LEDs Enter the Nanoscale

MicroLEDs, with pixels just micrometers across, have long been a byword in the display world. Now, microLED-makers have begun shrinking their creations into the uncharted nano realm. In January, a startup named Polar Light Technologies unveiled prototype blue LEDs less than 500 nanometers across. This raises a tempting question: How far can LEDs shrink?

We know the answer is, at least, considerably smaller. In the past year, two different research groups have demonstrated LED pixels at sizes of 100 nm or less.

These are some of the smallest LEDs ever created. They leave much to be desired in their efficiency—but one day, nanoLEDs could power ultra-high-resolution virtual reality displays and high-bandwidth on-chip photonics. And the key to making even tinier LEDs, if these early attempts are any precedents, may be to make more unusual LEDs.

New Approaches to LED

Take Polar Light’s example. Like many LEDs, the Sweden-based startup’s diodes are fashioned from III-V semiconductors like gallium nitride (GaN) and indium gallium nitride (InGaN). Unlike many LEDs, which are etched into their semiconductor from the top down, Polar Light’s are instead fabricated by building peculiarly shaped hexagonal pyramids from the bottom up.

Polar Light designed its pyramids for the larger microLED market, and plans to start commercial production in late 2026. But they also wanted to test how small their pyramids could shrink. So far, they’ve made pyramids 300 nm across. “We haven’t reached the limit, yet,” says Oskar Fajerson, Polar Light’s CEO. “Do we know the limit? No, we don’t, but we can [make] them smaller.”

Elsewhere, researchers have already done that. Some of the world’s tiniest LEDs come from groups who have foregone the standard III-V semiconductors in favor of other types of LEDs—like OLEDs.

“We are thinking of a different pathway for organic semiconductors,” says Chih-Jen Shih, a chemical engineer at ETH Zurich in Switzerland. Shih and his colleagues were interested in finding a way to fabricate small OLEDs at scale. Using an electron-beam lithography-based technique, they crafted arrays of green OLEDs with pixels as small as 100 nm across.

Where today’s best displays have 14,000 pixels per inch, these nanoLEDs—presented in an October 2025 Nature Photonics paper—can reach 100,000 pixels per inch.

Another group tried their hands with perovskites, cage-shaped materials best-known for their prowess in high-efficiency solar panels. Perovskites have recently gained traction in LEDs too. “We wanted to see what would happen if we make perovskite LEDs smaller, all the way down to the micrometer and nanometer length-scale,” says Dawei Di, engineer at Zhejiang University in Hangzhou, China.

Di’s group started with comparatively colossal perovskite LED pixels, measuring hundreds of micrometers. Then, they fabricated sequences of smaller and smaller pixels, each tinier than the last. Even after the 1 μm mark, they did not stop: 890 nm, then 440 nm, only bottoming out at 90 nm. These 90 nm red and green pixels, presented in a March 2025 Nature paper, likely represent the smallest LEDs reported to date.

Efficiency Challenges

Unfortunately, small size comes at a cost: Shrinking LEDs also shrinks their efficiency. Di’s group’s perovskite nanoLEDs have external quantum efficiencies—a measure of how many injected electrons are converted into photons—around 5 to 10 percent; Shih’s group’s nano-OLED arrays performed slightly better, topping 13 percent. For comparison, a typical millimeter-sized III-V LED can reach 50 to 70 percent, depending on its color.

Shih, however, is optimistic that modifying how nano-OLEDs are made can boost their efficiency. “In principle, you can achieve 30 percent, 40 percent external quantum efficiency with OLEDs, even with a smaller pixel, but it takes time to optimize the process,” Shih says.

Di thinks that researchers could take perovskite nanoLEDs to less dire efficiencies by tinkering with the material. Although his group is now focusing on the larger perovskite microLEDs, Di expects researchers will eventually reckon with nanoLEDs’ efficiency gap. If applications of smaller LEDs become appealing, “this issue could become increasingly important,” Di says.

What Can NanoLEDs Be Used For?

What can you actually do with LEDs this small? Today, the push for tinier pixels largely comes from devices like smart glasses and virtual reality headsets. Makers of these displays are hungry for smaller and smaller pixels in a chase for bleeding-edge picture quality with low power consumption (one reason that efficiency is important). Polar Light’s Fajerson says that smart-glass manufacturers today are already seeking 3 μm pixels.

But researchers are skeptical that VR displays will ever need pixels smaller than around 1 μm. Shrink pixels too far beyond that, and they’ll cross their light’s diffraction limit—that means they’ll become too small for the human eye to resolve. Shih’s and Di’s groups have already crossed the limit with their 100-nm and 90-nm pixels.

Very tiny LEDs may instead find use in on-chip photonics systems, allowing the likes of AI data centers to communicate with greater bandwidths than they can today. Chip manufacturing giant TSMC is already trying out microLED interconnects, and it’s easy to imagine chipmakers turning to even smaller LEDs in the future.

