Why is there a chip shortage again in 2026?

AI is buying memory chips faster than they can be made, and making more takes machines and factories that need years to build. In the US, a 32GB memory kit for a PC now costs about three times what it did a year ago.

This guide follows the HanaHow episode, which draws on Dwarkesh Patel’s podcast What will computer chips look like in 2045? (22 October 2026), where Intel’s chief technology officer, Pushkar Ranade, talks about how chipmakers cope with demand from AI. The interview was recorded around the middle of 2026, so some of its figures have since moved.

Figures are as of 23 October 2026. Memory prices change every week, so check before you buy.

What memory costs now

Tom’s Hardware’s price tracker says a standard 32GB DDR5 memory kit, two 16GB sticks, “sold for around $100 to $200 in October 2025”. Its table, last updated on 22 October 2026, lists the best US price for a 32GB DDR5-6000 kit at $447, against a lowest-ever price of $72, and a 32GB DDR5-5600 kit at $479. So the cheapest kits cost two to four and a half times what they did a year ago, or about three times the middle of that range.

Older DDR4 memory hasn’t escaped: “We’ve seen the pricing double to triple,” Tom’s Hardware says.

In the UK, Club386 followed one G.Skill Trident Z5 32GB kit that its writer bought for £116.66 in January 2025. It cost £499.99 by that December, and in August 2026 it was out of stock at most UK shops, with new stock listed at £1,199.99 at Overclockers.

Storage has gone up too. In January, Sandisk’s own store listed a WD Black SN7100 2TB SSD at $434.99, against a list price of $149.99, Tom’s Hardware reported.

What PC makers paid

Gartner said on 12 October 2026 that “DRAM costs surged five times year-over-year in the second quarter”. That’s what computer makers paid for memory chips between April and June, against the same months a year before.

TrendForce, which tracks the contract prices big buyers pay, counted rises of 45 to 50% for ordinary DRAM memory in the last three months of 2025 and 93 to 98% in the first three months of 2026. Put those together with its forecasts for the rest of 2026, rises of 58 to 63%, then 13 to 18%, then 10 to 15%, and prices end 2026 at about six times their level in July to September 2025, though the last three steps are forecasts.

Fewer computers and phones are going out. Between July and September, PC shipments fell 20.1% on a year before, IDC says, or 22.4% by Gartner’s count. Counterpoint says smartphone shipments fell 20%, “the lowest Q3 level since 2012”, and that phone makers have raised retail prices by 20% so far this year. These are shipments from makers, not sales to people, and IDC notes that retailers stocked up earlier in the year, ahead of price rises.

The price rises passed on

Some of the biggest names have raised prices and pointed at memory.

Product and date	UK	US
Apple 13-inch MacBook Air, 25 June 2026	£1,099 to £1,299	$1,099 to $1,299
Apple MacBook Neo, 25 June 2026	£599 to £699	$599 to $699
Apple 11-inch iPad Air, 25 June 2026	£599 to £749	$599 to $749
Xbox Series S 512GB, 1 August 2026	£299.99 to £429.99	$399.99 to $499.99
Xbox Series X with disc drive, 1 August 2026	£499.99 to £669.99	$649.99 to $799.99
Raspberry Pi 5 16GB, December 2025 to April 2026	Priced in dollars	$120 to $305

• Apple said: “We have never seen a component price increase this much, this quickly.” Reuters reported the rise as down to memory costs.
• Microsoft said “console storage and memory prices have increased by more than 2.5x” and that it expects “another doubling by the fall of 2027”.
• Raspberry Pi, a Cambridge company that sets its prices in dollars, has raised prices four times since December 2025. The 16GB Pi 5 went to $145 on 1 December 2025, $205 on 2 February and $305 on 1 April. It pointed to “a seven-fold increase over the last year in the price of the LPDDR4 DRAM used on Raspberry Pi 4 and 5”.

Nintendo raised the Switch 2 from £395.99 to £419.99, and $449.99 to $499.99, on 1 September, citing “various changes in market conditions”. Reuters had reported that higher component costs, particularly memory, and tariffs would add about ¥100 billion to its costs this financial year. Not every company names memory: Sony’s statement on April’s PS5 rise blamed pressures in the global economy, though Reuters reported in May that Sony expects memory prices to stay high next year.

