Written by Shaun David
Shaun David
Regulation • Trading Algorithms • Market Analysis
I’m the co-founder of CleaRank. I’ve spent my career analyzing markets, building trading algorithms and managing long-term portfolios, and I’ve learned that prices move on crowd psychology as much as on the numbers. Today I write market analysis on stocks, ETFs, forex and crypto, with a focus on AI infrastructure and the companies behind it. under our CLEAR™ Editorial Code The CLEAR™ Editorial Code is the standard every CleaRank market analysis and educational article follows. It keeps our work independent of advertisers, grounded in verifiable data, reviewed before publication and corrected promptly if we get something wrong. .
Last fact check on November 20, 2026 by
Jacob Bakshi
Jacob Bakshi
Senior Market Strategist
Trading CFDs and options has been my playground for years, and I love helping others understand these powerful tools and what makes the financial world tick. My work mostly focuses on giving traders the confidence to make informed decisions with unbiased reviews into platforms that prioritize fair pricing, advanced tools, and reliable execution because In fast-moving markets, every detail matters. I have a background in market analysis and risk management, and I’m always on the lookout for brokers that offer the right tools for serious traders.
The Memory Chip Shortage Has a Twist: Why Robots May Make Plain DRAM the Winner
Everyone is chasing the expensive memory built for AI data centers. The next wave of demand may come from the basic chip the industry is leaving behind.
Key Points
- Today's shortage began as a glut. In 2023 memory was oversupplied, prices collapsed and investment in new production lines froze. Clean rooms, the sterile production space where chips are made, take years to build, so new capacity is arriving late and slowly.
- Makers are moving wafers from standard DRAM to HBM, the stacked memory used in AI data centers, which needs roughly 2 to 4 times the wafer area per bit. That squeezes the plain DRAM used in PCs and phones. TrendForce says conventional DRAM contract prices rose roughly 93% to 98% quarter over quarter in Q1 2026 and forecasts another 10% to 15% in Q4 2026.
- The twist: Elon Musk wrote that "Robot RAM needs will exceed everything else on Earth," replying to a Counterpoint Research forecast that humanoid robot DRAM demand will grow about 20 times by 2030. The caveat is scale. By PC Gamer's estimate, even the 2030 figure is about 0.054% of current global DRAM capacity.
- The five big names split by time horizon. SK hynix leads HBM today, Samsung has the largest standard DRAM capacity, Micron gains from higher prices now and new US fabs later, CXMT is the most exposed if prices turn, and Sandisk is a NAND specialist with much of its output contracted years ahead.
- CXMT reported an 89.5% gross margin in Q3 2026, but SemiAnalysis says its DDR5 cost per bit is still over 30% higher than the three leaders and that the margin gain is driven mainly by pricing. High prices, not low costs, explain the number.
- Anthropic's IPO under the ticker ANT raised about $81.3 billion, the largest on record. Its filing points to compute and infrastructure as a use of proceeds, and memory is a growing share of compute cost, so part of that money ultimately supports memory demand. ANT is the customer, not a memory stock.
Why Is There a Memory Chip Shortage in 2026?
Short answer: A memory glut in 2023 crashed prices and froze investment in new production lines. Clean rooms take years to build, so supply could not respond quickly. Makers then moved wafers to HBM for AI data centers, which uses roughly 2 to 4 times the wafer area per bit, leaving standard DRAM short. TrendForce sees prices still rising in Q4 2026.
Memory chips are the least glamorous part of a computer, and in 2026 they are among the hardest to buy. Prices for the ordinary working memory inside PCs and phones have roughly doubled in a single quarter this year, and the companies that make it are sold out. The usual explanation is AI: data centers want a premium kind of memory, and the industry is rushing to supply it.
That explanation is correct, and it is also where the story gets interesting. If the next large buyer of memory is not a data center but a machine that walks around on a battery, it will want the opposite product: the plain, power-efficient chip the industry is currently moving away from. This article explains the shortage in simple terms, sets out that twist with its caveats, compares five memory makers over the short and long term, and looks at where this week's record ANT IPO fits in.
“The market is pricing memory as if the only customer that will ever matter is the data center. The customer after that has no power cord and no cooling loop, and it buys a different chip.”