But the tiniest nanoLEDs may have even more exotic applications, because they’re smaller than the wavelengths of their light. “From a process point of view, you are making a new component that was not possible in the past,” Shih says.

For example, Shih’s group showed their nano-OLEDs could form a metasurface—a structure that uses its pixels’ nano-sizes to control how each pixel interacts with its neighbors. One day, similar devices could focus nanoLED light into laser-like beams or create holographic 3D nanoLED displays.

Reference: https://ift.tt/aZz4Xvs

What the FDA’s 2026 Update Means for Wearables

As new consumer hardware and software capabilities have bumped up against medicine over the last few years, consumers and manufacturers alike have struggled with identifying the line between “wellness” products such as earbuds that can also amplify and clarify surrounding speakers’ voices and regulated medical devices such as conventional hearing aids. On January 6, 2026, the U.S. Food and Drug Administration issued new guidance documents clarifying how it interprets existing law for the review of wearable and AI-assisted devices.

The first document, for general wellness, specifies that the FDA will interpret noninvasive sensors such as sleep trackers or heart rate monitors as low-risk wellness devices while treating invasive devices under conventional regulations. The other document defines how the FDA will exempt clinical decision support tools from medical device regulations, limiting such software to analyzing existing data rather than extracting data from sensors, and requiring them to enable independent review of their recommendations. The documents do not rewrite any statutes, but they refine interpretation of existing law, compared to the 2019 and 2022 documents they replace. They offer a fresh lens on how regulators see technology that sits at the intersection of consumer electronics, software, and medicine—a category many other countries are choosing to regulate more strictly rather than less.

What the 2026 update changed

The 2026 FDA update clarifies how it distinguishes between “medical information” and systems that measure physiological “signals” or “patterns.” Earlier guidance discussed these concepts more generally, but the new version defines signal-measuring systems as those that collect continuous, near-continuous, or streaming data from the body for medical purposes, such as home devices transmitting blood pressure, oxygen saturation, or heart rate to clinicians. It gives more concrete examples, like a blood glucose lab result as medical information versus continuous glucose monitor readings as signals or patterns.

The updated guidance also sharpens examples of what counts as medical information that software may display, analyze, or print. These include radiology reports or summaries from legally marketed software, ECG reports annotated by clinicians, blood pressure results from cleared devices, and lab results stored in electronic health records.

In addition, the 2026 update softens FDA’s earlier stance on clinical decision tools that offer only one recommendation. While prior guidance suggested tools needed to present multiple options to avoid regulation, FDA now indicates that a single recommendation may be acceptable if only one option is clinically appropriate, though it does not define how that determination will be made.

Separately, updates to the general wellness guidance clarify that some non-invasive wearables—such as optical sensors estimating blood glucose for wellness or nutrition awareness—may qualify as general wellness products, while more invasive technologies would not.

Wellness still requires accuracy

For designers of wearable health devices, the practical implications go well beyond what label you choose. “Calling something ‘wellness’ doesn’t reduce the need for rigorous validation,” says Omer Inan, a medical device technology researcher at the Georgia Tech School of Electrical and Computer Engineering. A wearable that reports blood pressure inaccurately could lead a user to conclude that their values are normal when they are not—potentially influencing decisions about seeking clinical care.

“In my opinion, engineers designing devices to deliver health and wellness information to consumers should not change their approach based on this new guidance,” says Inan. Certain measurements—such as blood pressure or glucose—carry real medical consequences regardless of how they’re branded, Inan notes.

Unless engineers follow robust validation protocols for technology delivering health and wellness information, Inan says, consumers and clinicians alike face the risk of faulty information.

To address that, Inan advocates for transparency: companies should publish their validation results in peer-reviewed journals, and independent third parties without financial ties to the manufacturer should evaluate these systems. That approach, he says, helps the engineering community and the broader public assess the accuracy and reliability of wearable devices.

When wellness meets medicine

The societal and clinical impacts of wearables are already visible, regardless of regulatory labels, says Sharona Hoffman, JD, a law and bioethics professor at Case Western Reserve University.

Medical metrics from devices like the Apple Watch or Fitbit may be framed as “wellness,” but in practice many users treat them like medical data, influencing their behavior or decisions about care, Hoffman points out.

“It could cause anxiety for patients who constantly check their metrics,” she notes. Alternatively, “A person may enter a doctor’s office confident that their wearable has diagnosed their condition, complicating clinical conversations and decision-making.”

Moreover, privacy issues remain unresolved, unmentioned in previous or updated guidance documents. Many companies that design wellness devices fall outside protections like the Health Insurance Portability and Accountability Act (HIPAA), meaning data about health metrics could be collected, shared, or sold without the same constraints as traditional medical data. “We don’t know what they’re collecting information about or whether marketers will get hold of it,” Hoffman says.