In the UK, Alex Baldock, then chief executive of Currys, told reporters in July that AI and data centres were eating up the world’s supply of silicon, leaving less for phones and laptops, with “some cost price inflation coming through later this year”.

How the 2020 to 2021 shortage was different

The last chip shortage started with cancelled orders. “With panic, lockdowns, and general uncertainty rolling across the globe, automakers cancelled orders,” IEEE Spectrum wrote of 2020, while people stuck at home bought computers, TVs and games consoles. When car makers came back, the chips they needed had gone to other buyers.

The chips that ran short were cheap, older ones. A US Commerce Department survey found buyers’ median stock had fallen from 40 days in 2019 to under five in 2021 in some key US industries, worst for microcontrollers and analog chips made on old processes, and “the primary bottleneck across the board appears to be wafer production capacity.”

S&P Global Mobility reckons the shortage cost more than 9.5 million cars and other light vehicles in 2021, and another 3 million in 2022. That’s production lost against what would have been built: world vehicle production in 2021 was still about 3% up on 2020, according to the carmakers’ body OICA.

This time the demand comes from AI, and memory makers are putting it first. That has made even older kinds of memory scarce, such as the LPDDR4 in a Raspberry Pi and the DDR4 in older PCs.

Car chips in late 2025: Nexperia, not AI

Car makers had their own scare a year ago, and it was about politics. On 30 September 2025 the Dutch government took control of Nexperia, a chipmaker owned by China’s Wingtech. On 4 October China restricted exports of Nexperia chips made in China. They’re cheap chips for things like wipers and power windows.

Honda stopped its factory in Celaya, Mexico, on 28 October 2025 and halved output in Canada. China granted exemptions for civilian-use Nexperia chips on 9 November, and Honda said its North American plants would restart gradually from late November. Honda said the shortage would cut its operating profit by ¥150 billion, and in early January 2026 it halted two plants in Japan and extended a halt at three in China.

The AI squeeze is now reaching cars through older memory: Counterpoint said in August that “DDR4/LPDDR4 tightness for automotive customers will extend beyond 2027”. We found no confirmed reports of large car production cuts caused by memory in 2026.

Where the demand from AI comes from

Epoch AI, a research group that tracks AI, estimates that the computing power of all the AI chips sold has grown about 3.3 times a year since 2022, doubling about every seven months, with a likely range of 2.7 to 4.1 times. It counts sales from Nvidia, Google, AMD, Amazon, Huawei and others, measured against Nvidia’s H100 chip.

In the podcast, Patel says “the AI industry for a decade has been 3Xing compute year over year”. Epoch’s chip figure only goes back to 2022. Over a decade, the figure that fits is the computing used to train notable AI models, which Epoch puts at about 4.1 times a year from 2010 to 2024, and sums up as four to five times.

Nvidia’s quarterly filing for the three months to 26 July 2026 shows $279 billion of supply commitments, up from $119 billion three months before, “primarily memory and manufacturing facilities”. That’s what it has signed up to buy, not money already paid, and the filing says some of it “may be cancelable, rescheduled, or adjustable”.

Ranade told Patel: “the world will eat up as much compute as can possibly be made, and it’ll eat up as much memory as can possibly be made. I would argue that the supply crunch is larger for memory today than it is for compute.”

Why HBM leaves less memory for everything else

AI chips sit beside stacks of a special memory called HBM, or high-bandwidth memory. Micron’s chief executive, Sanjay Mehrotra, said in March 2025 that “HBM3E consumes three times the amount of silicon compared to D5 to produce the same number of bits”. D5 is DDR5, the memory in a new PC, and he said the ratio would pass four to one with later HBM. So making more HBM takes factory space from ordinary memory: TrendForce wrote in March that “DRAM suppliers are reallocating capacity toward HBM and server applications”.

The three big memory makers are making record profits:

• Samsung expects an operating profit of about 107.4 trillion won, roughly $80 billion, for July to September 2026, against 12.17 trillion a year earlier.
• SK hynix made an operating profit of 60.54 trillion won on sales of 79.32 trillion between April and June, a 76% margin.
• Micron had record revenue of $54.23 billion in the three months to 3 September, against $11.32 billion a year before.

They buy chipmaking machines too: memory made up 46% of ASML’s system sales between July and September.

