The chip engineers at SK hynix had a problem. Actually, they had the same problem chip engineers have wrestled with for decades: finding the right materials fast enough to keep pace with demand. Except this time, they claim to have cracked it—not through more aggressive hiring or bigger labs, but by turning AI loose on the periodic table.
In April 2026, the South Korean memory giant reported a significant reduction in the time needed to identify suitable materials for specific applications using AI. If that kind of acceleration holds up across the industry, and that's a meaningful "if," it would upend one of the semiconductor sector's most stubborn constraints: the glacial pace of materials discovery.
That discovery process has always lagged behind chip design itself, sometimes by years. You can draw up blueprints for a faster transistor in months; proving that an exotic compound can withstand the thermal stress of mass production can take a decade. But the economics of AI have made waiting untenable. Global chip revenue hit $791.7 billion in 2025, up 25.6% year-over-year, according to the Semiconductor Industry Association. Forecasts for 2026 range widely—from approximately $1 trillion to $1.32 trillion by Gartner, and up to $1.5 trillion by WSTS—but the direction is unmistakable. The industry is booming, and it needs materials innovation to keep up.
Enter a wave of startups, research consortia, and corporate labs betting that machine learning can compress what used to take human researchers years into a matter of months. Matforge, a Y Combinator-backed company from the spring 2026 cohort, bills itself as an "AI scientist" for semiconductor materials used in datacenters and fabs. Google DeepMind is building a Gemini-powered materials lab in the UK. Berkeley Lab has an autonomous synthesis robot. Applied Materials just opened a sprawling R&D center designed to bring materials, device, and packaging engineers into the same room—literally.
Whether any of this will work at scale remains an open question. But the fact that so much capital and talent is pouring into AI-driven materials discovery signals something deeper: the belief that the next bottleneck in chip manufacturing isn't transistor design. It's chemistry.
Following the Money
The numbers tell part of the story. The materials informatics market—a catch-all term for computational tools, databases, and machine learning workflows used in discovery—was valued at just under $247 million in 2026 by Fact.MR, with projections climbing to $1.35 billion by 2036. Those are vendor estimates, so treat them with appropriate skepticism. But even if the actual growth undershoots, the directional pull is clear.
More tangible is the investment flowing into AI infrastructure itself. A May 2026 report from the SIA and Deloitte projects cumulative datacenter capital expenditures of $4.0 trillion between 2023 and 2030, with up to $2.8 trillion earmarked for semiconductors and hardware. That's a staggering sum, and it's reshaping the technical requirements for everything from interconnects to the thermal pastes sandwiched between chips and heat sinks.
Data centers are a significant driver of electricity demand growth. The U.S. Energy Information Administration flagged them in January 2026 as a primary driver behind the strongest four-year growth in electricity demand since 2000. By 2035, according to the SIA-Deloitte analysis, roughly 30% of AI systems will rely on liquid cooling, a shift that places new and sometimes unforgiving demands on materials stability and thermal conductivity. Air cooling, it turns out, has limits.
The foundry market reflects this intensity. Revenue hit a record $320 billion in 2025, per Counterpoint Research. SEMI, the industry trade group, expects semiconductor equipment sales to reach $145 billion this year and $156 billion in 2027. But throwing money at equipment only gets you so far when the underlying materials can't keep up. Copper, the workhorse of chip interconnects, starts to lose efficiency at the smallest line widths due to surface scattering effects. Traditional dielectrics struggle with the speed and power requirements of sub-5 nanometer nodes. Photoresists—those light-sensitive chemicals used in lithography—demand tighter control over dose and line-edge roughness as features shrink.
In other words, the next generation of chips is running headlong into the limits of the periodic table.
The New Toolbox
So what does AI-assisted materials discovery actually look like in practice?
Take Imec, the Belgium-based nanoelectronics research hub. In February 2026, at the SPIE Advanced Lithography + Patterning conference, Imec researchers reported that carefully controlled oxygen injection during the post-exposure bake step for metal-oxide resists boosted photospeed by 15-20%. That's not a moonshot. It's the kind of incremental optimization—tweaking atmospheric conditions, adjusting bake times—that might once have taken months of trial and error. The difference now is the feedback loop between simulation, prediction, and lab validation has tightened.
Or consider IBM and Lam Research, which announced a five-year partnership in March 2026 to develop High-NA EUV dry resists capable of scaling beyond the 1 nanometer node. (Yes, "1 nanometer" is marketing shorthand, but the technical challenge is real.) Dry resists eliminate much of the liquid chemistry that complicates traditional lithography, potentially cutting material waste and energy use. Lam had already demonstrated 28 nanometer pitch back-end-of-line patterning with its Aether dry resist technology in early 2025; the IBM collaboration aims to push that into the High-NA regime, where precision margins grow even thinner.
On the interconnect side, researchers are hunting for alternatives to copper at the tiniest scales. A March 2026 arXiv preprint detailed high-throughput computational screening of ruthenium compounds for sub-5 nanometer lines. Another study that same month examined surface states and conductivity in ultrathin ruthenium wires. Neither claimed a production-ready solution, but both illustrate the intensity—and the computational firepower—now being brought to bear on candidate materials long before expensive fabrication trials begin.