International approaches

The European Union’s Artificial Intelligence Act designates systems that process health-related data or influence clinical decisions as “high risk,” subjecting them to stringent requirements around data governance, transparency, and human oversight. China and South Korea have also implemented rules that tighten controls on algorithmic systems that intersect with healthcare or public-facing use cases. South Korea provides very specific categories for regulation for technology makers, such as standards on labeling and description on medical devices and good manufacturing practices.

Across these regions, regulators are not only classifying technology by its intended use but also by its potential impact on individuals and society at large.

“Other countries that emphasize technology are still worrying about data privacy and patients,” Hoffman says. “We’re going in the opposite direction.”

Post-market oversight

“Regardless of whether something is FDA approved, these technologies will need to be monitored in the sites where they’re used,” says Todd R. Johnson, a professor of biomedical informatics at McWilliams School of Biomedical Informatics at UTHealth Houston, who has worked on FDA-regulated products and informatics in clinical settings. “There’s no way the makers can ensure ahead of time that all of the recommendations will be sound.”

Large health systems may have the capacity to audit and monitor tools, but smaller clinics often do not. Monitoring and auditing are not emphasized in the current guidance, raising questions about how reliability and safety will be maintained once devices and software are deployed widely.

Balancing innovation and safety

For engineers and developers, the FDA’s 2026 guidance presents both opportunities and responsibilities. By clarifying what counts as a regulated device, the agency may reduce upfront barriers for some categories of technology. But that shift also places greater weight on design rigor, validation transparency, and post-market scrutiny.

“Device makers do care about safety,” Johnson says. “But regulation can increase barriers to entry while also increasing safety and accuracy. There’s a trade-off.”

Reference: https://ift.tt/PAOhzct

Once-hobbled Lumma Stealer is back with lures that are hard to resist

Last May, law enforcement authorities around the world scored a key win when they hobbled the infrastructure of Lumma, an infostealer that infected nearly 395,000 Windows computers over just a two-month span leading up to the international operation. Researchers said Wednesday that Lumma is once again “back at scale” in hard-to-detect attacks that pilfer credentials and sensitive files.

Lumma, also known as Lumma Stealer, first appeared in Russian-speaking cybercrime forums in 2022. Its cloud-based malware-as-a-service model provided a sprawling infrastructure of domains for hosting lure sites offering free cracked software, games, and pirated movies, as well as command-and-control channels and everything else a threat actor needed to run their infostealing enterprise. Within a year, Lumma was selling for as much as $2,500 for premium versions. By the spring of 2024, the FBI counted more than 21,000 listings on crime forums. Last year, Microsoft said Lumma had become the “go-to tool” for multiple crime groups, including Scattered Spider, one of the most prolific groups.

Takedowns are hard

The FBI and an international coalition of its counterparts took action early last year. In May, they said they seized 2,300 domains, command-and-control infrastructure, and crime marketplaces that had enabled the infostealer to thrive. Recently, however, the malware has made a comeback, allowing it to infect a significant number of machines again.

Read full article

Comments

Reference : https://ift.tt/tsvpjSP

Tips for Using AI Tools in Technical Interviews

This article is crossposted from IEEE Spectrum’s careers newsletter. Sign up now to get insider tips, expert advice, and practical strategies, written in partnership with tech career development company Parsity and delivered to your inbox for free!

We’d like to introduce Brian Jenney, a senior software engineer and owner of Parsity, an online education platform that helps people break into AI and modern software roles through hands-on training. Brian will be sharing his advice on engineering careers with you in the coming weeks of Career Alert.

Here’s a note from Brian:

“12 years ago, I learned to code at the age of 30. Since then I’ve led engineering teams, worked at organizations ranging from five-person startups to Fortune 500 companies, and taught hundreds of others who want to break into tech. I write for engineers who want practical ways to get better at what they do and advance in their careers. I hope you find what I write helpful.”

Technical Interviews in the Age of AI Tools

Last year, I was conducting interviews for an AI startup position. We allowed unlimited AI usage during the technical challenge round. Candidates could use Cursor, Claude Code, ChatGPT, or any assistant they normally worked with. We wanted to see how they used modern tools.

During one interview, we asked a candidate a simple question: “Can you explain what the first line of your solution is doing?”

Silence.

After a long pause, he admitted he had no idea. His solution was correct. The code worked. But he couldn’t explain how or why. This wasn’t an isolated incident. Around 20 percent of the candidates we interviewed were unable to explain how their solutions worked, only that they did.

When AI Makes Interviews Harder

A few months earlier, I was on the other side of the table at this same company. During a live interview, I instinctively switched from my AI-enabled code editor to my regular one. The CTO stopped me.

“Just use whatever you normally would. We want to see how you work with AI.”

I thought the interview would be easy. But I was wrong.

Instead of only evaluating correctness, the interviewer focused on my decision-making process:

• Why did I accept certain suggestions?
• Why did I reject others?
• How did I decide when AI helped versus when it created more work?

I wasn’t just solving a problem in front of strangers. I was explaining my judgment and defending my decisions in real time, and AI created more surface area for judgment. Counterintuitively, the interview was harder.