Jacob Bakshi, CleaRank Senior Market Strategist
Because the industry stopped investing in 2023, and what capacity it has is being pointed at AI. The shortage has two causes that arrived one after the other.
The first is the hangover from the last downturn. In 2023 there was too much memory on the market. Prices collapsed and manufacturers froze investment in new production lines. Memory is made in clean rooms, which are sterile production spaces that take years to build. A decision not to build in 2023 becomes missing supply several years later, and a decision to build today arrives late and slowly.
The second cause is that the capacity that does exist is being redirected. Makers are shifting wafers away from standard memory and toward the premium memory that AI data centers need. The next section explains why that swap removes more supply than it adds.
The price data shows the result. According to TrendForce, conventional DRAM contract prices rose roughly 93% to 98% quarter over quarter in the first quarter of 2026. Its latest forecast, DRAM prices seen rising in 4Q26, calls for a further 10% to 15% increase in the fourth quarter, or 15% to 20% once HBM is included, with NAND flash contract prices up 15% to 20%. Server DRAM is expected to stay undersupplied.
Relief is not close. Most manufacturer forecasts point to 2028 before supply catches up, and demand for HBM in 2027 is seen far above 2026.
How a 2023 Glut Became a 2026 Shortage
Memory makers stopped building when prices collapsed. Then AI took the factories that were left.
-
2023 The Glut
Oversupply crashes prices. Investment in new production lines freezes.
-
2024 to 2025 The Lag
Clean rooms take years to build, so new capacity arrives late and slowly.
-
2025 to 2026 The Shift to HBM
Makers move wafers to HBM for AI servers. HBM needs roughly 2 to 4 times the wafer area of standard DRAM.
-
2026 The Squeeze
Standard DRAM for PCs and phones runs short. Contract prices rose over 90% in one quarter early in the year and are still climbing 10% to 15% in Q4.
-
Late decade The Twist
Robots need memory on board, and HBM runs too hot and power-hungry for them. Standard low-power DRAM becomes the growth product.
Today's profits sit in HBM, the premium memory stacked beside AI processors. A robot cannot carry a data center's power and cooling, so its memory has to be the efficient standard kind. If robot volumes arrive, the product everyone is deprioritizing today becomes the one in shortest supply.
Source: CleaRank analysis; TrendForce, Counterpoint Research, company filings, Elon Musk on X. Data as of November 20, 2026.
HBM vs DRAM: What Is the Difference?
Standard DRAM is the simple, power-efficient working memory in almost every device. HBM is a stacked, far more complex version built for AI data centers. They come from the same factories, which is the root of the problem.
Standard DRAM. This is the memory a PC or phone uses to hold whatever it is working on at that moment. It is the basic product of the memory industry: relatively simple, efficient with power and sold in very large volumes.
HBM. High bandwidth memory stacks memory layers on top of one another so that AI processors in data centers can be fed data fast enough. It is far more complex to make, and it runs hot and draws a lot of power, which a data center can handle because it has mains electricity and cooling systems.
The important number is wafer area. By the analyst framing we use here, HBM needs roughly 2 to 4 times the wafer area per bit of standard DRAM. A wafer moved from standard DRAM to HBM therefore produces far fewer bits of memory than it did before. Every such move makes makers more money and leaves the world with less memory overall.
That is why the shortage shows up in ordinary products. Standard DRAM for PCs and phones is short, device makers are paying more for it, and those higher prices are being passed on to consumers.
The Twist: Why Could Robots Change Which Memory Wins?
Because a robot cannot use the premium chip. This is our thesis, not a consensus forecast: if humanoid robots are produced in large numbers, they will need the standard low-power DRAM that the industry is currently giving up capacity in.
The logic is physical. A robot acts in the real world in real time. It cannot pause mid-step to wait for an answer from the cloud, so its memory has to sit on board. And unlike a server, a robot has no mains power and no cooling pipes. Hot, power-hungry HBM does not fit inside a machine that runs on a battery. Standard low-power DRAM does. In that sense, plain DRAM is the robotics product.
The idea reached a wide audience in late October 2026, when Elon Musk replied on X to a Counterpoint Research forecast with the words: "Robot RAM needs will exceed everything else on Earth." The exchange is covered in PC Gamer's report on Musk's robot RAM comment.
