleneck
It is worth pausing on why this problem has no quick fix, because that is the structural part. Building a new semiconductor factory —a fab— costs between 20 and 30 billion dollars and takes four to five years to become operational. That means the additional capacity that would ease the shortage will not arrive in volume until 2027 or 2028, at best.
And there is a second factor that prolongs the pain: the asymmetry between how memory prices rise and how they fall. Historically, memory-chip prices fall more slowly than they rise. That implies that, even as new capacity begins to come online in late 2027, price normalisation will lag behind supply. Gartner expects high prices through the end of 2027; Intel’s chief executive, Lip-Bu Tan, went further, stating that “there’s no relief until 2028”.
The bluntest line in Gartner’s analysis sums up the knock-on effect: memflation, said its senior principal analyst Rajeev Rajput, “will destroy, or at least delay, non-AI demand into 2028”. In other words, the priority given to AI not only makes the rest of electronics more expensive: it can come to postpone their very existence, because certain products cease to be viable at those component prices.
The two readings, with comparable weight
The phenomenon admits two legitimate interpretations, and it is worth setting them out without tipping the scale.
On one reading, memflation is the expected —and temporary— cost of a large-scale technological transition. Its defenders argue that every industrial revolution first concentrates resources in the new infrastructure and then redistributes them as capacity expands: the factories being built today will ease the shortage, prices will fall, and the consumer will have borne a transitory surcharge in exchange for an AI infrastructure that, they hold, will generate broad economic value. Under this reading, the market itself corrects the imbalance: high prices are the signal that incentivises investment in new capacity, and Gartner itself stresses that memflation “is profound, but it is not perennial”.
On the contrary reading, the episode reveals a cost transfer that warrants scrutiny. Its defenders point out that the mechanism distributes the benefits and costs unequally: the AI companies and memory manufacturers capture the value —semiconductor revenue will grow 64% in 2026 and exceed 1.3 trillion dollars, with memory revenue nearly tripling— while the ordinary consumer absorbs the surcharge without having participated in the decision or, in many cases, in the benefit. For this position, calling “transitory” a phenomenon that will last until 2028 and hit hardest those who buy the cheapest devices is to minimise a real regressive redistribution. And it raises an open question the analysts themselves leave unanswered: if hyperscalers already have their memory costs baked into their infrastructure, will they pass that cost on to the prices of their cloud services, making everything built on top of them more expensive down the chain?
It is not for this outlet to rule which reading is correct. It is to note the fact both share: the physical infrastructure of artificial intelligence is not free, and its bill is not paid solely by those who use it. It is spread, quietly and through the supply chain, until it reaches the price of the phone someone will buy this year without suspecting that part of that sum is, in fact, the toll of a technology they may never come to touch.
Methodological note. The cover chart shows percentage changes forecast for 2026 versus 2025, not absolute values nor figures comparable with each other on the same magnitude: the first two bars (DRAM +125%, NAND +234%) measure the wholesale price of memory components per Gartner; the last two (PC +17%, phone +13%) measure the final consumer price of the complete device, also per Gartner, and are smaller because memory is only a part of the device’s total cost. The four figures come from Gartner’s public forecasts of February and April 2026. The shipment data (−10.4% PC, −8.4% phones), the production concentration (95% in three manufacturers), data-centre memory consumption (≈70%), the cost and timeline of a factory (20–30 billion USD, 4–5 years) and the observations on DDR4, LPDDR4 and HBM3e come from the cited analyses by Gartner, TrendForce, Counterpoint Research and IDC, as well as public statements gathered in specialist press. The figures are forecasts and may be revised. Data cutoff: 23 May 2026.