The Shift in Interview Evaluation

Most engineers now use AI tools in some form, whether they write code, analyze data, design systems, or automate workflows. AI can generate output quickly, but it can’t explain intent, constraints, or tradeoffs.

More importantly, it can’t take responsibility when something breaks.

As a result, major companies and startups alike are now adapting to this reality by shifting to interviews with AI. Meta, Rippling, and Google, for instance, have all begun allowing candidates to use AI assistants in technical sessions. And the goal has evolved: interviewers want to understand how you evaluate, modify, and trust AI-generated answers.

So, how can you succeed in these interviews?

What Actually Matters in AI-Enabled Interviews

Refusing to use AI out of principle doesn’t help. Some candidates avoid AI to prove they can think independently. This can backfire. If the organization uses AI internally—and most do—then refusing to use it signals rigidity, not strength.

Silence is a red flag. Interviews aren’t natural working environments. We don’t usually think aloud when deep in a complex problem, but silence can raise concerns. If you’re using AI, explain what you’re doing and why:

• “I’m using AI to sketch an approach, then validating assumptions.”
• “This suggestion works, but it ignores a constraint we care about.”
• “I’ll accept this part, but I want to simplify it.”

Your decision-making process is what separates effective engineers from prompt jockeys.

Treat AI output as a first draft. Blind acceptance is the fastest way to fail. Strong candidates immediately evaluate the output: Does this meet the requirements? Is it unnecessarily complex? Would I stand behind this in production?

Small changes like renaming variables, removing abstractions, or tightening logic signal ownership and critical thinking.

Optimize for trust, not completion. Most AI tools can complete a coding challenge faster than any human. Interviews that allow AI are testing something different. They’re answering: “Would I trust this person to make good decisions when things get messy?”

Adapting to a Shifting Landscape

Interviews are changing faster than most candidates realize. Here’s how to prepare:

Start using AI tools daily. If you’re not already working with Cursor, Claude Code, ChatGPT, or CoPilot, start now. Build muscle memory for prompting, evaluating output, and catching errors.

Develop your rejection instincts. The skill isn’t using AI. It’s knowing when AI output is wrong, incomplete, or unnecessarily complex. Practice spotting these issues and learning known pitfalls.

Your next interview might test these skills. The candidates who’ve been practicing will have a clear advantage.

—Brian

Was 2025 Really the Year of AI Agents?

Around this time last year, CEOs like Sam Altman promised that 2025 would be the year AI agents would join the workforce as your own personal assistant. But in hindsight, did that really happen? It depends on who you ask. Some programmers and software engineers have embraced agents like Cursor and Claude Code in their daily work. But others are still wary of the risks these tools bring, such as a lack of accountability.

Read more here.

Class of 2026 Salary Projections Are Promising

In the United States, starting salaries for students graduating this spring are expected to increase, according to the latest data from the National Association of Colleges and Employers. Computer science and engineering majors are expected to be the highest paying graduates, with a 6.9 percent and 3.1 percent salary increase from last year, respectively. The full report breaks down salary projections by academic major, degree level, industry, and geographic region.

Read more here.

Go Global to Make Your Career Go Further

If given the opportunity, are international projects worth taking on? As part of a career advice series by IEEE Spectrum’s sister publication, The Institute, the chief engineer for Honeywell lays out the advantages of working with teams from around the world. Participating in global product development, the author says, could lead to both personal and professional enrichment. Read more here. Reference: https://ift.tt/wxobd5V

How Do You Define an AI Companion?

For a different perspective on AI companions, see our Q&A with Brad Knox: How Can AI Companions Be Helpful, not Harmful?

AI models intended to provide companionship for humans are on the rise. People are already frequently developing relationships with chatbots, seeking not just a personal assistant but a source of emotional support.

In response, apps dedicated to providing companionship (such as Character.ai or Replika) have recently grown to host millions of users. Some companies are now putting AI into toys and desktop devices as well, bringing digital companions into the physical world. Many of these devices were on display at CES last month, including products designed specifically for children, seniors, and even your pets.

AI companions are designed to simulate human relationships by interacting with users like a friend would. But human-AI relationships are not well understood, and companies are facing concern about whether the benefits outweigh the risks and potential harm of these relationships, especially for young people. In addition to questions about users’ mental health and emotional well being, sharing intimate personal information with a chatbot poses data privacy issues.

RELATED: How Can AI Companions Be Helpful, not Harmful?

Nevertheless, more and more users are finding value in sharing their lives with AI. So how can we understand the bonds that form between humans and chatbots?

Jaime Banks is a professor at the Syracuse University School of Information Studies who researches the interactions between people and technology—in particular, robots and AI. Banks spoke with IEEE Spectrum about how people perceive and relate to machines, and the emerging relationships between humans and their machine companions.

Defining AI Companionship

How do you define AI companionship?

