MR. Memory, Explained

The Global Memory Shortage

The world is running low on memory.
Here's the whole picture.

Every phone, car, game console, and AI data center runs on memory chips — and demand is racing ahead of what the world can build. This is the five-minute tour of who makes it, who needs it, and why the squeeze isn't ending soon.

Start the tour → ↓ scroll
The short version

01 — The Basics

What is memory, really?

Think of any computer — your phone, a car, an AI server — as someone working at a desk.

Memory is the desk itself: the space where you spread out whatever you're working on right now. It's blazing fast and instantly reachable, but small — and it's wiped clean the moment the power goes off. Storage is the filing cabinet behind the desk: huge and permanent, but slower to dig through. Every device constantly shuttles things between the two.

Processor CPU / GPU Memory DRAM — the desk fast · small · temporary Storage NAND / SSD — the cabinet slower · huge · permanent instant load / save The processor lives or dies by how fast memory can feed it.
Memory is the working surface; storage is the archive. This guide is about the working memory — the part the world is running short of.
Working memory

DRAM

The desk. Holds whatever the chip is actively crunching, and feeds it at enormous speed.

  • SpeedExtremely fast
  • SizeGigabytes
  • Keeps data without power?No
Long-term storage

NAND flash

The filing cabinet. Keeps your photos, apps and files even when the device is off.

  • SpeedSlower
  • SizeTerabytes
  • Keeps data without power?Yes

The squeeze everyone's talking about is mostly about the desk — DRAM. As AI, cars and robots all demand bigger, faster desks at once, the world can't build them quickly enough. The rest of this guide unpacks why.

02 — The Players

Who actually makes the world's memory?

Almost all of it comes from a handful of companies — and just three make nearly every DRAM chip on Earth.

Memory is brutally hard and capital-intensive to manufacture, so the industry has consolidated into a tight oligopoly. In DRAM — the "desk" memory from section 1 — three names control roughly 90% of the market.

DRAM revenue share

Q1 2026
Samsung 38.6% SK hynix 28.8% Micron 22.4% Others 10.2%

NAND flash share

Q1 2026
Samsung 29% SK hynix 18% Kioxia 14% Micron 13% SanDisk/WD 13% YMTC 13%

HBM share

2026
SK hynix 57% Samsung + Micron 43%
~90% of DRAM is made by Samsung, SK hynix and Micron combined. There is no fourth option at scale.

DRAM & HBM: Counterpoint, Omdia via TechTimes; NAND: Counterpoint. Shares vary slightly by firm and quarter.

The hierarchy flips for AI memory

The ranking above is for ordinary DRAM. For HBM — the ultra-fast stacked memory that AI accelerators depend on — the order changes. SK hynix leads with about 57%, having moved first and locked up supply to Nvidia. HBM is so profitable it pushed SK hynix to a roughly 72% operating margin in Q1 2026 — several times the margins on commodity memory. That's why all three makers are racing to shift capacity toward it (more on what that does to everyone else in section 5).

$440B+ projected total memory market in 2026.
$54.6B HBM market in 2026 — up about 58% in a single year (BofA estimate).

HBM market & margins: SK hynix outlook, CNBC, Silicon Analysts.

03 — Demand

Who's buying it all?

Everything that's getting smarter needs more memory — and it all got smarter at once.

~70%
of the world's high-end DRAM in 2026 is going to AI data centers. They sit at the front of the line; everyone else waits.

Where the world's memory goes

Memory made Samsung·SK hynix·Micron AI data centers ~70% Everything else phones·PCs·cars·gaming ~30%

The same factories feed both — so when AI takes the lion's share, everything else competes for what's left.

How much working memory one device holds — to scale:
Smartphone ~12 GB Laptop ~16 GB Game console ~16 GB Self-driving car ~300 GB AI server node ~2 TB
One AI server node can hold as much working memory as roughly 160 phones — and data centers run them by the million. That's the demand curve in one picture. (Approximate, typical configurations.)

AI data centers

The dominant driver · hyperscaler capex +60% in 2026

Every AI model lives in memory while it runs. Training and serving them devours HBM and high-capacity server DRAM faster than fabs can supply it — and cloud giants are spending more than ever to grab it.

Autonomous & electric cars

16 GB → 300 GB+ per car

Today's car uses about 16 GB of DRAM. A Level-4 self-driving car needs 300 GB+ to fuse cameras, radar and AI in real time. Micron calls automotive a "20-year growth vector."

Robotics

~300 GB per humanoid robot

A humanoid robot needs compute on par with a self-driving car — roughly ~300 GB of DRAM plus heavy storage. Every robot that ships is another data-center-class memory customer.

Gaming

Memory ≈ 20% of a PS5's bill of materials

Consoles and GPUs lean on fast graphics memory — PS5 uses GDDR6, Switch 2 uses LPDDR5. With makers prioritizing servers, gaming hardware is getting pricier and harder to find.

Phones & PCs

PC prices +15–20% in early 2026

The everyday squeeze. As memory is diverted to AI, PC and phone makers eat higher costs — IDC expects 2026 PC shipments down 11.3% and smartphones down 12.9%.

Today's car ~16 GB Robotaxi (L4) 300 GB+
One self-driving car can hold as much working memory as a small rack of servers — and the world wants millions of them.