The Counterpoint forecast he was replying to projects that DRAM demand from humanoid robots will rise from about 8,300 terabits today to about 172,000 terabits by 2030, an increase of about 20 times. Memory per robot is expected to go from 19GB to 39GB.
There is already a sign that memory is a constraint for robot makers. Musk has also said that Tesla cut the memory on its AI5 chip in half, and on its AI6 chip by a third, in order to get enough volume for its Optimus robot. AI5 was reportedly later set at 96GB.
Now the caveat, which matters as much as the thesis. Even at the 2030 level in the Counterpoint forecast, robot demand would equal about 0.054% of current global DRAM capacity, by PC Gamer's estimate (172,000 terabits against roughly 40 exabytes of capacity). Twenty times a tiny number is still a small number. Robots are not driving memory prices in 2026 and will not in 2027. If this story becomes real, it is a story for after 2028.
“Robot demand is about 0.05% of capacity today, so nobody should buy a memory stock on it this year. The point is direction. When you are planning a fab that opens in 2030, the mix decision you make today matters.”
Jacob Bakshi, CleaRank Senior Market Strategist
Which Memory Stocks Win in the Short Term and the Long Term?
In the short term, the HBM leaders and anyone selling into rising prices. In the long term, the answer depends on who has low-cost standard DRAM capacity when new fabs open. Here is how five companies line up.
SK hynix (SKHY)
SK hynix is the HBM leader. Counterpoint put its HBM share at 50% in the second quarter of 2026, ahead of Samsung at 33% and Micron at 18%. IDC data cited in the company's SEC filing gave it 56.4% of HBM revenue in the first quarter of 2026. Its sales are backed by long-term contracts. SK hynix has filed for a US listing on Form F-1 and trades as SKHY.
Short term: the strongest position in the product in shortest supply. Long term: the company most tied to data center demand continuing, and the least exposed to the robot thesis.
Samsung
Samsung's HBM share is rising with its HBM4 generation, and it benefits from the NAND shortage as well. It also has the largest standard DRAM capacity in the industry.
Short term: gaining ground in HBM while earning more on everything else. Long term: if standard low-power DRAM becomes the scarce product, the largest capacity in that product is the asset to own.
Micron (MU)
Micron benefits from price rises across memory. Its distinguishing feature is a pipeline of new US fabs. In May 2026 it began making 1-alpha DRAM, the most advanced memory ever produced in the United States, at its Manassas, Virginia fab, with qualified output expected by the end of 2026. That line serves long-life DDR4 for cars, defense and industry; Micron's leading-edge US capacity comes later, from new fabs in Boise, Idaho and Clay, New York. More detail is in the Micron (MU) stock analysis on CleaRank.
Short term: higher prices on every product line. Long term: new capacity arriving in stages across the period when robot demand would begin to matter.
CXMT
China's CXMT is the newcomer. It listed on the Shanghai STAR Market on July 27, 2026, raised about 57.9 billion yuan, and its shares rose about 470% on debut. Counterpoint put its share of global DRAM revenue at 10% in the second quarter of 2026, up from 4% a year earlier. Its gross margin reached 89.53% in the third quarter of 2026, up from 87.59% in the second quarter, a level covered in TrendForce's report on CXMT's gross margin.
The margin needs context. SemiAnalysis estimates that CXMT's DDR5 cost per bit is still over 30% higher than the three leaders, and that the margin gain is driven mainly by pricing. In other words, the number reflects what memory sells for today, not how cheaply CXMT makes it.
Short term: a large beneficiary of shortage pricing. Long term: the most exposed. A higher-cost producer with a technology gap has no cushion when prices fall.
Sandisk (SNDK)
Sandisk is a NAND specialist. NAND is storage memory, a different product from DRAM. It is moving its business onto multi-year supply agreements with data center customers, and management expects those agreements to cover more than half of its bits in fiscal 2027 and about two-thirds in fiscal 2028. Micron has warned that memory supply stays tight into 2028. Our thesis is that NAND supply stays tight as rivals prioritize DRAM and HBM.
Short term: much of its output contracted ahead, with pricing power. Long term: dependent on NAND staying constrained. The robot DRAM thesis does not apply to it directly.