Jaime Banks: My definition is evolving as we learn more about these relationships. For now, I define it as a connection between a human and a machine that is dyadic, so there’s an exchange between them. It is also sustained over time; a one-off interaction doesn’t count as a relationship. It’s positively valenced—we like being in it. And it is autotelic, meaning we do it for its own sake. So there’s not some extrinsic motivation, it’s not defined by an ability to help us do our jobs or make us money.

I have recently been challenged by that definition, though, when I was developing an instrument to measure machine companionship. After developing the scale and working to initially validate it, I saw an interesting situation where some people do move toward this autotelic relationship pattern. “I appreciate my AI for what it is and I love it and I don’t want to change it.” It fit all those parts of the definition. But then there seems to be this other relational template that can actually be both appreciating the AI for its own sake, but also engaging it for utilitarian purposes.

That makes sense when we think about how people come to be in relationships with AI companions. They often don’t go into it purposefully seeking companionship. A lot of people go into using, for instance, ChatGPT for some other purpose and end up finding companionship through the course of those conversations. And we have these AI companion apps like Replika and Nomi and Paradot that are designed for social interaction. But that’s not to say that they couldn’t help you with practical topics.

Jaime Banks customizes the software for an embodied AI social humanoid robot.Angela Ryan/Syracuse University

Different models are also programmed to have different “personalities.” How does that contribute to the relationship between humans and AI companions?

Banks: One of our Ph.D. students just finished a project about what happened when OpenAI demoted GPT-4o and the problems that people encountered, in terms of companionship experiences when the personality of their AI just completely changed. It didn’t have the same depth. It couldn’t remember things in the same way.

That echoes what we saw a couple years ago with Replika. Because of legal problems, Replika disabled for a period of time the erotic roleplay module and people described their companions as though they had been lobotomized, that they had this relationship and then one day they didn’t anymore. With my project on the tanking of the soulmate app, many people in their reflection were like, “I’m never trusting AI companies again. I’m only going to have an AI companion if I can run it from my computer so I know that it will always be there.”

Benefits and Risks of AI Relationships

What are the benefits and risks of these relationships?

Banks: There’s a lot of talk about the risks and a little talk about benefits. But frankly, we are only just on the precipice of starting to have longitudinal data that might allow people to make causal claims. The headlines would have you believe that these are the end of mankind, that they’re going to make you commit suicide or abandon other humans. But much of those are based on these unfortunate, but uncommon situations.

Most scholars gave up technological determinism as a perspective a long time ago. In the communication sciences at least, we don’t generally assume that machines make us do something because we have some degree of agency in our interactions with technologies. Yet much of the fretting around potential risks is deterministic—AI companions make people delusional, make them suicidal, make them reject other relationships. A large number of people get real benefits from AI companions. They narrate experiences that are deeply meaningful to them. I think it’s irresponsible of us to discount those lived experiences.

When we think about concerns linking AI companions to loneliness, we don’t have much data that can support causal claims. Some studies suggest AI companions lead to loneliness, but other work suggests it reduces loneliness, and other work suggests that loneliness is what comes first. Social relatedness is one of our three intrinsic psychological needs, and if we don’t have that we will seek it out, whether it’s from a volleyball for a castaway, my dog, or an AI that will allow me to feel connected to something in my world.

Some people, and governments for that matter, may move toward a protective stance. For instance, there are problems around what gets done with your intimate data that you hand over to an agent owned and maintained by a company—that’s a very reasonable concern. Dealing with the potential for children to interact, where children don’t always navigate the boundaries between fiction and actuality. There are real, valid concerns. However, we need some balance in also thinking about what people are getting from it that’s positive, productive, healthy. Scholars need to make sure we’re being cautious about our claims based on our data. And human interactants need to educate themselves.

Jaime Banks holds a mechanical hand.Angela Ryan/Syracuse University

Why do you think that AI companions are becoming more popular now?

Banks: I feel like we had this perfect storm, if you will, of the maturation of large language models and coming out of COVID, where people had been physically and sometimes socially isolated for quite some time. When those conditions converged, we had on our hands a believable social agent at a time when people were seeking social connection. Outside of that, we are increasingly just not nice to one another. So, it’s not entirely surprising that if I just don’t like the people around me, or I feel disconnected, that I would try to find some other outlet for feeling connected.

More recently there’s been a shift to embodied companions, in desktop devices or other formats beyond chatbots. How does that change the relationship, if it does?

Banks: I’m part of a Facebook group about robotic companions and I watch how people talk, and it almost seems like it crosses this boundary between toy and companion. When you have a companion with a physical body, you are in some ways limited by the abilities of that body, whereas with digital-only AI, you have the ability to explore fantastic things—places that you would never be able to go with another physical entity, fantasy scenarios.

But in robotics, once we get into a space where there are bodies that are sophisticated, they become very expensive and that means that they are not accessible to a lot of people. That’s what I’m observing in many of these online groups. These toylike bodies are still accessible, but they are also quite limiting.

