Why Is RAM So Expensive Now? The AI Shortage Explained

Transcript

How a $100 RAM Kit Became a $400 RAM Kit

NovaWelcome to YourTalks.

NovaA year ago you could build a solid gaming PC and drop $100 on 32 gigabytes of DDR5. That same kit today runs $379 to $500, depending on speed and brand. And that's on a good day, when it's even in stock. J.P. Morgan says DRAM prices will have risen more than 400% from the start of 2024 to the end of this year. Tom's Hardware called it an AI-induced RAM apocalypse, and they're not wrong.

AtlasWhat happened is straightforward. AI data centers created a demand shock that the three companies who make nearly all the world's DRAM — Samsung, SK Hynix, and Micron — responded to rationally. They shifted production toward the highest-margin product, which is high-bandwidth memory for AI chips. Consumer RAM got squeezed out. That's not a failure, that's a market working.

NovaA market working? PC builders are paying $500 for RAM that cost $90 eighteen months ago, laptop prices are up 15 to 30 percent across the board, and the average smartphone is heading toward a $200 price increase. Calling that "the market working" is a pretty comfortable position to hold when you're not the one getting squeezed.

AtlasThe price signals are real, which means the information is flowing correctly. Nobody promised consumers cheap RAM forever. The underlying economics changed.

NovaThe economics changed because three companies with a collective 90-plus percent market share decided server margins were more important than everything else. That's not exactly a free and open market making an efficient choice.

AtlasA concentrated market is still a market. The three producers are competing against each other for the most lucrative contracts. Nobody is colluding to keep consumer prices high — AI is just paying more.

NovaAnd every person buying a laptop, a phone, or a gaming console is paying the difference. That's the part that keeps getting buried in the supply-chain language.

Why HBM Eats the Factory: The Wafer Math Nobody Talks About

AtlasTo understand why this is so hard to fix quickly, you need to understand what high-bandwidth memory actually is and why it consumes so much manufacturing capacity. HBM is built by stacking DRAM dies vertically and connecting them with microscopic wires called through-silicon vias. That stacking process means each gigabyte of HBM requires roughly three to four times the wafer area of a standard DDR5 gigabyte. Micron confirmed that ratio in its own earnings materials. TrendForce puts it at four times.

NovaSo when a hyperscaler orders a rack full of Nvidia AI accelerators, they're not just buying a few extra chips. Each of those accelerators comes loaded with HBM, and each gigabyte of that HBM consumed the factory space that could have produced four gigabytes of the DDR5 sitting in your laptop. That's the direct trade-off. It's not abstract.

AtlasThe scale of the demand reallocation is striking. TrendForce data shows HBM consumed about 19 percent of total global DRAM wafer output in 2025. By the end of this year it's at 22 percent. By the end of 2027, that share is projected to hit 30 percent. One in three DRAM wafers is going to AI memory, period.

NovaAnd this is before you factor in that OpenAI reportedly locked up agreements in 2025 for up to 900,000 DRAM wafers per month for the Stargate project alone. CSIS puts that at roughly 40 percent of global DRAM output. Forty percent. Committed to one AI infrastructure project.

AtlasThat is what an aggressive capital allocation looks like. The AI buildout required it.

NovaIt required it for the companies bankrolling Stargate. For the consumer trying to buy a mid-range laptop, it means memory now makes up 30 to 40 percent of the manufacturing cost of their machine, up from the historical 16 to 20 percent. Gartner says average PC prices are up 17 percent this year directly because of this. That's a tax on everyone who isn't a hyperscaler.

AtlasAnd the manufacturers are responding. Samsung is targeting 150,000 wafers a month of 1c DRAM capacity primarily for HBM4. SK Hynix has sold out its entire 2026 HBM production slate and stopped taking new orders. Micron's 2026 HBM output is gone too. Capacity is being added — it just takes years to build a fab.

NovaWhich is exactly the problem.

The Big Three's Triage: Why Servers Win and Consumers Lose

NovaWhat I want people to understand is that this isn't just scarcity. This is a deliberate production triage by three companies who control the entire global supply. Samsung, SK Hynix, and Micron have all been converting existing DDR4 and DDR5 production lines into HBM lines — what analysts call brownfield conversions. Every line that gets converted is a line that is no longer making the chips that go into phones and laptops.

AtlasBecause HBM commands dramatically higher margins. SK Hynix's CEO said the broader industry was in for its worst shortage in history and that even with new facilities, demand would outpace supply through 2030. When your highest-margin product is also the one where you're sold out, you make more of it. That calculation is not mysterious.

