Sold Out Until 2027: How AI Bought All the Memory
Reading time: 8 minutes
TL;DR
- Three companies (Samsung, SK Hynix, Micron) control over 95% of global DRAM production, and all three have shifted capacity toward high bandwidth memory (HBM) for AI accelerators
- The math is brutal: one wafer of HBM displaces roughly three wafers’ worth of the ordinary memory that goes into phones, laptops, and cars
- Data centers are forecast to consume about 70% of all memory chips produced in 2026, up from roughly 20 to 30% as recently as 2022
- Consumer DRAM prices have risen 300 to 600% from their 2024 to 2025 lows; a 32GB DDR5 kit that cost around $90 in mid 2025 listed at $375 in June 2026
- On June 25, 2026, Apple raised prices globally on every Mac, iPad, HomePod, Apple TV, and the Vision Pro. The MacBook Pro rose $300 and the Mac Studio rose $1,300.
- Micron’s CEO expects the shortage to last through 2027, with supply only gradually improving in 2028. Some analysts say 2030
- New fabs are coming, but meaningful new output does not arrive before 2027, and most of it is already earmarked for AI memory anyway
1. This Is Not a Shortage. It Is a Reallocation.
The last great chip crisis, from 2020 to 2023, was an accident. A pandemic broke supply chains, automakers canceled orders at the wrong moment, and everyone spent two years untangling the mess.
What is happening in 2026 is not an accident. It’s a decision.
The memory industry is effectively three companies: Samsung, SK Hynix, and Micron, which together control more than 95% of the world’s DRAM output. Starting in 2024 and accelerating through 2025, all three deliberately pivoted their limited fab capacity away from the commodity memory that goes into consumer devices and toward high bandwidth memory, the specialized stacked DRAM that feeds AI accelerators.
The pivot is a rational business decision. HBM sells at multiples of commodity DRAM pricing and comes with massive guaranteed purchase orders. Micron disclosed that its entire HBM output for calendar 2026 was sold out. SK Hynix said in October 2025 that it had already sold its full 2026 production capacity, and not just HBM, but DRAM and NAND too. Hyperscalers including Google, Amazon, Microsoft, and Meta reportedly placed open ended orders, telling suppliers they would take as much as could be made, at whatever price.
When the entire output of an oligopoly is pre-sold to the richest buyers in history, everyone else shops from the leftovers. In 2026, the leftovers are thin.
2. The Wafer Math
Here is the constraint, stripped down to its physics.
A fab produces a fixed number of silicon wafers per month. Fabs take years and tens of billions of dollars to build, so in the short run, wafer supply is fixed. Every product the memory industry makes is a claim on those same wafers.
HBM is a premium product in performance and a terrible one in wafer efficiency. Because it stacks dies and uses larger die sizes with more complex processing, a single wafer yields roughly one third the memory bits as the same wafer running commodity DRAM. Some estimates put HBM’s wafer area cost at about four times per gigabyte.
The Wafer Trade, Visualized
ONE FAB. ONE MONTH. FIXED WAFER SUPPLY.
─────────────────────────────────────────────
Path A: Commodity DRAM Path B: HBM for AI
┌─────────────────────┐ ┌─────────────────────┐
│ ███ ███ ███ ███ ███ │ │ ███ │
│ ███ ███ ███ ███ ███ │ │ (1x bits, │
│ ███ ███ ███ ███ ███ │ │ 3-5x margin) │
│ (3x memory bits) │ │ │
└─────────────────────┘ └─────────────────────┘
Phones, PCs, cars, AI accelerators,
routers, consoles pre-sold to 2027+
Every wafer moved from A to B removes ~3 units of
consumer memory from the market. The fab chooses B.
Multiply that choice across every leading edge line at three companies and the outcome is math, not mystery. IDC projects 2026 DRAM supply growth of just 16% year over year, well below the historical norm of 20 to 30%, at the exact moment demand is exploding. Data centers, which consumed roughly 20 to 30% of global memory as recently as 2022, are forecast to take about 70% of everything produced in 2026.
One project alone illustrates the scale: OpenAI’s Stargate buildout, backed by supply agreements with Samsung and SK Hynix, has been reported to require up to 40% of global DRAM output, on the order of 900,000 wafers per month.