One risk hangs over all five. Every maker is responding to the shortage by building, and those fabs open late in the decade. If they arrive together, they could flood the market and push prices down, exactly as happened before 2023. Robot demand is the possible offset: a new growth engine arriving at about the time the new supply does. Whether it is large enough by then is unknown.
Investors who do not want to choose between individual makers sometimes use a diversified route instead. Our SMH semiconductor ETF analysis on CleaRank covers one such fund.
Five Memory Makers: Short Term Versus Long Term
Who is paid today, and who is positioned for the robot era.
-
01
SK hynix (SKHY)
- Now
- About half of the HBM market (50% in Q2 2026, Counterpoint), locked in by long-term contracts
- Later
- Contracts cushion a standard DRAM downturn; less standard-DRAM volume than Samsung
Outlook STRONG NOW -
02
Samsung
- Now
- Full mix: HBM share rising to about 33%, NAND shortage, largest standard DRAM capacity
- Later
- Most standard DRAM capacity if robot demand arrives
Outlook STRONG BOTH -
03
Micron (MU)
- Now
- Gains from price rises across every memory type
- Later
- New US fabs: Idaho from 2027, New York around 2030, a strategic domestic supply
Outlook STRONG BOTH -
04
CXMT (Shanghai STAR)
- Now
- Gross margin near 90% (89.5% in Q3) selling standard DRAM into the shortage
- Later
- Cost per bit still over 30% above the leaders; margin rests on price, not technology
Outlook STRONG NOW, EXPOSED LATER -
05
Sandisk (SNDK)
- Now
- NAND storage specialist while rivals move clean rooms to DRAM and HBM
- Later
- Pricing power lasts as long as rivals stay away from NAND
Outlook STRONG NOW
Short Term: HBM Contracts Pay. Long Term: Whoever Can Make the Most Standard DRAM, Cheaply.
CXMT's near-90% gross margin is the highest of the group and the least durable. It comes from shortage pricing on a product where its costs are higher than its rivals'. When new fabs open later this decade and prices fall, the makers with the lowest cost per bit and contracted demand keep their profits. The rest give them back.
Source: CleaRank analysis; Counterpoint Research, IDC via SK hynix filings, TrendForce, SemiAnalysis, company reports. Editorial assessment, not a rating. November 2026.
What Does the Anthropic (ANT) IPO Mean for Memory Stocks?
It adds a very large, newly funded buyer to the demand side. Anthropic does not make memory. It buys compute, and memory is a growing share of what compute costs.
The facts first. Anthropic, PBC priced its IPO at $83 a share on November 18, 2026, above the $70 to $78 range. The deal was upsized to 980 million shares from 825 million and raised about $81.3 billion, the largest IPO on record, at a valuation of about $1.5 trillion at pricing, as reported in CNBC's coverage of the Anthropic IPO pricing. The shares trade on the Nasdaq Global Select Market under the ticker ANT.
ANT opened on November 19 at $111.50, traded between $92.72 and $115.10, and closed at $97.17, up 17% from the IPO price. CNBC put the fully diluted value at about $1.8 trillion at that close. In its second session, on November 20, ANT fell 8.5% to $88.91 after dipping as low as $83.05, close to the IPO price, on a day the wider market sold off. Current figures are on the Anthropic (ANT) stock tracker on CleaRank.
The S-1 says proceeds are for general corporate purposes, including compute and infrastructure. The scale of that spending is the link to memory. According to Fortune's analysis of the Anthropic S-1, the company plans future compute and infrastructure spending of $542 billion to $567 billion. The amended filing of November 17 added the third quarter, in which compute and infrastructure cost $9.86 billion against revenue of $16.96 billion.
Two more details from the filing matter here. Compute was 48% of revenue in the second quarter of 2026 and about 58% in the third, and the filing says agentic workloads consume significantly more compute per interaction. The compute agreements also carry take-or-pay terms, meaning the company pays whether or not it uses the capacity.
Our analysis of the link runs as follows. Capital raised by AI labs flows into compute. Memory is a growing share of compute cost. So IPO money ultimately supports demand for HBM. The same link works in reverse as a risk: high memory prices raise the cost of compute, which is a cost risk for AI labs, including this one.
