Why RAM Prices Are Skyrocketing: The AI Memory Crisis Explained
The PC you were planning to build in 2026 just became significantly more expensive and it's going to stay that way for years. RAM prices have soared way too much and the consumer memory is now a low-priority commodity while AI infrastructure consumes RAM production capacity at high profit margins.
You might remember when scalpers were buying all the latest GPUs to mine crypto and they caused GPU prices to skyrocket and made NVIDIA one of the most profitable companies of the world. For RAM, it is similar but replace the scalpers with multibillion dollar enterprises. Unlike previous chip crises that resolved within 12-18 months, this one might as well persist well into 2028 or beyond.
TL;DR
You know that companies need GPUs for AI, but they also need a ton of high-end memory. And only three companies really mass produce it, Samsung, Micron, and SK Hynix. And these companies have to decide, okay, we have all these memory wafers made by TSMC. What type of memory should we make them into? We could do VRAM. We could do memory for phones. We could do memory for gaming PCs. or should we make it into High Bandwidth Memory (HBM) for AI? Hmmm…..Which is the most profitable, the most in demand? I could just sell it to these multi-billion dollar companies. I'll let you guess which one they picked. So, uh, it's going to be rough out there.
What Is RAM?
RAM, or random-access memory, is your computer's short-term memory. Every application running on your device, from the web browser you’re reading this on to your video editor processing footage, everything depends on RAM to temporarily hold some data while the CPU works on it. Unlike Hard Drives or SSDs that permanently retain files even when powered off, RAM loses all data the moment your device shuts down. The speed of RAM (measured in nanoseconds) matters far more, because computers need to rapidly read and write data millions of times per second.
Modern systems have come a long way with several generations of RAM. DDR5 is the latest standard for consumer and business PCs, offering higher speed and better efficiency than its predecessor DDR4. Server systems use specialized memory like RDIMM (Registered DIMM) for reliability. And now this is important — increasingly, high-bandwidth memory (HBM) powers data centers and AI accelerators, providing the high speed and huge storage required to feed GPUs and TPUs for AI processing. Manufacturers are supposedly sidelining DRAM to prioritize newer, higher-margin HBM and consumers are caught in the squeeze.
Surely, people are smart, why didn’t they realize they can just make more RAM? To understand why we can’t just manufacture more RAM, let’s look at how it’s made.
What Is RAM Made Up of and How Is It Manufactured?
RAM is fundamentally built from sand. Silica (silicon dioxide, SiO₂) is extracted from sand and processed into pure silicon, the same material used to make desktop CPUs but fabricated using very different architectures. The manufacturing process begins with a purified silicon ingot, a single-crystal cylinder roughly the diameter of a dinner plate, which is sliced into ultra-thin, highly polished wafers less than a millimeter thick. (Crucial)
As you read this, you may have thought, “Okay wow, this sounds hard and expensive. Yes, it is hard and expensive. But the next process is much, much harder: extreme ultraviolet (EUV) lithography, the cutting-edge technology that makes 12nm and smaller DRAM nodes. New DRAM has circuits so small that normal photolithography doesn’t work as the wavelength of light itself is too large.
The actual chip production happens in a thoroughly controlled clean environment through EUV lithography. Engineers first coat the wafer with many layers: a glass layer (silicon dioxide) created by exposing silicon to oxygen at 900°C for hours. Then comes a light-sensitive chemical called photoresist. But unlike older UV systems, EUV uses 13.5-nanometer wavelength light. EUV doesn’t exist naturally on earth so it is generated by blasting microscopic Tin (Sn) droplets with a high-power CO₂ laser 50,000 times per second twice, creating plasma 40 times hotter than the sun's surface.
This EUV light reflects off precisely engineered multilayer mirrors (no lenses, as EUV absorbs into glass) and projects through a patterned mask (acting as a stencil) onto the wafer. The extreme precision chemically alters the exposed photoresist at atomic scales. Developers then remove the exposed portions, leaving a pattern on the wafer that matches the intended circuit design. It is all done in vacuum by the way, as any air would absorb or scatter the EUV.
The next step is etching. The exposed areas are now chemically weakened, so powerful acids etch away unprotected material, leaving precise grooves according to the mask pattern. This etching step repeats dozens of times as engineers build transistors, capacitors, resistors, and tiny electrical connections layer by layer.