What it means for you. When data centers buy most of the memory, the squeeze reaches your shopping cart. PC makers have flagged price rises of +15–20%, and 2026 PC and smartphone shipments are expected to fall 11.3% and 12.9%. Translation: pricier laptops, phones, consoles and graphics cards — and thinner stock when you go to buy.

Data-center share & AI demand: CNBC, IEEE Spectrum; auto/robotics: Tom's Hardware, TrendForce; gaming: Game Developer; PCs/phones: IDC.

04 — The Toolkit

Not all memory is the same.

"Memory" is really a toolkit. Each type trades off speed, cost and power for a different job. Tap one to explore it.

Specs & use cases: IntuitionLabs, Rambus, Semiconductor Engineering. Bandwidth bar is relative to the fastest (HBM4, ~2 TB/s).

05 — The Squeeze

Why the shortage is real — and lasting.

Here's the part that surprises people: even if the makers build more, relief is years away. Three forces lock it in.

23% of all DRAM wafers go to HBM in 2026 (up from 19% in 2025) — and HBM needs ~3× the silicon per gigabyte. Every wafer moved to AI removes 3× as much ordinary memory.
2–3 yrs to build a new $15–20B DRAM fab. Micron's new HBM plant breaks ground in 2026 but won't ship a chip until 2028.
9% projected HBM shortfall in 2027 — widening from 5% in 2025. The gap gets worse before new supply arrives.

Why HBM eats so much silicon

8–12 dies logic die + stacked DRAM = one HBM cube
HBM stacks 8–12 DRAM dies on a logic die — which is why it needs roughly 3× the silicon per gigabyte of ordinary memory, and why moving wafers to HBM drains so much capacity.
HBM supply shortfall (% of demand) 5% 2025 6% 2026 9% 2027 new fabs first output 2028
The deficit keeps growing through 2027 — the first chips from newly-built fabs don't arrive until 2028.
Relief watch First new-fab output expected ~2028.

Why this cycle is different. In past memory crunches, makers added capacity and the shortage cleared in 12–18 months. This time demand is structural — AI, cars and robots all scaling at once — and the only real fix, new fabs, takes years. New capacity won't reach volume until 2027 and beyond.

The result is already in prices: DRAM is up roughly 172% year-over-year by late 2025 and has more than doubled since early 2025. More supply is coming — just not soon, and not before demand climbs again.

DRAM contract-price index (early 2024 = 100)

100 (early 2024) Q1 '24 Q1 '26 ≈430
Approximate index derived from the cited price moves — DRAM has gone from a baseline of 100 to roughly 430 (≈4.3×) in about two years.

Wafer mix & shortfall: Tech-Insider, Tom's Hardware, TrendForce; fabs & prices: IDC.

On the radar

What to watch next

Moments that could move the memory market in the weeks ahead.

7
days
EARNINGS
Samsung Q2 2026 preliminary earnings guidance
Released during Korean market hours — first read on the memory recovery and HBM mix.
2026-07-07
15
days
PRICING
TrendForce Q3 DRAM & NAND contract price update
Quarterly contract-price reset; sets the tone for spot moves.
2026-07-15
24
days
EARNINGS
SK Hynix Q2 2026 earnings call
HBM3E/HBM4 demand and pricing commentary from the HBM leader.
2026-07-24
28
days
SUPPLY
Samsung Q2 2026 full results & capex detail
Watch fab capex guidance and HBM qualification progress.
2026-07-28

Editorial calendar — dates may shift. Deeper coverage in the daily briefs.

FAQ

Common questions

Why is computer memory so expensive right now?

AI data centers are buying roughly 70% of the world's high-end DRAM, and memory makers have shifted capacity toward high-margin HBM for AI chips. That leaves less for everyone else — DRAM prices have more than doubled since early 2025 (up about 172% year over year).

When will the memory shortage end?

Not soon. A new memory fab costs $15–20B and takes 2–3 yrs to build, so plants breaking ground now won't ship chips until 2028. The projected HBM shortfall actually widens through 2027 before new supply catches up.

Will my next phone or PC cost more?

Likely yes. PC makers have signaled price rises of +15–20% in early 2026, and 2026 PC and smartphone shipments are expected to fall 11.3% and 12.9% as memory costs climb.

What is HBM, and why does it matter?

HBM (High Bandwidth Memory) is ultra-fast DRAM stacked right next to an AI chip. It's the bottleneck for AI: SK hynix leads with about 57% share, and HBM now consumes 23% of all DRAM wafers while using roughly 3x the silicon per gigabyte — which is a big reason ordinary memory is scarce.

Who makes the world's memory?

Just three companies — Samsung, SK hynix and Micron — make about 90% of all DRAM. That concentration is part of why supply can't flex quickly when demand jumps.

The takeaway

The squeeze is structural — and it's just getting started.

A handful of makers, a wave of demand from AI, cars and robots, and fabs that take years to build. Put together, the world will be short on memory for years — even as more gets made. Now you can see every piece of the puzzle.

Read the daily briefs →

Figures here are compiled from public industry sources — Counterpoint, TrendForce, Omdia, IDC and company filings — cited inline, and reviewed before they're published. The daily briefs are generated from live sources, each listed at the foot of the brief. 2025–2026 data.

Market figures as of Q1 2026 · last updated June 2026