Micron’s exit from Crucial

On 3 December 2025, Micron said it would stop selling its Crucial memory and SSDs to the public, citing “AI-driven growth in the data center”. It said it was leaving “in order to improve supply and support for our larger, strategic customers in faster-growing segments”. Shipments were planned to end in February 2026, warranty support carries on, and Micron still sells memory to businesses.

Intel left memory earlier. It sold its flash memory business to SK hynix in a deal agreed in October 2020 and completed in March 2025. Ranade: “Clearly, in hindsight, probably not the smartest time to sell off a flash memory business.”

Why the machines hold everything back

So why not make more? Every maker of advanced chips, memory makers included, needs the same machines. Only one company, ASML in the Netherlands, makes the extreme ultraviolet (EUV) lithography machines that print the finest patterns on a chip. Our guide to how computer chips are made so small explains how they work.

ASML’s 2025 annual report says it sold 44 standard EUV machines in 2025, and four of its newest High-NA ones, up from 42 and two in 2024. On 14 October 2026 it said it expects to ship about 65 standard machines in 2026. It plans at least 80 in 2027 and about 30% more capacity in 2028, and its chief executive, Christophe Fouquet, said that “our supply as currently planned for both 2027 and 2028 will not fully meet the demand of our customers”. Reuters, citing JPMorgan analysts who spoke to ASML’s finance chief, reported in September that ASML is exploring how to make more than 110 a year by 2028. ASML hasn’t announced that figure.

In the podcast, Patel says ASML “produces fewer than 100 of them a year” and doesn’t plan to produce more than 100 a year “even by the end of the decade”. The first part still holds. The second no longer fits: ASML’s own plan reaches just over 100 in 2028, and it says it is looking at ways to go beyond that.

ASML doesn’t publish prices. Yole Group put a standard machine at around €170 million ($183 million) in 2024, and Reuters put the High-NA version at about €350 million ($370 million). Ranade says “a High-NA EUV scanner and all its associated equipment costs like $400 million”.

EUV isn’t the only hold-up. “EUV or ASML is not the only bottleneck,” Ranade says: “we’re seeing all equipment seeing tremendous supply shortage.” And a factory can’t easily switch suppliers. Once a process is set, he says, “we need a specific tool. We can’t just go and get a different tool for the same process.”

Patel asked whether there are enough tools in the world for Intel to make 10 times as many leading-edge wafers. Ranade: “Not 10X, but probably 2X, yes.” That was about Intel’s own output, not the whole industry’s.

How long new tools and factories take

Ranade: “Let’s say you place an enormous order for tools today, and you’re looking at 18 months to get those tools. You’re already into basically ’28. From that point on, if you have to build a new shell, build a new fab, you’re basically looking at the end of the decade.”

Georgetown University’s Center for Security and Emerging Technology (CSET) studied 635 new chip factories built between 1990 and 2020. On average they took 682 days, about 1.86 years, from the start of construction to the start of production, not counting planning and permits. The Americas averaged 736 days, and in the US the average rose from 665 days in the 1990s to 918, about two and a half years, in 2010 to 2020. Most of those factories don’t make cutting-edge chips.

TSMC announced its first Arizona factory in May 2020, planning to start building in 2021, and reached high-volume production at the end of 2024, about four and a half years from announcement. Micron’s first New York memory factory isn’t expected to make wafers until 2030.

Fewer companies can make the newest chips at all. In 2001, 18 chipmakers had factories that could make the most advanced chips of the day, according to the consultancy IBS. Now there are three: TSMC, Samsung and Intel. IBS estimates that a 2 nm factory making 50,000 wafers a month costs about $28 billion, against about $20 billion for 3 nm, mainly because of the extra EUV machines. Ranade says the cost of building a factory has been “doubling every 4-5 years”.

The money is going in. TSMC plans to spend $60 billion to $64 billion on factories and tools in 2026, up from the $52 billion to $56 billion it planned in January. Intel expects to spend “more than $20 billion”.

Racing the growth of AI against the machines

The slider runs those growth rates forward from 2025 to 2035: the chips AI would want if its demand for computing kept growing, against the chips ASML’s machines could make if their output kept growing. A third slider lets each chip do more computing every year, which cuts the number of chips AI would want.

At the sourced figures, the gap widens. Demand for computing grows 3.3 times a year. Each chip does about 1.35 times more a year, so the number of chips AI would want grows about 2.44 times a year. The machines grow about 1.3 times a year. So the gap grows about 1.88 times a year, passes 2 times in 2027 and 10 times in 2029, and reaches about 550 times by 2035. For the machines to keep up, their output would have to grow as fast as the number of chips AI wants.

Where the starting figures come from:

• Demand for computing, 3.3 times a year: Epoch AI’s estimate of how fast the computing power of AI chips sold has grown since 2022.
• EUV machines made, 1.3 times a year: growing from ASML’s expected 65 standard machines in 2026 to the reported aim of more than 110 by 2028 takes 1.3 times a year, though the report doesn’t say whether the 110 counts only standard machines. ASML’s own plan, at least 80 in 2027 and about 30% more in 2028, works out at about 1.26.
• Computing per chip, 1.35 times a year: Epoch AI found the computing performance of 47 AI chips from 2010 to 2023 doubled every 2.3 years. That’s at one level of precision; moving AI work to smaller number formats has given chips bigger one-off jumps.

It isn’t a forecast. A gap of hundreds of times can’t actually open, because AI can’t use chips that were never made. AI’s computing has outgrown the machines so far partly by taking a bigger share of what factories make: Patel says most of TSMC’s 3 nm production now goes to AI chips, and a share can’t grow past all of it. When demand outruns supply, prices rise, buyers wait and AI’s growth slows to what can be made, or more machines and factories arrive. As Ranade told Patel, from the equipment suppliers down to their sub-suppliers, “unless everybody produces more, you can’t increase the output.”

Terafab

Some buyers are building their own. In Austin, Texas, on 21 March 2026, Elon Musk announced Terafab: two advanced chip factories at one site, Reuters reported, making chips for Tesla’s “cars and humanoid robots” and SpaceX’s “AI satellites in space”. In a clip played in the episode, Musk says: “We either build the Terafab or we don’t have the chips. And we need the chips, so we’re going to build the Terafab.”

Intel said on 7 April that it had joined, bringing its chip design, manufacturing and packaging, though its exact role hasn’t been set out.

Terafab still needs ASML’s machines. On SpaceX’s results call, reported on 21 October, Musk said SpaceX is aiming for volume production by the third or fourth quarter of 2028, Bloomberg reported. He called the dates “very aggressive” and thanked ASML and his other suppliers for speeding up equipment deliveries. That’s Musk’s account: ASML hasn’t confirmed it.

When prices may come down

The companies that make the memory don’t expect it soon.

• Micron, 30 September 2026: “we expect memory and storage supply-demand conditions to be much tighter in fiscal 2027 and 2028 than they were in 2026”, and “we do not have line of sight to when supply and demand will return to balance.”
• SK hynix’s chief executive, Kwak Noh-jung, told Reuters in July: “We forecast that next year will be the worst year in the industry’s history from the supply perspective.” He said “customer demand will remain higher than our supply capacity even beyond 2030.”
• Samsung said on its results call in July that it expects the shortage to get worse in 2027, with supply staying tight until at least 2028, TechCrunch reported.

Their new factories mostly open in 2027 and 2028. Micron expects wafers from a new Idaho factory in mid-2027, from a second one in late 2028, from its expansion in Japan in the second half of 2028, and from New York in 2030. SK hynix brought its first Yongin factory in South Korea forward to February 2027, Reuters reported, and Samsung’s P5 at Pyeongtaek is reported to be in full operation in 2028.

The analysts don’t expect relief before then. TrendForce expects DRAM prices to keep rising in 2027, with “substantial output contributions” from new factories “not expected to materialize until 2028”. IDC said in June it sees no relief “before the end of 2027”, and that “pricing is unlikely to return to 2025 levels.”

Storage may ease first. TrendForce expects NAND flash, the kind in SSDs, to move to “a looser supply environment in 2H27”, with “increasing downward pressure on prices”. For now, prices are still going up: it forecasts rises of 10 to 15% for DRAM and 15 to 20% for NAND contracts in the last three months of 2026. NAND wafer spot prices slipped 1.28% in the week to 19 October, while DDR5 supply stayed tight, TrendForce said.