Once the transistor logic is complete, aluminum or copper is deposited across the wafer and patterned into interconnect "wires" that connect components and route electrical signals. A final layer seals the circuit. The wafer is then tested chip by chip; defective dies are marked and discarded while functional ones advance to assembly. A diamond saw cuts the wafer into individual dies, which are mounted on lead frames, connected with microscopic gold wires, and encased in plastic or ceramic packages to create the finished memory modules you buy.
The critical bottleneck here are yield rates. A single defect in a circuit can render an entire die worthless. When Samsung or SK Hynix transitions to a new technology node (like the 1c process, a 10-nanometer-class standard), initial yields often drop to 50-70% as engineers debug process variations, mask defectivity, and other manufacturing challenges. With time, production matures, yields climb toward 85-90%, but this learning curve takes months and diverts enormous R&D resources. Early runs of advanced nodes can take 12-18 months to approach commercial viability.
Although for it’s complexity, they make it up with the scale of production. A single 30cm wafer can contain thousands of memory dies. A large wafer fab produces 15,000-30,000 wafers per month. A single fab decision, like allocating 10,000 wafers monthly toward HBM rather than DDR5 represents the loss of billions of DRAM bits hitting the consumer market. This is what is happening right now.
TSMC and ASML on Supply
The semiconductor industry mostly depends on a single technology company and a single machine supplier, which creates a lot of leverage concentrated in two corporations. It is also the reason for any shortage as they control the basic supply.
TSMC (Taiwan Semiconductor Manufacturing Company) is the world's premier chip foundry, operating the most advanced fabrication facilities on Earth. But TSMC primarily manufactures logic chips (processors) rather than memory. What makes TSMC important to this story is its control over cutting-edge process nodes. TSMC's 3nm and 2nm production lines set the pace for the industry's technological frontier. Every consumer-grade DRAM manufacturer must license or develop process technology that follows TSMC's progress to remain competitive. TSMC now has over 70% market share and other competitors like Intel and Micron who have their own foundry cannot compete on the cutting edge.
I ask you to listen to Acquired podcast episode on TSMC to learn more about TSMC.
ASML, a Dutch company, holds a total monopoly more absolute than TSMC's dominance. Anytime TSMC or any other fab foundry for that matter, needs an EUV lithography machine they go to ASML. ASML manufactures the lithography machines that etch circuit patterns onto wafers using extreme ultraviolet (EUV) light. After decades of R&D costing billions, ASML has (kind of) perfected EUV technology. The EUV light bounces off multilayer Zeiss mirrors (costing millions apiece also) onto the wafer, burning patterns so small they're nearly invisible to optical microscopes.
A single ASML EUV machine costs about >$250 million and takes years to ship. Their new lineup of High NA machines is close to $400 million per piece. Intel bought it recently. No other company on Earth can manufacture these machines at the moment. Every memory manufacturer, every processor fab, competes for ASML machine allocation. The Dutch government recently restricted ASML exports to China, tightening the global supply even further. This single bottleneck means that even when a manufacturer wants to expand DRAM production, they cannot simply order machines and start fabricating chips. ASML contracts advance years into the future, and production capacity remains stagnant by the number of EUV machines ASML can manufacture annually, approximately 50-60 units worldwide.
When memory manufacturers shift towards HBM (which requires advanced process nodes and EUV machines), they're necessarily taking capacity away from commodity DRAM. And once they do, adding DRAM capacity back requires ASML machines that may not arrive until 2027 or 2028.
HBM vs DDR5
High Bandwidth Memory (HBM) and DDR5 are made up of same DRAM chips but their architecture makes a huge difference. They serve different computing niches, with HBM dominating AI workloads due to its 3D-stacked architecture versus DDR5's traditional planar design.
Architecture: HBM stacks many DRAM dies vertically using through-silicon vias (TSVs) on a silicon interposer—3D design yielding 1024-bit interfaces. DDR5 uses traditional 2D side-by-side chips with 64-bit memory channels.
Bandwidth: HBM2E delivers 410+ GB/s per stack; a single NVIDIA H100 GPU accesses terabytes/second. DDR5 tops out at ~33-50 GB/s per channel—adequate for gaming but starved by AI workloads. Each second, terabytes of information can be sent to the CPU for processing with HBM.
