Sociogencia · Case 05 · Adam Roberts

The Memory
Supercycle

The supply story is land, power and capex. This is the bill – how the data centre’s hunger for memory is quietly repricing every device you own.

Bottom line

AI data centres now consume roughly 70% of the world’s memory output, and the three firms that make it have swung the bulk of production to the high-bandwidth memory that wraps an AI chip. The conventional memory left for your laptop, phone and console is scarcer and far dearer. Micron just printed an 84.9% gross margin. The cascade that begins in a server hall ends at a checkout. Estimated

cascade_trace.exe · liveSELECT LINK

Who drives it
Who is exposed

Every device gets pricier

Memory is now too large a share of a device’s build cost to absorb. On 25 June 2026 – the day this analysis was filed – Apple, Microsoft and others moved at once. Here is what each move did to the shelf price.

The gistthe whole argument, in six lines
  1. The server hall starves the laptop. Every AI accelerator is wrapped in premium memory, and making it crowds out the cheap kind.
  2. The numbers moved fast. With supply re-pointed to HBM, conventional memory prices have jumped at a pace the industry has not seen.
  3. One income statement tells it. Micron’s blowout quarter is the clearest read on who is capturing the cycle.
  4. The budget device disappears. The cascade reprices not just flagships but who can afford to compute at all.
  5. Memory is a strategic weapon. Only two countries make HBM at scale, which turns scarcity into leverage between states.
  6. Two makers, two strategies. SK Hynix ramps and sells out; Samsung holds back because plain DRAM now pays better.

Why a server hall starves your laptop

The takeawayEvery AI chip needs stacked high-bandwidth memory that sells for ten to twenty times the price of the ordinary kind. Making it eats the capacity that would have gone to your laptop.

Every AI accelerator is wrapped in stacked high-bandwidth memory. It sells for ten to twenty times the price of the memory in your laptop, and the maths is the whole story.

Making it eats the capacity that would have gone to your laptop.

High-bandwidth memory, or HBM, is the specialist stacked DRAM that rings an AI chip such as NVIDIA’s Blackwell or the coming Vera Rubin. Each accelerator needs several stacks, and as training and inference have scaled, demand has turned vertical: TrendForce puts HBM demand growth above 130% year on year in 2026, with delivery lead times stretching from ten weeks to twenty-four. Verified

The incentive is brutal in its simplicity. An HBM module sells for $60–100 against $5–10 for the equivalent DDR5 in a consumer machine – a twelve to twenty times premium – and each gigabyte of it consumes three times the wafer capacity. Every HBM stack a fab makes is several consumer modules it does not. Verified

So the world’s memory has been quietly re-pointed. In 2022, data centres took 20–30% of all memory produced. By 2026 they take about 70%. The same three firms – Micron, SK Hynix and Samsung – have moved the overwhelming majority of combined output to HBM for AI. Estimated

This is the crowd-out at the heart of the cycle. It is not a shortage of factories or a supply shock from a disaster. It is a deliberate, rational reallocation of fixed capacity toward the highest-margin product the industry has ever sold – and consumer memory is what gets displaced.

Interactive · the share

Where the memory goes, charted.

Where the memory goes Data-centre share of global memory output
2022before the buildout
~25%
2024scale-up
~45%
2026the supercycle
~70%
Shares are indicative, drawn from TrendForce and industry commentary; the trajectory, not the decimal, is the point. Graded Estimated.
Interactive · the numbers

The three numbers that tell the story.

130%+
HBM demand growth, year on year
TrendForce · 2026
24 wk
HBM lead time, up from 10 weeks
to Dec 2025
12–20×
HBM price premium over DDR5
$60–100 vs $5–10
Wafer capacity per GB vs DDR5
the crowd-out

The numbers, in full

The takeawayOnce supply re-pointed to HBM, conventional DRAM and NAND prices rose faster than at any point in the modern cycle.

With supply re-pointed, conventional memory prices have moved at a speed the industry has not seen before. Morgan Stanley counted a roughly six-fold rise across a year and gave it a name: chipflation.

Interactive · the numbers

The six-fold surge, charted in full.

The squeeze in numbers Memory price change · period over period
HBM3E per chipearly → Dec 2025
+68%
Contract DRAMQ3 → Q4 2025
+30%
Overall DRAMfull year 2025
+172%
NAND flashQ1 2026, QoQ
+90%
Smartphone DRAMQ1 2026, QoQ
+50%
iPhone 12GB chipDec 2025 → now
+230%
Bars scaled within this set for legibility; figures compiled from TrendForce, Counterpoint and analyst notes. Mixed periods – read the labels. Graded Verified to Estimated.