Do you have any favorite examples from popular culture to help explain AI companionship, either how it is now or how it could be?

Banks: I really enjoy a lot of the short fiction in Clarkesworld magazine, because the stories push me to think about what questions we might need to answer now to be prepared for a future hybrid society. Top of mind are the stories “Wanting Things,” “Seven Sexy Cowboy Robots,” and “Today I am Paul.” Outside of that, I’ll point to the game Cyberpunk 2077, because the character Johnny Silverhand complicates the norms for what counts as a machine and what counts as companionship.

Reference: https://ift.tt/tEw4smP

How and When the Memory Chip Shortage Will End

If it feels these days as if everything in technology is about AI, that’s because it is. And nowhere is that more true than in the market for computer memory. Demand, and profitability, for the type of DRAM used to feed GPUs and other accelerators in AI data centers is so huge that it’s diverting away supply of memory for other uses and causing prices to skyrocket. According to Counterpoint Research, DRAM prices have risen 80-90 precent so far this quarter.

The largest AI hardware companies say they have secured their chips out as far as 2028, but that leaves everybody else—makers of PCs, consumer gizmos, and everything else that needs to temporarily store a billion bits—scrambling to deal with scarce supply and inflated prices.

How did the electronics industry get into this mess, and more importantly, how will it get out? IEEE Spectrum asked economists and memory experts to explain. They say today’s situation is the result of a collision between the DRAM industry’s historic boom and bust cycle and an AI hardware infrastructure build-out that’s without precedent in its scale. And, barring some major collapse in the AI sector, it will take years for new capacity and new technology to bring supply in line with demand. Prices might stay high even then.

To understand both ends of the tale, you need to know the main culprit in the supply and demand swing, high-bandwidth memory, or HBM.

What is HBM?

HBM is the DRAM industry’s attempt to short-circuit the slowing pace of Moore’s Law by using 3D chip packaging technology. Each HBM chip is made up of as many as 12 thinned-down DRAM chips called dies. Each die contains a number of vertical connections called through silicon vias (TSVs). The dies are piled atop each other and connected by arrays of microscopic solder balls aligned to the TSVs. This DRAM tower—well, at about 750 micrometers thick, it’s more of a brutalist office-block than a tower—is then stacked atop what’s called the base die, which shuttles bits between the memory dies and the processor.

This complex piece of technology is then set within a millimeter of a GPU or other AI accelerator, to which it is linked by as many as 2,048 micrometer-scale connections. HBMs are attached on two sides of the processor, and the GPU and memory are packaged together as a single unit.

The idea behind such a tight, highly-connected squeeze with the GPU is to knock down what’s called the memory wall. That’s the barrier in energy and time of bringing the terabytes per second of data needed to run large language models into the GPU. Memory bandwidth is a key limiter to how fast LLMs can run.

As a technology, HBM has been around for more than 10 years, and DRAM makers have been busy boosting its capability.

As the size of AI models has grown, so has HBM’s importance to the GPU. But that’s come at a cost. SemiAnalysis estimates that HBM generally costs three times as much as other types of memory and constitutes 50 percent or more of the cost of the packaged GPU.

Origins of the memory chip shortage

Memory and storage industry watchers agree that DRAM is a highly cyclical industry with huge booms and devastating busts. With new fabs costing US $15 billion or more, firms are extremely reluctant to expand and may only have the cash to do so during boom times, explains Thomas Coughlin, a storage and memory expert and president of Coughlin Associates. But building such a fab and getting it up and running can take 18 months or more, practically ensuring that new capacity arrives well past the initial surge in demand, flooding the market and depressing prices.

The origins of today’s cycle, says Coughlin, go all the way back to the chip supply panic surrounding the COVID-19 pandemic . To avoid supply-chain stumbles and support the rapid shift to remote work, hyperscalers—data center giants like Amazon, Google, and Microsoft—bought up huge inventories of memory and storage, boosting prices, he notes.

But then supply became more regular and data center expansion fell off in 2022, causing memory and storage prices to plummet. This recession continued into 2023, and even resulted in big memory and storage companies such as Samsung cutting production by 50 percent to try and keep prices from going below the costs of manufacturing, says Coughlin. It was a rare and fairly desperate move, because companies typically have to run plants at full capacity just to earn back their value.

After a recovery began in late 2023, “all the memory and storage companies were very wary of increasing their production capacity again,” says Coughlin. “Thus there was little or no investment in new production capacity in 2024 and through most of 2025.”

The AI data center boom

That lack of new investment is colliding headlong with a huge boost in demand from new data centers. Globally, there are nearly 2,000 new data centers either planned or under construction right now, according to Data Center Map. If they’re all built, it would represent a 20 percent jump in the global supply, which stands at around 9,000 facilities now.

If the current build-out continues at pace, McKinsey predicts companies will spend $7 trillion by 2030, with the bulk of that—$5.2 trillion—going to AI-focused data centers. Of that chunk, $3.3 billion will go toward servers, data storage, and network equipment, the firm predicts.