ANT is not a memory stock. It is the customer. Anyone looking at it for memory exposure is looking at the buyer's side of the table.
Who Buys the Memory: Today's Customer and Tomorrow's
Pink nodes are today's AI data center demand. Green nodes are the robot demand that may follow.
- TODAY
Anthropic (ANT)
Raised about $81.3B in the largest IPO on record; proceeds go to compute and infrastructure
- TODAY
Cloud Giants
Long-term contracts for HBM with SK hynix, Samsung and Micron
- TODAY
AI Compute Commitments
Anthropic alone reports over $540B of future compute spending
- TODAY
HBM
Stacked memory beside AI chips; 2027 demand seen far above 2026
Global Memory Supply
ONE SET OF CLEAN ROOMS, TWO KINDS OF DEMAND
- LATER
Standard DRAM
Low-power memory for phones, PCs and, later, robots
- LATER
Humanoid Robots
DRAM demand seen rising about 20 times by 2030, from a tiny base
- LATER
On-Board Memory
A robot acts in real time and cannot wait for the cloud
- LATER
Tesla Optimus
Musk: robot RAM needs will exceed everything else on Earth
A large share of what AI labs raise is spent on compute, and a growing share of compute cost is memory. That is the link between the record Anthropic IPO and the memory shortage: today's buyer is the data center. The robot buyer is real but small for now, about 0.05% of global DRAM capacity by one estimate, so it is a story for the end of the decade.
Source: CleaRank analysis; Anthropic S-1 filings as reported by Fortune, CNBC, Counterpoint Research, TrendForce. Data as of November 20, 2026.
“An $81 billion raise by a company whose largest cost is compute is, one step removed, an order for memory. But the same filing that excites memory bulls also warns of take-or-pay commitments. If AI demand disappoints, the cancellation risk runs straight back to the memory makers.”
Jacob Bakshi, CleaRank Senior Market Strategist
Memory Market Scenarios: Bear, Base and Bull
The table below sets out three ways the rest of the decade could develop. These are illustrative scenarios, not forecasts, and they contain no price targets.
Scenario | What happens | HBM leaders | Standard DRAM makers |
|---|---|---|---|
Bear | AI spending slows, new fabs open into weak demand, prices fall sharply | Contracts cushion but volumes get renegotiated | Hit hardest; high-cost producers lose margin first |
Base | Shortage eases around 2028 as capacity arrives; robots grow from a small base | Earnings stay strong, growth slows | Prices normalize; low-cost scale producers do best |
Bull | AI demand stays ahead of supply and robot volumes ramp late in the decade | Sold out for years | A second shortage in standard low-power DRAM |
Bear. AI spending slows and the new fabs open into weak demand. Prices fall sharply. Long-term contracts cushion the HBM leaders, but volumes get renegotiated. Standard DRAM makers are hit hardest, and high-cost producers lose margin first.
Base. The shortage eases around 2028 as capacity arrives, and robots grow from a small base. HBM leaders keep strong earnings with slower growth. Standard DRAM prices normalize and low-cost producers with scale do best.
Bull. AI demand stays ahead of supply and robot volumes ramp late in the decade. HBM leaders are sold out for years, and a second shortage appears in standard low-power DRAM.
What Could Go Wrong?
Several things, and the most important one is that memory has been here before.
The cycle. Memory is one of the most cyclical industries in technology. 2023 is the reminder: oversupply, collapsing prices and frozen investment came only three years before the current shortage. High prices fund the capacity that ends them.
China and export policy. CXMT is growing quickly, and policy on either side could change what it can buy, make or sell. That affects global supply and the pricing every other maker depends on.
Robot timing. The robot thesis rests on projections for 2030 and a current base of about 0.054% of DRAM capacity. Humanoid robots may arrive later, in smaller numbers, or with less memory each than forecast.
AI capital spending. HBM demand depends on AI labs and cloud companies continuing to spend. Take-or-pay commitments show how much has been promised, and also how much could be renegotiated if demand disappoints.
ANT is days old. The stock has traded for two sessions and has already moved sharply in both directions. Its price says little yet about long-term AI demand, and nothing about memory.
Today's Shortage Is in the Premium Chip. Tomorrow's May Be in the Basic One.