Size & Power: HBM's compact stacks use 75% less space with lower power via short TSV connections. Basically they are placed closer to the CPU. DDR5 spreads across larger PCBs, consuming more energy for equivalent capacity.
Cost: HBM costs $40-50/GB (3x DDR5 wafer area, due to complex stacking). DDR5: $5-10/GB for consumers—until AI demand triples prices.
Why It Matters: One HBM stack equals 10-15 consumer DDR5 modules in bandwidth but consumes vastly more production capacity. Manufacturers are prioritizing HBM's larger margins.
Comparison via Semiconwiki
Now that we know how expensive and hard it is to develop chips, let’s shift our focus from technology to the businesses driving the price increase.
AI and HBM, GPUs and NVIDIA
I might sound like a broken record here but the same large enterprises on frontier of AI development are again responsible for inflated RAM prices.
In data centers, where all the supposed magic of AI (✨) happens………rolling back, where all the computation and processing happens, we need fast data storage. From previous sections we know RAM is used as working memory, anything that the computer is working on, like active programs are loaded into it.
Training data thus needs to be loaded in the RAM for fast access in order to train the models. DDR5 does this well, but HBM does it even better. When you ask ChatGPT to make a million-dollar SaaS (without mistakes of course) , they use the trained model loaded on the RAM to run inference and send you the answer. HBM reduces those memory bottlenecks and the answer is processed faster. Inference happens millions of times daily.
Training a large language model like OpenAI's GPT-4 requires roughly 900GB of high-speed video RAM on data center GPUs, while inference (running the trained model) consumes slightly less. But training occurs rarely; inference happens millions of times daily. The real demand driver is NVIDIA's GPU market domination. NVIDIA holds approximately 61% of HBM consumption as of 2025, expanding to 68% by 2025 forecast. Each NVIDIA H100 GPU ships with 80GB of HBM3 memory; the newer Blackwell architecture will feature HBM4 with even higher capacity. A single data center rack housing many GPUs can contain 100GB-500GB+ of HBM in a single rack which is memory capacity equal to several consumer PCs,. Read more about HBM usage in this report.
Each new data center requires massive amounts of HBM, DDR5 server memory, and storage. The trillion dollar companies have money to spend (Nvidia's Q3 2025 revenue exceeded $35 billion annually) and will pay premium prices to secure memory allocation. They offer long-term contracts worth billions. Memory manufacturers face a choice between selling commodity DDR5 to consumers at $10-15 per gigabyte or selling HBM to companies at $40-50+ per gigabyte, and obviously choose the latter.
OpenAI on Demand
OpenAI has secured approximately 40% of the world's DRAM supplies through major agreements signed with Samsung and SK Hynix in early October 2025. These contracts represent roughly 900,000 DRAM wafers per month, equivalent to nearly half of the projected global memory production capacity in 2025 all for the Stargate data center project, which carries an estimated budget of approximately $500 billion.
The procurement model is what sets this arrangement apart from traditional enterprise memory purchasing. OpenAI purchases raw DRAM wafers, not pre-fabricated memory modules. These wafers are not sliced and do not adhere to certain specifications such as DDR5 or HBM. Basically, OpenAI can do whatever it wants with the modules once they are created.
The immediate consequence is that monthly ~900,000 wafers are now exclusively committed to OpenAI's infrastructure. This big chunk of supply removed from market explains the current shortages and price increase observed throughout late 2025. Suppliers report wait times for new DDR5 shipments stretching up to one year, while high-capacity consumer kits have increased many times in price since the beginning of last month. Some retailers have abandoned fixed pricing entirely, acknowledging that prices fluctuate almost daily based on OpenAI's allocation decisions and forward contracting behavior.
Read more about this here.
What's Happening Now: Crucial, Samsung, and Others
The most dramatic shifts are unfolding in real-time among the three major DRAM manufacturers—Samsung, SK Hynix, and Micron as they rapidly phase out consumer product lines and redirect capacity toward lucrative enterprise contracts and AI infrastructure like OpenAI's Stargate project.