NovaAnd data centers are projected to consume roughly 70 percent of all memory devices produced worldwide this year. Seventy percent. Before the AI buildout, servers took about 30 percent. The consumer share of a market that was already small got cut nearly in half in the span of two years.

AtlasWhich is why Q4 DRAM contract prices are still projected to rise another 10 to 15 percent over Q3, even as Q3 slowed down from the 60 percent quarterly jumps we saw in Q2. TrendForce's read is that the slowdown is consumers hitting an affordability ceiling, not supply improving. The shortage itself has not changed.

NovaAnd Micron has already committed more than 75 percent of its 2027 output. The CEO said customer conversations are now mostly about 2028. So the three companies who make all the RAM are sold out through next year, and when they do have conversations, they're talking to hyperscalers and server builders, not PC makers.

AtlasPC and smartphone manufacturers are at the back of the queue because they pay less per gigabyte. The market is ranking buyers by willingness to pay. That's textbook allocation under scarcity.

NovaThe textbook doesn't mention that regular people trying to buy a phone or fix a broken laptop have no alternative supplier to go to. There are three vendors. Three. That's not a market, that's a tollbooth.

What This Costs You: Laptops, Phones, Consoles, and the New Price Floor

NovaThe downstream effects on consumer hardware are already visible and documented. Apple raised MacBook and iPad prices by nearly 20 percent. Sony raised the PlayStation 5 by $100 to $150. Microsoft raised Xbox prices by a similar range. Valve priced its new Steam Machine at $1,049 for the base model. Dell, HP, Lenovo, Acer, and Asus have all confirmed price increases of 15 to 30 percent. The BBC reported that console prices are now 30 to 40 percent higher than they were a year ago.

AtlasOn smartphones specifically, IDC projects the average selling price will hit an all-time high of $523 this year, a 14 percent increase. The sub-$100 smartphone is gone. Smartphone production is tracking roughly 2 percent below what manufacturers planned because they simply can't get enough memory at the volumes they need. That's not price gouging — that's a supply constraint propagating through the chain.

NovaAnd the people hit hardest are the people who can least absorb it. Budget smartphones in price-sensitive markets are being downgraded to 4GB RAM configurations to preserve any kind of accessible price point. IDC flagged that explicitly. The premium user loses some money. The entry-level buyer loses access to a usable device.

AtlasAnd yet manufacturers are still finding buyers. Apple's margins are holding up fine. The premium segment is resilient. The lesson is that if you make a product people can't live without and supply is constrained, pricing power is real.

NovaApple has long-term supply contracts that insulate them through 2026. The brands without that leverage — the mid-tier OEMs, the white-box PC makers — are the ones who can't keep the lights on. And even Apple flagged memory costs as a significant line-item pressure on its last earnings call. Tim Cook called the shortage a "100-year flood." When Tim Cook is the one warning you, something has gone structurally wrong.

AtlasHe was warning investors. The stock absorbed it. That's not the same as the sky falling.

NovaTell that to the teacher who needed a replacement laptop this fall and found that the $600 machine she bought two years ago now starts at $850.

2027, 2028, 2029: The Debate Over When This Ends

AtlasOn the relief timeline, the honest answer is that every credible forecast keeps getting pushed further out. At the start of 2026, the consensus was late 2027. By mid-year, UBS moved it to Q2 2028. Micron's CEO said in his most recent earnings call that supply-demand conditions will be tighter in 2027 and 2028 than they were in 2026, and that he has no line of sight to when balance returns. Samsung's memory sales chief told analysts this month the shortage deepens through 2027 and stays tight into 2028.

NovaAnd even the optimistic case isn't a return to 2025 prices. Counterpoint Research projects DRAM prices falling roughly 23 percent in 2028 compared to 2027 — but still up around 230 percent versus 2025. Bernstein's forecast has the price per gigabit dropping sharply in 2028, but landing at roughly twice the pre-crisis floor. Deloitte published an analysis saying new fab supply won't come online in volume until 2029 or 2030. This is a multi-year structural reset of what memory costs.

AtlasMicron's first new Idaho fab delivers initial wafer output around mid-2027. A second follows in late 2028. Samsung's Pyeongtaek P5 megafab targets mass production in late 2028. SK Hynix's Yongin Y2 facility doesn't start until 2029. You can see the supply additions on paper — they just require patience and they're all denominated in years, not months.

NovaAnd every one of those fabs is being built primarily for HBM and server DRAM, not consumer DDR5. TrendForce projects HBM will consume 30 percent of total DRAM wafer capacity by 2027. If AI demand keeps accelerating, the new capacity gets absorbed before it can reach retail shelves.