3. The Price Ledger
The numbers coming out of this squeeze read like typos.
| What | Then | Now (mid 2026) | Change |
|---|---|---|---|
| 32GB DDR5 kit (US retail) | $80 to $120 (mid 2025) | $375 (June 3, 2026) | +300 to 600% from lows |
| 16GB DDR4 module | $137 (Q1 2026) | $207 (Q2 2026) | +51% in one quarter |
| LPDDR5X 12GB (phone memory) | $77 (Q1 2026) | $146 (Q2 2026) | +89% in one quarter |
| DRAM overall, 2025 | baseline | reported +172% through the year | structural |
| Memory share of a PC’s build cost | 15 to 18% | 35% (HP, Q1 2026 earnings) | roughly doubled |
The consumer fallout has moved from spreadsheets to store shelves. Micron exited its Crucial consumer memory brand entirely. Retailers in Tokyo’s Akihabara district rationed DDR5 purchases to stop hoarding. Nvidia reportedly cut RTX 50 series gaming GPU production by 30 to 40% because the GDDR7 memory those cards need is being outbid by data centers. And then there is Apple, which earns its own section, because what happened on June 25 is the clearest signal yet of how deep this goes.
Even the language of the market changed. Reports in early 2026 described DRAM moving to something close to hourly pricing, with supplier quotes valid for days instead of quarters.
4. June 25: The Day Even Apple Gave In
If any company was built to ride out a component shortage, it was Apple. Long term supply contracts, unmatched purchasing power, and margins deep enough to absorb almost anything. Apple had reportedly locked in DRAM agreements that shielded it into early 2026, and for months it held the line on pricing while the rest of the industry hiked.
On June 25, it stopped holding. Apple raised prices globally on every Mac, every iPad, its home devices, and the Vision Pro, effective immediately on its online store. In its statement, the company said it had “reached a point where we need to begin raising prices” after shielding customers for as long as it could.
The New Stickers (US pricing, June 25, 2026)
| Product | Before | After | Increase |
|---|---|---|---|
| MacBook Neo (launched at $599 in March) | $599 | $699 | +17% |
| MacBook Air | $1,099 | $1,299 | +18% |
| MacBook Pro (base) | $1,699 | $1,999 | +18% |
| iPad (base) | $349 | $449 | +29% |
| iPad Air | $599 | $749 | +25% |
| iPad Pro | $999 | $1,199 | +20% |
| iPad Mini | $499 | $599 | +20% |
| Apple TV | $129 | $199 | +54% |
| HomePod | $299 | $349 | +17% |
| Apple Vision Pro | $3,499 | $3,699 | +6% |
| Mac Studio (M3 Ultra) | $3,999 | $5,299 | +33% |
Many of the most popular models rose 20% or more overnight. The iPhone, Apple Watch, and AirPods were spared for now, though Apple hinted that further adjustments may come, and Counterpoint Research estimates rising component costs could add roughly $200 of cost per iPhone.
The market read the signal instantly: Apple shares fell more than 6%, the stock’s worst single day since April 2025. And the warning signs had been accumulating for weeks. In May, Apple quietly discontinued the $599 base Mac mini configuration, and high memory Mac configurations became hard to find in its store.
Here is why this matters beyond Cupertino. Apple is the canary with the best air supply in the mine. If the company with the strongest supply chain leverage on Earth can no longer shield its customers from the memory squeeze, nobody downstream of a DRAM contract can. Microsoft proved the point within days, announcing Xbox price increases of $100 to $150 starting in August, citing the same component costs.
5. Who Actually Pays
Follow the memory and you find the tax spreading through the whole economy.
Phone buyers, especially at the low end. Memory is 15 to 20% of a mid range phone’s bill of materials, versus 10 to 15% for a flagship. The cheaper the device, the less room to absorb a doubling of memory price. IDC warns of 10 to 20% device price increases by the end of 2026 and projects smartphone shipments could fall by double digits.
PC buyers, at the worst possible time. The shortage collided with the Windows 10 end of life refresh cycle. Lenovo, Dell, HP, Acer, and ASUS have all warned of 15 to 20% price hikes and reset contracts.
Drivers. A modern car carries up to 3,000 chips, and automakers hold less than 10% of the DRAM market, which means almost no bargaining power against hyperscalers. Tesla’s CEO described the company’s position in January 2026 as a choice to “hit the chip wall or make a fab.” Expect memory costs to quietly push automakers toward higher margin trims and away from entry level models.
Anyone buying anything with a plug. Routers, consoles, TVs, appliances, medical devices. Standard DRAM is the working memory of the physical world, and the physical world is now bidding against GPT training runs for it.
6. Why This Doesn’t Fix Itself Quickly
The classic memory cycle goes: prices spike, producers add capacity, glut follows, prices crash. Three things break that script this time.