Micron Technology made the boldest move; they’re shutting down its Crucial consumer brand entirely by February 2026, exiting retail memory markets after nearly 30 years. Micron explicitly stated it can meet only "half to two-thirds" of demand from key customers, prioritizing high-margin HBM and enterprise applications. They warn shortages will "persist beyond 2026”.
Early October, Samsung signed with OpenAI for part of the 900,000 raw DRAM wafers/month deal (40% of global supply), removing unsliced wafers from consumer markets entirely. Samsung Electronics now scaled back HBM production to redirect toward DDR5 RDIMM modules which are being sold at a higher price.
OpenAI's 900,000 wafers/month commitment for its $500 billion Stargate project pushes this transition, resulting in DDR5 wait times of up to a year and daily price swings as shops leave set pricing. Competing AI startups (Anthropic, xAI, and Meta) face a two-tier market: OpenAI/Microsoft achieves optimal scalability through direct manufacturer relationships, while others adjust for memory limits.
How Will This Affect The Normal Consumer
Everything that uses a RAM stick will get more expensive of course.
Desktop PC Builders are already absorbing immediate shocks. A 32GB DDR5 memory kit that cost $60-80 in early 2025 now costs $180-250, with premium brands asking $250-350. Some retailers report individual 32GB kits priced above $300. For a budget PC build, memory now represents 20-25% of total component cost, up from 8-12% a year ago. A typical gaming PC refresh that cost $1,200 in 2024 now costs $1,500-1,700 even before accounting for GPU or CPU price increases. It’s like if you wanted just a RAM upgrade on your build, you’d be better buying a new prebuilt.
Laptops don’t see immediate surge as they stockpiled memory earlier. But as their reserves exhaust, we’ll see a jump in their prices as well. Likely by mid 2026. Apple, meanwhile, operates in a completely different paradigm. The company that faced mockery for charging $400 to upgrade from 8GB to 32GB of RAM suddenly looks less like a price gouger and more like a company that understood where the memory market was headed.
Pre-built PCs left in stock should be available at a lower cost. Although I think shops will also increase the prices seeing the market trend.
Mobile Device Users might expect reprieve, but LPDDR (mobile memory) faces similar constraints. LPDDR5X prices are rising, though less dramatically than DDR5 because smartphone manufacturers contract direct supply from Samsung/SK Hynix at their desired rates. However, the next generation of flagship phones with more on-device processing may require more LPDDR5X RAM, putting pressure on supply. Budget and mid-range phones may see memory configurations from 2018 (4GB instead of 6GB, 8GB instead of 12GB) as manufacturers optimize costs. Smartphone prices might increase altogether.
Can I Be Optimistic They'll Lower Prices?
This is the hardest section to write because the evidence is unambiguous: prices will not significantly decline until 2028 at the earliest, and even that assumes no major surprises.
Prices as they stand now will probably become the norm, there will be discounts sure, but as we saw in the last 5 years with GPU prices, how there was a shortage, MSRP prices increased overall and mostly stayed there, then GPU companies brought mid tier models with less VRAM or compute to add products and fill the price gap between high end and budget CPUs.
The bottom line: if you can delay a PC purchase until late 2027 or 2028, you should. If you need a system now, expect elevated pricing to persist for 24+ months. For context, delaying a $1,500 PC purchase by two years and saving only $200-300 through price reductions might not be worth the opportunity cost of using an underpowered system. The rational play is to buy now at inflated prices if you need the system immediately, or delay a lot if you can work around current hardware.
The era of cheap commodity RAM might just have ended. The semiconductor industry has fully transitioned to a premium model where AI infrastructure captures the vast majority of memory production, and consumers compete for scraps at prices that would have seemed impossible in 2020. There is no relief in sight, and hoping for price cuts is a strategy doomed to disappointment.
Thank you for reading 💓, you've reached so far so please like my blog which keeps me motivated to write more content, and also don't forget to leave your views down below 👇 . Here have two cookies🍪🍪. Happy Holidays! ☃️
External Links:
Manufacturer Insights & Products:
Technical Deep-Dives & Manufacturing:
Industry News & Financials:
Samsung and SK Hynix Join Stargate Infrastructure Initiative
Intel Installs Industry's First Commercial High-NA EUV Tool – Tom’s Hardware
Additional Resources:
