Last updated: June 2026
Key Takeaways
- The viral claim is essentially true. Analysts at IDC, reported by Tom's Hardware and the Wall Street Journal, forecast that AI data centers will consume roughly 70 percent of the world's memory output in 2026, up from 20 to 30 percent in 2022. The alarming number survives scrutiny.
- The squeeze hits everything, not just RAM. Three companies make about 90 percent of the world's DRAM and have redirected it toward AI, pushing up prices on memory kits, SSDs, graphics cards, prebuilt PCs, phones, and consoles alike.
- There is a real lawsuit alleging a price-fixing "cartel," and these same firms were convicted of exactly that twenty years ago. But the new allegations are unproven, and a market this concentrated can keep prices high without an illegal agreement. Either way, relief is not expected before late 2027 or 2028.
A post claiming that AI data centers will use 70 percent of the world's memory chips in 2026 collected millions of views in late June, and the replies were bleak. Say goodbye to affordable gaming. Say goodbye to home computers, to consoles, to indie game development, unless someone breaks up the "RAM cartels."
It is worth slowing down on a viral statistic before repeating it. So this piece answers three questions in order: is the 70 percent figure real, is this actually a cartel, and what should you do about the prices you are seeing right now. The short answers are, roughly, yes, nobody outside the companies knows yet, and not what you would hope.
Yes, the number is basically real
The 70 percent figure traces to the research firm IDC, reported by Tom's Hardware and echoed by the Wall Street Journal and others: data centers are forecast to absorb about 70 percent of all memory produced worldwide in 2026. As recently as 2022, that share sat between 20 and 30 percent. The reversal has been fast and structural.
One honest caveat belongs on that statistic. The 70 percent is a share of memory output. TrendForce frames the same dynamic differently, estimating that AI will consume closer to 20 percent of DRAM wafer capacity, because the high-bandwidth memory that AI accelerators use eats roughly four times the wafer area per gigabyte of ordinary DRAM. Both figures can be true; they measure different things. What no analysis disputes is the direction and the severity. Memory makers have tilted hard toward AI, and consumer supply is now whatever is left over.
How AI ate the memory
Three companies, Samsung, SK Hynix, and Micron, manufacture roughly 90 to 95 percent of the world's DRAM. Through 2025 and into 2026 they redirected wafer capacity toward high-bandwidth memory and high-capacity server DDR5 built for AI data centers, both of which carry far richer margins than consumer modules. Every wafer that becomes an HBM stack for an AI accelerator is a wafer that does not become a stick of desktop RAM. The market is close to zero-sum.
The shortage stuck because supply cannot turn on a dime. After the memory downturn of 2022 and 2023, when makers cut output and skipped new investment, the industry was poorly positioned for the AI demand that arrived in 2025. IDC puts 2026 DRAM supply growth at about 16 percent and NAND at 17 percent, both below historical norms, and new fabrication plants from Micron and SK Hynix are not expected to reach volume production until 2027 at the earliest.
The clearest signal of the new priorities arrived in late 2025, when Micron announced it was shutting down its consumer Crucial brand to serve the AI market instead. Memory, historically the most volatile commodity in technology, has started to behave like a contracted infrastructure product, with multiyear supply agreements that used to be almost unheard of in the industry. Hyperscalers such as Google, Microsoft, Meta, and Amazon lock in those long-term contracts at premium but stable prices, which leaves the consumer retail channel to absorb the volatility.
What it means for the gear you buy
The effect is not confined to memory sticks. Because DRAM and NAND sit inside nearly everything, the squeeze shows up across the entire shopping list. By one market tracker's count, a 64GB DDR5 kit that cost around 195 dollars in 2025 reached 788 dollars in some markets during the worst of the spike. Here is how the pressure lands component by component.
| Component | What happened | Why | What to do |
|---|---|---|---|
| System RAM (DDR5/DDR4) | Consumer RAM rose roughly 110 percent in Q1 2026 | The same DRAM lines diverted to AI server memory | Buy only what you need; do not over-spec to future-proof |
| SSDs (NAND flash) | Client SSD prices up over 40 percent; NAND up around 147 percent in Q1 | NAND output cut while makers chased HBM and DRAM | A real storage need is worth filling now; capacity will not get cheaper soon |
| Graphics cards (GDDR) | Phased hikes from AMD (January) and Nvidia (February); RTX 50 Super refresh slipped to Q3 | GDDR6 and GDDR7 contracts expired into a tight spot market; memory is now most of a high-end card's cost | The used market is the relief valve; treat "5,000 dollar GPU" claims as unconfirmed projections |
| Prebuilt PCs and laptops | Dell announced increases up to 30 percent on some systems; ASUS raised prices in January | Component costs passed through to finished machines | A complete system dilutes the markup versus buying parts separately |
| Phones | Apple raised prices in late June; IDC expects smartphone sales to fall up to 5 percent | Mobile DRAM and NAND compete for the same capacity | Hold serviceable devices longer if you can |
| Consoles | Valve's Steam Machine was delayed; Bloomberg reports the Switch 2 and PS5 could rise, and Xbox hardware already has | Consoles use the same memory and storage | Treat specific console hikes as reported but not final |
Processors are the exception. CPUs ride more mature manufacturing and are not the AI bottleneck, so they have been comparatively insulated from the squeeze.
Shortage, or cartel?