Raspberry Pi said in October that it looks forward to reversing its own price rises once memory costs come down.

These figures are as of 23 October 2026. Prices change every week, and more results from memory and chip makers are due at the end of October.

Sources

• Dwarkesh Patel: What will computer chips look like in 2045?, Dwarkesh Podcast, 22 October 2026, also on YouTube. The quotes from Pushkar Ranade and Patel are from its official transcript.
• Shop prices: Tom’s Hardware’s RAM price index (last updated 22 October 2026) and Sandisk’s SSD price rises (31 January 2026), and Club386 on one G.Skill kit’s price (18 August 2026).
• What manufacturers pay: Gartner (12 October 2026), TrendForce on DRAM contract prices for late 2025, early 2026, April to June, July to September and October to December, IDC (8 October 2026) and Counterpoint (12 October 2026).
• Price rises: Reuters on Apple and Which? on Apple’s UK prices (June 2026), Xbox Wire and Pure Xbox’s UK and US price table, Raspberry Pi on its rises of December 2025, February, April and October, Reuters on Nintendo and Sony (8 May 2026), Sony’s PS5 price statement (27 March 2026), Nintendo Life on the Switch 2’s UK price, HWBusters on Nintendo’s notice and Reuters on Currys (2 July 2026).
• The 2020 to 2021 shortage: IEEE Spectrum on how it started, The Register on the US Commerce Department’s findings (25 January 2022), S&P Global Mobility (13 July 2023) and OICA’s 2021 production table (archived).
• Nexperia and Honda: Reuters on China’s export curbs and exemptions, Bloomberg on the Honda stoppages (30 October 2025), CBT News on the restart, Dow Jones on Honda’s forecast, Automotive World on the January halts and Counterpoint on memory for cars (22 August 2026).
• Demand from AI: Epoch AI’s Global AI computing capacity is doubling every 7 months (9 January 2026), Training compute of frontier AI models grows by 4-5x per year (28 May 2024) and trends dashboard, Nvidia’s 10-Q for the quarter to 26 July 2026 and TrendForce on Nvidia’s commitments (27 August 2026).
• HBM and memory makers: Blocks & Files on Micron’s HBM silicon figure (21 March 2025), TrendForce on capacity moving to HBM, Samsung’s guidance (8 October 2026) and Pulse 2.0’s dollar conversion, SK hynix’s results (29 July 2026) and Blocks & Files on Micron’s results (1 October 2026).
• Crucial: Micron’s announcement, reprinted by StorageNewsletter, and StorageReview (3 December 2025). Intel’s flash sale: Intel’s announcement (20 October 2020) and DigiTimes on its completion (28 March 2025).
• ASML: its 2025 annual report, MarketBeat’s summary of its third-quarter 2026 call and TrendForce on the same call (14 October 2026), Seeking Alpha on its 2027 plan, Evertiq on Fouquet’s comments (15 October 2026), Techzine on Reuters’ report of more than 110 by 2028 (15 September 2026) and Yole Group on machine prices, including Reuters’ High-NA figure.
• Factories: CSET’s No Permits, No Fabs (John VerWey, October 2021), TSMC’s Arizona announcement (15 May 2020) and the Arizona Tech Council on its high-volume start, Evertiq on Micron’s new factories (6 October 2026), Semiconductor Engineering on IBS’s count of leading-edge chipmakers, the Hoover Institution’s Silicon Triangle, TechSpot on IBS’s factory costs, Focus Taiwan on TSMC’s spending (15 October 2026), TrendForce on its January plan (15 January 2026) and MarketBeat on Intel’s (23 July 2026).
• Computing per chip, for the slider: Epoch AI’s Trends in machine learning hardware (Hobbhahn, Heim and Aydos, 9 November 2023).
• Terafab: Reuters on the announcement (22 March 2026), Evertiq on Intel joining (7 April 2026) and Bloomberg, via The Edge Malaysia, on Musk’s latest comments (21 October 2026).
• When prices may ease: Reuters on SK hynix’s outlook (11 July 2026) and its Yongin factory (15 January 2026), Korea JoongAng Daily on Samsung’s P5 (27 September 2026), TechCrunch on Samsung (31 July 2026), TrendForce’s 2027 outlook (30 July 2026), IDC (2 June 2026) and TrendForce on spot prices (21 October 2026).