The iPhone line is the one to sit with. The 12GB LPDDR5X chip in a premium handset went from $25–29 at the start of 2025 to about $70 – a 230% rise on a single component. Memory now runs past 10% of an iPhone Pro Max’s bill-of-materials, and above 20% at the high-memory end. Estimated

Morgan Stanley’s 66-page June 2026 report found memory prices up roughly six times across the year as producers favoured AI silicon, and warned of chipflation spreading from the server hall to the wider economy. Bernstein expects HBM itself to more than double again in 2027. Announced

One caveat for reading these numbers going forward: spot and contract have decoupled. By mid-July 2026 TrendForce data showed DDR5 spot trading gone quiet, in a wide $32 to $65 band, while mainstream contract chips rose only modestly. The spot softness does not yet reach the contract and enterprise pricing this piece tracks, and the two should be read separately from here. Contested

Micron’s 84.9% quarter

The takeawayMicron’s quarter is the single clearest proof that the money is flowing to whoever controls memory supply, not compute.

The clearest read on the cycle is a single income statement. As the only American-headquartered memory maker, Micron sits at the centre of the story – and its most recent quarter is one no memory company has printed before.

Interactive · Micron

The 84.9% quarter, graphed quarter by quarter.

Micron, quarter by quarter Revenue, $bn · gross margin, %
Revenue ($bn)Gross margin (%)
$9.3Yr-prior
Q3 FY25
$23.9Q2 FY26
$41.5Q3 FY26
~$50Q4 FY26
guide
39% 62% 84.9% ~86%
Revenue bars and margin line to scale within the panel. Q4 FY2026 figures are company guidance. Source: Micron IR, June 2026. Graded Verified; guidance Announced.
$41.5B
Q3 FY2026 revenue, up 346% YoY
beat by ~$7bn
84.9%
Gross margin, from 39% a year prior
guiding ~86%
$18.3B
Free cash flow, single-quarter record
Q3 FY2026
$1,000
Bull price target, stock past $600
D.A. Davidson
Interactive · bull vs bear

Weigh both sides of the reallocation call.

The reallocation looks durable while hyperscalers keep buying HBM on long-term contracts, locking conventional buyers out of a shrinking pool. UBS sees HBM revenue at $32.7bn in 2026; Aletheia has memory rising from ~45% of AI hardware value toward 70%-plus by 2027. New fab capacity will not bite until 2027 and beyond: SK Group’s chairman now warns of a wafer shortfall above 20 per cent against a four-to-five-year lead time to add capacity, and Micron is guiding to ~$50bn revenue at ~86% margin – numbers without precedent. Announced

Memory is the most cyclical business in technology, and every boom has ended in a bust. The 39-analyst consensus target sits near $478, far below the $1,000 calls – a vote for cyclicality. And demand destruction is already visible: PS5 unit sales fell 46%, PC and smartphone markets are shrinking to decade lows. As one observer put it, memory has become the new gold – and gold is cyclical too. Contested

The disappearing budget device

The takeawayThe squeeze hits the bottom of the market hardest: the entry-level phone and the school laptop are where the price rise actually bites.

The cascade does not stop at flagships. It is reshaping who can afford to compute at all – and the most striking forecast is about the bottom of the market.

The cascade does not stop at flagships.

Gartner sees PC prices up 17% and smartphone prices up 13% in 2026, with shipments down to their worst in over a decade. Memory’s share of a PC’s build cost rises from 16% to 23% in a single year. Announced

The line that should worry policymakers: Gartner expects the market for sub-$500 PCs to disappear by 2028. France’s INSEE has already logged a 6.34% rise in laptop and tablet prices in the first five months of 2026. Announced

The clearest signal came from the supplier side. In December 2025 Micron stopped selling its Crucial consumer brand of RAM and drives, to focus entirely on AI demand. The world’s only American memory maker walked out of the consumer aisle. Verified

Downstream, the effects compound: upgrade cycles lengthen as people hold devices 15–20% longer, used and refurbished markets boom, and even cloud providers pass memory costs on – OVH has guided to 5–10% increases. Estimated

Interactive · the fallout

What the squeeze does to devices.

−10%
2026 PC shipments, worst in a decade
Gartner / IDC
2028
The year sub-$500 PCs disappear
Gartner forecast
−46%
PS5 unit sales, year on year
demand destruction
Exit
Micron leaves the consumer aisle
Crucial, Dec 2025

HBM as a strategic weapon

The takeawayHBM is made in just two countries, so export controls on it now shape who can build frontier AI at all. Memory has become an instrument of foreign policy.