When prices climb roughly 700 percent over four years and the suppliers number three, "cartel" is a natural suspicion, and it is no longer only a forum theory. On June 25, seventeen plaintiffs sued Samsung, SK Hynix, and Micron in federal court in California, accusing the three under the Sherman Act of coordinating to restrict supply and inflate prices. The complaint's sharpest claim is that the makers used the shift to HBM as cover to wind down cheaper DDR4 and DDR5 production, and it singles out Micron shutting Crucial at what it calls the most profitable price point in the company's history as a move that defies ordinary business logic. Apple's recent price increases are cited as the downstream trigger.
This is not a hypothetical pattern for these companies. Samsung and SK Hynix both pleaded guilty to criminal DRAM price-fixing in a US Justice Department case covering 2002 to 2006, paying a combined total north of 700 million dollars in fines, with one Samsung executive sentenced to eight months in prison. European regulators later fined nine memory makers a total of 331 million euros over the same cartel. A separate 2018 class action over later conduct failed.
Here is the line worth holding. What is established: three companies control the market, they reallocated capacity to AI, prices spiked at record speed, and two of them carry prior criminal convictions for fixing these exact prices. What is alleged but unproven: that the current increases are the product of illegal coordination rather than independent decisions. The companies say they are operating independently, and as one executive's widely quoted line about the market put it, nobody wants to be the first to lower prices.
That last point matters more than it sounds. A market with three players does not need a secret meeting to behave like a cartel. When each producer can see that cutting prices would only start a war that helps no one, restraint becomes the rational individual choice, and from the outside the result looks identical to collusion. Economists call this conscious parallelism, and it is generally legal. The Crucial shutdown is the genuinely hard fact for the defense to explain, since exiting a market at peak prices runs against the textbook, and the courts will weigh it. But the takeaway for a buyer holds either way: a handful of suppliers facing irresistible incentives to chase AI margins is enough to keep your prices high, lawsuit or no lawsuit.
When it ends, and what to do until then
The consensus is that this is not a dip to wait out. Analysts at Jefferies expect DRAM prices to keep rising through the second half of 2026, and most forecasts place the earliest meaningful relief in late 2027 or 2028, when new fabrication capacity ramps. Even then, the likely result is a higher new normal rather than a return to 2023 prices. Two wildcards could ease things sooner, a pullback in AI spending or memory-efficiency advances in AI software that reduce demand, though those efficiency claims are contested. For the RAM-specific timing and the build math, our companion piece on when RAM prices will come down goes deeper, and the same forces are already raising router and networking prices too.
So what do you actually do. First, do not panic-buy or speculate, because buying memory you do not need at peak prices is how the shortage wins twice. But since waiting carries no near-term payoff, a genuine need is better filled sooner than later. A few principles hold up: buy complete systems where the memory markup is diluted across a finished machine rather than paying spot prices for a discrete kit; lean on the used market, especially for graphics cards; resist over-speccing for models or workloads you do not actually run; and avoid the gray-market and counterfeit modules that always surface in a shortage by sticking to established brands from authorized sellers.
For local AI specifically, which is the relevant case for many readers here, a complete mini PC with the memory already onboard sidesteps buying a discrete kit at today's worst prices and gives you a working, low-power machine for small models.
Check Price on Amazon: Beelink EQR6 (32GB)
If your constraint is model storage rather than compute, a single high-capacity SSD is the one purchase that is hard to regret, since model files only grow and downloading them repeatedly wastes the bandwidth you are paying for.
Check Price on Amazon: Samsung 990 EVO Plus 2TB
For the full range of options at every budget, our guides to the best hardware for running local AI and the best mini PCs for local AI lay out the trade-offs in detail.
The squeeze underneath the squeeze
The deeper pattern is worth naming. The same AI build-out that has put the most capable models behind government and corporate gates is also making the hardware of self-reliance more expensive. The companies that own the infrastructure capture both ends at once: they shape what you are allowed to run in the cloud, and their demand sets what you pay to run anything at home. Local-first computing is still the strongest hedge against that dependence, but the hedge now carries a price it did not a year ago. Whoever controls the infrastructure controls the experience, and in 2026 that control reaches all the way down to the memory module.
Frequently Asked Questions
Is AI really using 70 percent of the world's memory?
Essentially yes. IDC forecasts data centers consuming about 70 percent of memory output in 2026, up from 20 to 30 percent in 2022. A stricter wafer-capacity measure from TrendForce puts AI's share nearer 20 percent, but every analysis agrees the reallocation toward AI is large and is driving consumer prices up.
Is the RAM shortage a scam or a cartel?
A class action filed in June 2026 alleges illegal price-fixing, and the same makers were convicted of it in the 2000s. But the new allegations are unproven, and a three-company market can keep prices high through independent, legal restraint. The honest answer is that nobody outside the companies knows yet.
When will memory prices come down?
Not soon. Most forecasts point to late 2027 or 2028 at the earliest, when new fabrication capacity ramps, and even then to a higher baseline than 2023 rather than a full reset.
Should I buy RAM, an SSD, or a GPU now, or wait?
If you have a real need, sooner generally beats later, because prices are projected to keep climbing into 2027. If you do not need it, do not buy at peak prices just to future-proof. Used parts and complete systems offer the best value right now.
Why did Micron shut down Crucial?
Micron exited its consumer Crucial brand in late 2025 to focus on higher-margin AI memory. It is the clearest example of makers prioritizing data centers over consumers, and the price-fixing lawsuit cites it as evidence, since exiting a market at peak prices is unusual.
Does this affect phones and consoles too?
Yes. The same DRAM and NAND go into phones, tablets, and consoles. Apple raised prices in June, IDC expects smartphone sales to dip, Valve delayed its Steam Machine, and reports suggest the Switch 2 and PS5 could see increases.