The memory shortage of 2026 has a clear cause. An industry that stopped building in 2023 is now sending its limited clean room space to HBM, a product that uses roughly 2 to 4 times the wafer area per bit, and the plain DRAM in PCs and phones has become scarce and expensive as a result. Most forecasts do not see supply catching up before 2028.
The twist is about what comes after. A machine that works in the physical world needs its memory on board, and without mains power or cooling it cannot carry the premium chip. If robots scale, they will buy standard low-power DRAM. The numbers are tiny today, about 0.054% of capacity even on a 2030 projection, so this is a question of direction and not of this year's earnings. It matters because the fabs being planned now will open into that world.
Meanwhile, the largest IPO on record has just put about $81.3 billion in the hands of a company whose biggest cost is compute. That supports memory demand today and concentrates risk if AI spending slows. The scenarios above are illustrations, not forecasts, and nothing in this article is investment advice.
Frequently Asked Questions
Why is there a memory chip shortage?
A memory glut in 2023 collapsed prices and froze investment in new production lines, and clean rooms take years to build. Makers then shifted wafers from standard DRAM to HBM for AI data centers, which needs roughly 2 to 4 times the wafer area per bit. TrendForce reports conventional DRAM contract prices rose roughly 93% to 98% in Q1 2026 and forecasts further increases in Q4 2026.
What is the difference between HBM and DRAM?
Standard DRAM is the simple, power-efficient working memory used in PCs and phones. HBM, or high bandwidth memory, is stacked memory built for AI data centers. It is far more complex, runs hotter, uses more power and takes roughly 2 to 4 times the wafer area per bit.
Will robots really need that much memory?
Possibly, but not soon. Counterpoint Research projects humanoid robot DRAM demand rising about 20 times by 2030, from about 8,300 terabits to about 172,000 terabits. By one analysis, that 2030 figure is still about 0.054% of current global DRAM capacity. It is a story for after 2028.
Which company leads in HBM?
SK hynix. Counterpoint put its HBM share at 50% in Q2 2026, with Samsung at 33% and Micron at 18%. IDC data cited in an SEC filing by SK hynix gave it 56.4% of HBM revenue in Q1 2026.
Why is CXMT's margin so high?
CXMT reported a gross margin of 89.53% in Q3 2026, up from 87.59% in Q2. SemiAnalysis says the gain is driven mainly by pricing, and that its DDR5 cost per bit is still over 30% higher than the three leaders. The margin reflects shortage prices, not a cost advantage.
What is the ANT stock IPO?
ANT is the Nasdaq ticker of Anthropic, PBC. The company priced its IPO at $83 a share on November 18, 2026, sold 980 million shares and raised about $81.3 billion, the largest IPO on record. On its first day, November 19, the stock opened at $111.50 and closed at $97.17, up 17%.
Where can I read more on CleaRank?
For the newly listed AI lab, see the Anthropic (ANT) Stock Tracker on CleaRank. For a memory maker, read the Micron (MU) Stock Analysis on CleaRank. For a diversified view of the chip sector, see the VanEck Semiconductor ETF (SMH) Analysis on CleaRank.
Disclaimer: The memory industry is highly cyclical, and prices, margins and supply conditions can reverse quickly. Forecasts and market-share estimates cited in this article differ by source and may change. Projections of memory demand from robots are speculative, and the thesis that robots will favor standard DRAM is CleaRank analysis, not an established fact. Newly listed shares such as ANT are especially volatile, and the price figures cited reflect only the first days of trading. Chinese-listed shares carry access and policy risk for foreign investors. The scenarios described are illustrative, not forecasts. This article is for informational purposes only and does not constitute financial, legal or investment advice. CleaRank and its analysts may hold positions in securities mentioned in this article. Past performance is not indicative of future results. Always consult with a licensed financial advisor before making investment decisions.
I’m the co-founder of CleaRank. I’ve spent most of my life studying finance, from analyzing market trends to spotting early adoption in the crypto industry, and I’ve learned that markets move on the psychology and sentiment of the crowd as much as on the numbers. Over my career I’ve built trading algorithms, managed long-term investment portfolios and helped traders avoid shady brokers before they knew they were at risk. Today I write market analysis on stocks, ETFs, forex and crypto, with a focus on AI infrastructure and the companies behind it.