The biggest beneficiary so far of the AI data center boom is unquestionably GPU-maker Nvidia. Revenue for its data center business went from barely a billion in the final quarter of 2019 to $51 billion in the quarter that ended in October 2025. Over this period, its server GPUs have demanded not just more and more gigabytes of DRAM but an increasing number of DRAM chips. The recently released B300 uses eight HBM chips, each of which is a stack of 12 DRAM dies. Competitors’ use of HBM has largely mirrored Nvidia’s. AMD’s MI350 GPU, for example, also uses eight, 12-die chips.

With so much demand, an increasing fraction of the revenue for DRAM makers comes from HBM. Micron—the number three producer behind SK Hynix and Samsung—reported that HBM and other cloud-related memory went from being 17 percent of its DRAM revenue in 2023 to nearly 50 percent in 2025.

Micron predicts the total market for HBM will grow from $35 billion in 2025 to $100 billion by 2028—a figure larger than the entire DRAM market in 2024, CEO Sanjay Mehrotra told analysts in December. It’s reaching that figure two years earlier than Micron had previously expected. Across the industry, demand will outstrip supply “substantially… for the foreseeable future,” he said.

Future DRAM supply and technology

“There are two ways to address supply issues with DRAM: with innovation or with building more fabs,” explains Mina Kim, an economist with the Mkecon Insights. “As DRAM scaling has become more difficult, the industry has turned to advanced packaging… which is just using more DRAM.”

Micron, Samsung, and SK Hynix combined make up the vast majority of the memory and storage markets, and all three have new fabs and facilities in the works. However, these are unlikely to contribute meaningfully to bringing down prices.

Micron is in the process of building an HBM fab in Singapore that should be in production in 2027. And it is retooling a fab it purchased from PSMC in Taiwan that will begin production in the second half of 2027. Last month, Micron broke ground on what will be a DRAM fab complex in Onondaga County, N.Y. It will not be in full production until 2030.

Samsung plans to start producing at a new plant in Pyeongtaek, South Korea in 2028.

SK Hynix is building HBM and packaging facilities in West Lafayette, Indiana set to begin production by the end of 2028, and an HBM fab it’s building in Cheongju should be complete in 2027.

Speaking of his sense of the DRAM market, Intel CEO Lip-Bu Tan told attendees at the Cisco AI Summit last week: “There’s no relief until 2028.”

With these expansions unable to contribute for several years, other factors will be needed to increase supply. “Relief will come from a combination of incremental capacity expansions by existing DRAM leaders, yield improvements in advanced packaging, and a broader diversification of supply chains,” says Shawn DuBravac , chief economist for the Global Electronics Association (formerly the IPC). “New fabs will help at the margin, but the faster gains will come from process learning, better [DRAM] stacking efficiency, and tighter coordination between memory suppliers and AI chip designers.”

So, will prices come down once some of these new plants come on line? Don’t bet on it. “In general, economists find that prices come down much more slowly and reluctantly than they go up. DRAM today is unlikely to be an exception to this general observation, especially given the insatiable demand for compute,” says Kim.

In the meantime, technologies are in the works that could make HBM an even bigger consumer of silicon. The standard for HBM4 can accommodate 16 stacked DRAM dies, even though today’s chips only use 12 dies. Getting to 16 has a lot to do with the chip stacking technology. Conducting heat through the HBM “layer cake” of silicon, solder, and support material is a key limiter to going higher and in repositioning HBM inside the package to get even more bandwidth.

SK Hynix claims a heat conduction advantage through a manufacturing process called advanced MR-MUF (mass reflow molded underfill). Further out, an alternative chip stacking technology called hybrid bonding could help heat conduction by reducing the die-to-die vertical distance essentially to zero. In 2024, researchers at Samsung proved they could produce a 16-high stack with hybrid bonding, and they suggested that 20 dies was not out of reach.

Reference: https://ift.tt/lgI2hUj

IEEE Honors Global Dream Team of Innovators

Meet the recipients of the 2026 IEEE Medals—the organization’s highest-level honors. Presented on behalf of the IEEE Board of Directors, these medals recognize innovators whose work has shaped modern technology across disciplines including AI, education, and semiconductors.

The medals will be presented at the IEEE Honors Ceremony in April in New York City. View the full list of 2026 recipients on the IEEE Awards website, and follow IEEE Awards on LinkedIn for news and updates.

IEEE MEDAL OF HONOR

Sponsor: IEEE

Jensen Huang

Nvidia

Santa Clara, Calif.

“For leadership in the development of graphics processing units and their application to scientific computing and artificial intelligence.”

IEEE FRANCES E. ALLEN MEDAL

Sponsor: IBM

Luis von Ahn

Duolingo

Pittsburgh

“For contributions to the advancement of societal improvement and education through innovative technology.”