AtlasWhich is the scenario where we talk about the AI bubble. CNBC reported this week that memory stocks are getting hit partly on fears of a 2029 oversupply, but also on concerns about Chinese competitor CXMT entering the server memory market. If a burst or even a plateau in AI infrastructure spending pulled HBM demand back sharply, conventional DRAM supply would loosen fast — because that wafer capacity would suddenly be freed up for commodity production. That is the wild-card scenario every analyst acknowledges but can't time.

NovaA bubble pop would help RAM prices, but the damage to everything else makes it a brutal way to get there. If hyperscalers stop spending, data center workers lose jobs, AI startups collapse, equity markets take a hit. Counting on a crash to make your laptop affordable again is not a policy. It's a hostage situation with the whole economy in the room.

AtlasNobody said it was clean. The structural read is that prices stay elevated through 2027, begin a slow descent in 2028 as fabs ramp, but do not return to the pre-2025 floor — probably ever. Analysts use the term "higher equilibrium" for a reason. The days of $90 for a 32GB DDR5 kit were a product of overcapacity during the DRAM down-cycle. That cycle is over.

NovaTwo decades of falling RAM prices wiped out in eighteen months. Tom's Hardware ran the analysis — DDR5 now costs what it cost in 2007 in nominal dollars. If you're planning around the idea that a $400 RAM kit will be $120 again someday soon, the data does not support that. Plan around the new floor, fight for the policies that build domestic fab capacity faster, and understand that the three companies at the top of this stack are making record profits while everyone downstream absorbs the cost. That asymmetry is the story, and it doesn't end in 2027 just because a few fabs break ground.

Sources

1. RAM price tracking 2026 — lowest price on DDR5 and DDR4 ...tomshardware.com
2. Corsair's Vengeance 32GB DDR5-5200 RAM becomes the cheapest 2x16GB kit— DDR5 kit still $379tomshardware.com
3. The AI-Driven Memory Shortage: DRAM Prices, Inflation and Market ...jpmorgan.com
4. One year into the AI-induced RAM apocalypse — how much does memory actually cost, and is there hope for a more affordable future?tomshardware.com
5. Memory price surge begins to cool as consumers hit affordability limit — AI demand still keeps DRAM and NAND prices climbing through Q3 2026tomshardware.com
6. Corsair Vengeance 32GB DDR5 kit price-drops in the big memory saletechradar.com
7. Corsair has slashed $103 off this 32GB Vengeance RAM — the cheapest kit on the market at this speedtomshardware.com
8. The memory crisis is getting so bad that even retro RAM prices are going to the Moontheregister.com
9. Rampant AI demand for memory is fueling a growing chip ...fortune.com
10. Research publicly available data on DRAM and HBM ...useluminix.com
11. Samsung's own memory chief says the RAM shortage won't peak until 2028 - Startup Fortunestartupfortune.com
12. Buying a new phone or laptop in 2026? Here's why it'll cost morestuff.tv
13. Laptops, phones and even cars are facing price hikes. ...koat.com
14. AI Boom Drives RAM Shortage, Threatening PC Sales and Console ...ground.news
15. Investigation: RAM prices are falling. Don’t fall for itpcworld.com
16. RAM prices keep going up — what is RAMageddon, and why is it getting worse?tomsguide.com
17. RAM Prices May Ease in 2028. High-VRAM GPUs Are a Different Storyimplicator.ai
18. DRAM ETF Records Biggest Daily Outflow as AI Bubble Fears Riselongbridge.com
19. AI Is Keeping DRAM Prices High — Here's Exactly When the Market ...finance.yahoo.com
20. DRAM and NAND Prices Rise 20-30% Quarter-on- ...news.futunn.com
21. DRAM crisis: Analysts expect drastic price drop in 2028notebookcheck.net
22. DRAM Buyers Face Two More Years Of Scarcity As NAND Tilts ...techtimes.com
23. AI Wiped Out Two Decades of Falling RAM Prices: Full DDR5 ...techtimes.com

More episodes

• How Do Waymo Cars Work? Sensors, AI, and Safety DataTechnology · 11 min
• Tesla Roadster Flying Car ExplainedTechnology · 7 min
• The Social Reckoning: True Story Vs. Sorkin's FilmTechnology · 10 min
• Why OpenAI Paused Training on Its Most Advanced ModelsTechnology · 5 min
• CXMT DRAM: China's Memory Chip Breakthrough, ExplainedTechnology · 7 min

Browse every episode