The same scarcity that reprices your laptop is a lever between states. Only two countries make HBM at scale, and that concentration is now a tool of policy.

hbm_endgame.log · the memory frontMOVE 1 / 6
Washington & alliesvsBeijing

Commercial HBM is made in South Korea (SK Hynix, Samsung) and, in a limited way, Japan (Micron’s Hiroshima fab). For every other nation, that geography is a dependency. Washington’s export controls block China from leading-edge GPUs, advanced HBM, EUV lithography and the packaging tools that go with them. Verified

The result is a roughly three-generation gap: Chinese firms stockpile older HBM2E and push a domestic programme that faces steep equipment barriers. The strategic syllogism is blunt – control HBM supply, and you shape who can build frontier AI at all. Estimated

Industrial policy has mobilised to match: the US CHIPS Act at $52.7bn, the EU Chips Act at €43bn, South Korea’s national strategy pledging $450bn, and China’s “Big Fund” deploying $140bn+ against the controls. Verified

For the markets this report covers, the implication is sharp. Memory bandwidth, not raw compute, is becoming the binding constraint – and nations have started to stockpile HBM the way they once stockpiled oil.

The capital is moving to match. SK Hynix has committed roughly $15bn in 2026 to expand advanced memory focused on HBM, and plans to double wafer capacity into the shortage. Private money is now chasing the exact scarcity this piece describes, which tends to deepen a squeeze before it relieves one. Verified

Two makers, two strategies.

The takeawayThe makers disagree on strategy, but the effect on you is identical: whichever path they take, consumer memory gets squeezed.

The easy story is that Samsung, SK Hynix and Micron all pivoted to HBM in lockstep. The truer one, by late 2025 into 2026, is a split: SK Hynix ramping hard and selling out, Samsung deliberately holding back because conventional DRAM now pays better in a rising market. The crowd-out is real either way, but the reasoning diverges.

Interactive · the faceoff

See how the two makers split the market.

SK Hynix
The split
Samsung
150k to 200k wafers/month on HBM, ramped through the shortage.
HBM output
Held flat at 150k wafers/month, no aggressive add.
2026 HBM capacity fully sold out.
2026 book
Capacity kept for high-end conventional DRAM.
Bet that HBM demand and price justify the ramp.
The wager
Bet that DRAM margins now beat HBM in this price-strengthening cycle.
Squeezes consumer DRAM by reallocating to HBM.
Effect on you
Squeezes consumer DRAM by chasing premium DRAM margin: same shortage, opposite motive.

Source: SK Hynix and Samsung guidance, Nov 2025 to 2026 · Verified

Sources: HBM demand, lead times and capacity reallocation from TrendForce and industry reporting; price-change series compiled from TrendForce, Counterpoint and analyst notes (HBM3E, DRAM, NAND, smartphone DRAM); Micron Q3 FY2026 results and Q4 guidance from Micron Investor Relations and earnings coverage (CNBC, Yahoo Finance, Investing.com), 24 June 2026; chipflation framing from Morgan Stanley’s June 2026 report; device pricing from Apple, Microsoft, Sony and Nintendo statements and Reuters, CNBC, TechCrunch and Business Insider coverage of 25 June 2026; market and shipment forecasts from Gartner and IDC; budget-PC and CPI points from Gartner and INSEE; geopolitics and industrial-policy figures from public policy records. Claims are graded Verified, Announced, Contested or Estimated; price bars are scaled for legibility and projections are framed as such. Analysis and framing are the Entelligencia desk’s; “Adam Roberts” is a pen name of the Entelligencia desk.

The read

The cascade ends at a checkout

Adam Roberts
Entelligencia · Adam Roberts

For a house that reads where compute lands, the memory supercycle is the cleanest proof of a thesis we keep returning to: the data centre is no longer a thing apart. Its appetite now sets the price of a child’s first laptop.

Three implications travel directly into the markets we cover. First, memory bandwidth, not compute, is the constraint that now governs an AI build, and it is a constraint a country cannot quickly manufacture its way out of. Second, the top five US AI firms are projected to spend around $700bn on data-centre infrastructure in 2026, nearly double the prior year; that capex is what is bidding memory away from consumers in the first place.

Third, and most consequential for the ten markets in this report, emerging-market AI projects face an asymmetric access risk. With HBM locked up by hyperscaler contracts, a data centre in São Paulo or Lagos competes for the same scarce, premium-priced silicon as one in Virginia, without the same balance sheet. New fab capacity will not meaningfully ease this until 2027 at the earliest, with some analysts pointing to 2028 and beyond.

Until then the cycle is self-reinforcing: the more compute the world builds, the dearer every other memory-bearing device becomes. The boring middle will report the price rises one device at a time. The story is that they are all the same story.

In conclusion
Where that compute actually lands

The memory squeeze decides what a build costs. The next study decides where the build goes, and which tier captures the value once it arrives.

Read The Continuum →
The Next Hotspot · dropping
Entelligencia

One study down. The map is next.

The Memory Supercycle is a Sociogencia case inside The Next Hotspot, an interactive read on where the world actually builds AI infrastructure. Chapters land through the summer, the full edition on 4 August, and new original analysis is already going live.

Live now · Sociogencia Original analysis is dropping already. Read the studies. Read the studies →