IEEE ALEXANDER GRAHAM BELL MEDAL

Sponsor: Nokia Bell Labs

Scott J. Shenker

University of California, Berkeley

“For contributions to Internet architecture, network resource allocation, and software-defined networking.”

IEEE JAGADISH CHANDRA BOSE MEDAL IN WIRELESS COMMUNICATIONS

Sponsor: Mani L. Bhaumik

Co-recipients:
Erik Dahlman
Stefan Parkvall
Johan Sköld

Ericsson

Stockholm

“For contributions to and leadership in the research, development, and standardization of cellular wireless communications.”

IEEE MILDRED DRESSELHAUS MEDAL

Sponsor: Google

Karen Ann Panetta

Tufts University

Medford, Mass.

“For contributions to computer vision and simulation algorithms, and for leadership in developing programs to promote STEM careers.”

IEEE EDISON MEDAL

Sponsor: The Edison Medal Fund

Eric Swanson

PIXCEL Inc.

MIT

“For pioneering contributions to biomedical imaging, terrestrial optical communications and networking, and inter-satellite optical links.”

IEEE MEDAL FOR ENVIRONMENTAL AND SAFETY TECHNOLOGIES

Sponsor: Toyota Motor Corp.

Wei-Jen Lee

University of Texas at Arlington

“For contributions to advancing electrical safety in the workplace, integrating renewable energy and grid modernization for climate change mitigation.”

IEEE FOUNDERS MEDAL

Sponsor: IEEE Foundation

Marian Rogers Croak

Google

Reston, Va.

“For leadership in communication networks, including acceleration of digital equity, responsible Artificial Intelligence, and the promotion of diversity and inclusion.”

IEEE RICHARD W. HAMMING MEDAL

Sponsor: Qualcomm, Inc.

Muriel Médard

MIT

“For contributions to coding for reliable communications and networking.”

IEEE NICK HOLONYAK, JR. MEDAL FOR SEMICONDUCTOR OPTOELECTRONIC TECHNOLOGIES

Sponsor: Friends of Nick Holonyak, Jr.

Steven P. DenBaars

University of California, Santa Barbara

“For seminal contributions to compound semiconductor optoelectronics, including high-efficiency visible light-emitting diodes, lasers, and LED displays.”

IEEE MEDAL FOR INNOVATIONS IN HEALTHCARE TECHNOLOGY

Sponsor: IEEE Engineering Medicine and Biology Society

Rosalind W. Picard

MIT

“For pioneering contributions to wearable affective computing for health and wellbeing.”

IEEE JACK S. KILBY SIGNAL PROCESSING MEDAL

Sponsor: Apple

Biing-Hwang “Fred” Juang

Georgia Tech

“For contributions to signal modeling, coding, and recognition for speech communication.”

IEEE/RSE JAMES CLERK MAXWELL MEDAL

Sponsor: ARM, Ltd.

Paul B. Corkum

University of Ottawa

“For the development of the recollision model for strong field light–matter interactions leading to the field of attosecond science.”

IEEE JAMES H. MULLIGAN, JR. EDUCATION MEDAL

Sponsor: IEEE Life Members Fund and MathWorks

James H. McClellan

Georgia Tech

“For fundamental contributions to electrical and computer engineering education through innovative digital signal processing curriculum development.”

IEEE JUN-ICHI NISHIZAWA MEDAL

Sponsor: IEEE Jun-ichi Nishizawa Medal Fund

Eric R. Fossum

Dartmouth College

Hanover, N.H.

“For the invention, development, and commercialization of the CMOS image sensor.”

IEEE ROBERT N. NOYCE MEDAL

Sponsor: Intel Corp.

Chris Malachowsky

Nvidia

Santa Clara, Calif.

“For pioneering parallel computing architectures and leadership in semiconductor design that transformed artificial intelligence, scientific research, and accelerated computing.”

IEEE DENNIS J. PICARD MEDAL FOR RADAR TECHNOLOGIES AND APPLICATIONS

Sponsor: RTX

Yoshio Yamaguchi

Niigata University

Japan

“For contributions to polarimetric synthetic aperture radar imaging and its utilization.”

IEEE MEDAL IN POWER ENGINEERING

Sponsors: IEEE Industry Applications, Industrial Electronics, Power Electronics, and Power & Energy societies

Fang Zheng Peng

University of Pittsburgh

“For contributions to Z-Source and modular multi-level converters for distribution and transmission networks.”

IEEE SIMON RAMO MEDAL

Sponsor: Northrop Grumman Corp.

Michael D. Griffin

LogiQ, Inc.

Arlington, Va.

“For leadership in national security, civil, and commercial systems engineering and development of elegant design principles.”

IEEE JOHN VON NEUMANN MEDAL

Sponsor: IBM

Donald D. Chamberlin

IBM

San Jose, Calif.

“For contributions to database query languages, particularly Structured Query Language, which powers most of the world’s data management and analysis systems.”

Reference: https://ift.tt/nvySKWA