First, fab lead times. Micron has announced well over $150 billion in investments across Idaho, New York, Virginia, and Taiwan, and SK Hynix is building a $15 billion advanced packaging plant in Indiana. Almost none of it produces meaningful output before 2027, and several facilities will not reach mass production until 2028.
Second, the new capacity is spoken for. Most incremental supply is earmarked for HBM and server grade memory, because that is where the margin is. New fabs relieve the AI bottleneck first and the consumer bottleneck last.
Third, demand is structural, not cyclical. AI inference workloads grow with every deployed application, and each new accelerator generation carries 20 to 30% more memory than the last. The producers have no incentive to break the sellers’ market they finally, after decades of boom and bust, get to enjoy. Lenovo has told customers to expect elevated memory prices to be the norm until at least 2030. Micron’s CEO said in June 2026 that shortage conditions persist through 2027 with gradual improvement in 2028. AMD said at Computex that DDR5 pricing does not normalize before 2028.
The honest planning assumption for anyone budgeting hardware: current prices through at least the end of 2026, relief measured in years, not quarters.
7. What Would Actually Change the Curve
Three levers exist, and only one is fully in the industry’s control.
An AI capex pullback would free capacity fastest. If hyperscaler spending, projected to exceed $600 billion in 2026 across the top cloud providers, decelerates hard, wafers flow back to commodity memory within quarters. That is a bet on AI disappointment, not a plan.
Architectural workarounds could reduce HBM dependency. There is early industry discussion of AI systems built on cheaper commodity DRAM pools instead of maximum HBM per accelerator. If it works at scale, it eases the squeeze from the demand side.
And new capacity eventually lands. The 2027 to 2028 fab wave is real. The question is only how much of it consumer memory ever sees.
The Bottom Line
The AI industry is taxing the entire electronics economy to fund its own buildout, and the tax is collected at the wafer. Three companies control the supply, the richest buyers in history have pre-purchased the output through 2027, and every HBM stack that ships subtracts roughly three units of ordinary memory from the world.
This is what a hard constraint looks like in practice. Not a dramatic failure, just a quiet repricing of everything with a chip in it, set in motion by capacity decisions made years before you noticed your laptop got expensive. The software wants to scale exponentially. The wafers do not care, yet. The honest counterweight: the 2027 to 2028 fab wave is real, the architectural workarounds are being engineered right now, and the same forces that created this squeeze are funding the largest memory buildout in the industry’s history. The constraint is genuine. So is the response.
Sources
- Micron Technology, fiscal Q1 2026 earnings call (HBM capacity sold out through calendar 2026); CEO Sanjay Mehrotra remarks, June 2026, on shortage duration
- Everstream Analytics, “Global Memory Chip Shortage Worsens,” January 2026 (SK Hynix 2026 capacity sold out; DRAM inventory drawdown)
- IDC, “Global Memory Shortage Crisis,” February 2026 (95%+ DRAM concentration; 16% supply growth; BOM shares; device price and shipment forecasts)
- Tom’s Hardware daily RAM price tracker, June 3, 2026 (32GB DDR5 kit pricing); Tom’s Hardware reporting on hourly DRAM pricing, March 2026
- SigmaIntel Q2 2026 consumer memory pricing data, via TweakTown and Wccftech, June 2026
- HP Inc., Q1 2026 earnings call (memory as share of PC bill of materials)
- Fortune, “AI’s memory chip shortage is quietly taxing the entire economy,” March 19, 2026 (wafer displacement economics)
- TrendForce, “Memory Wall” insights, January 2026 (DRAM price projections; HBM demand growth)
- Reuters reporting on hyperscaler open ended memory orders and OpenAI Stargate supply agreements with Samsung and SK Hynix, October 2025
- IDC via Tom’s Hardware (data centers consuming ~70% of 2026 memory production)
- Bloomberg (Mark Gurman), “Apple Hikes Mac, iPad Prices on Memory Shortage; Shares Fall,” June 25, 2026
- CNN Business and Al Jazeera, June 25 to 26, 2026 (model by model Apple pricing, including the Mac Studio increase and 20%+ hikes on popular models)
- CNBC, June 25, 2026 (Counterpoint Research iPhone component cost estimate; May Mac mini configuration cut; worst trading day since April 2025)
- Company statements and reporting: Micron Crucial brand exit, Nvidia RTX 50 production cuts (December 2025), Microsoft Xbox price increases (August 2026), Lenovo and AMD pricing outlooks