Then there's the packaging challenge. SK hynix's iHBM architecture, unveiled in May 2026, integrates improved thermal interface materials to manage the heat generated by stacked high-bandwidth memory. The company claims a roughly 30% reduction in thermal resistance, a significant figure given that AI accelerators are pushing memory ever closer to logic. Liquid cooling helps, but it introduces its own materials puzzle: the compounds need to remain stable under prolonged contact with coolants, a requirement that rules out some otherwise promising candidates.
The Ecosystem Takes Shape
Perhaps the most revealing development isn't any single material or breakthrough, but the way the industry is reorganizing itself around the problem.
Applied Materials' EPIC Center, which opened in 2026, is designed to co-locate device makers, equipment suppliers, and academic researchers. Samsung joined in February, SK hynix and Micron in March, TSMC in May. Gary Dickerson, Applied Materials' CEO, has framed the center's mission around "energy-efficient performance for AI systems"—a telling shift in emphasis. It's no longer just about cramming more transistors onto a die. It's about whether those transistors can run without melting the chip or bankrupting the datacenter operator on electricity bills.
The model echoes, in some ways, what's happening in the broader AI-for-science community. Google DeepMind's GNoME platform, announced back in November 2023, predicted 2.2 million crystal structures; roughly 380,000 were flagged as potentially stable. The database has since undergone corrections and academic scrutiny—science in real time, messier than press releases suggest—but the ecosystem it helped catalyze continues to evolve. Berkeley Lab's autonomous A-Lab can now synthesize compounds with minimal human intervention. DeepMind is building a Gemini-powered materials lab in the UK. In February 2026, Berkeley announced a multi-institution push to develop an open-source AI assistant explicitly aimed at energy materials, with applications in both batteries and semiconductors.
A flurry of 2026 research papers—published on arXiv between March and June—captures the breadth of activity: AI-discovered electronic metal phosphide semiconductors, agentic language models reasoning through self-driving labs for air-sensitive lithium halide spinels, inverse design of complex oxide thin films, extreme-k dielectrics via physics-validated generative models. Each tackles a different corner of the materials landscape, but the architecture is the same: tighter loops between simulation and experiment, algorithms learning from smaller datasets, predictions validated in hours rather than weeks.
Synopsys, the EDA giant, showcased NVIDIA-accelerated quantum chemistry simulations at GTC 2026, claiming speedups of up to 30×. Those numbers come from controlled benchmarks, so real-world mileage may vary. But even a 10× or 15× improvement would matter in an R&D environment where time-to-market can mean the difference between capturing a product cycle and missing it entirely.
What Comes Next
If the optimists are right, the materials discovery cycle will keep compressing, particularly in domains where datasets are rich and the underlying physics well understood. Interconnects, dielectrics, photoresists—all have decades of prior research to draw on, and all face urgent scaling pressures. Companies that can weave AI workflows into their existing R&D pipelines, whether through in-house tools or partnerships, will likely outpace those still relying on traditional methods.
But there's a countervailing force: complexity. The convergence of materials science and systems engineering is accelerating, and that means optimizing a single material in isolation misses the point. Applied Materials' EPIC Center model reflects an industry recognition that tomorrow's chips will be limited less by transistor density and more by how effectively materials manage heat, signal integrity, and power delivery at the package and system level. The SIA-Deloitte report calls thermal and packaging co-design a "system bottleneck," which is consultant-speak for: this is where the next wave of problems will cluster.
Energy efficiency, once a secondary concern, is now front and center. With data centers driving substantial electricity demand growth—the U.S. EIA projects commercial-sector sales (which include data centers) will rise 2.2% in 2026 and 5.3% in 2027—materials that reduce power consumption or improve thermal management have direct economic and environmental implications. Lower-loss interconnects, high-k/low-k dielectric engineering, advanced thermal interface materials: these are marginal gains that, multiplied across millions of servers, translate into material cost savings and measurable carbon reductions.
The geopolitical backdrop adds another layer of friction. U.S. export controls on semiconductor equipment to China remain in effect, with one-year licenses granted to Samsung and SK hynix for certain tools in December 2025. China, for its part, has restricted exports of gallium, germanium, and antimony—critical inputs for compound semiconductors—since 2023. The U.S. CHIPS and Science Act continues funding domestic capacity; June 2026 awards included up to $1.6 billion for USA Rare Earth to shore up critical materials supply chains. The Department of Energy's Genesis Mission, announced in March with $293 million in funding, explicitly targets AI-for-science applications, including materials discovery.
There's something almost circular about the current moment. The chips that made generative AI possible are now being designed, at least in part, by generative AI. Whether that loop accelerates semiconductor innovation or simply moves the bottleneck elsewhere—to fab capacity, supply chain logistics, or some other constraint we haven't hit yet—is still unclear. Skepticism is warranted. AI hasn't eliminated the hard work of verifying predictions in the lab, scaling synthesis processes, or ensuring that a material performing beautifully in simulation doesn't degrade after six months in a humid fab environment.
But for now, the industry is betting that the slowest part of the chip development cycle can be sped up. That the materials enabling the next generation of AI will themselves be discovered with AI's help.
Maybe that's less irony than inevitability. Or maybe it's just the kind of story the industry tells itself when the old ways stop scaling.
