The bottleneck isn't where anyone expected it to be.
While the world fixates on transistor counts and power consumption, the semiconductor industry is quietly running into a more fundamental problem: the stuff chips are actually made from. Metals. Dielectrics. Thermal compounds. The unglamorous materials that connect, insulate, and cool silicon are now the limiting factor in how fast—and how far—AI computing can advance.
And discovering a better interconnect material? That traditionally takes a decade, maybe longer. Painstaking lab work. Trial and error. The AI datacenter boom, needless to say, doesn't have a decade to spare.
Which brings us to Matforge, a two-person startup fresh out of Y Combinator's spring 2026 cohort. Their pitch: collapse ten-plus years of semiconductor materials R&D into months using what they describe as "a swarm of AI agents." Audacious doesn't quite cover it, particularly in an industry where materials innovation has always moved at a glacial pace—by necessity, not choice.
But the pressure mounting across the industry is very real. The clock, as they say, is ticking.
A $73 Billion Problem Nobody's Talking About
The semiconductor materials market hit $73.2 billion in 2025, climbing 6.8 percent year-over-year, according to SEMI's reporting this past May. Nearly all that growth traces back to AI infrastructure, which devoured datacenter electricity at a rate 17 percent higher than the year before, per International Energy Agency figures.
Demand isn't the issue. Delivery is.
Advanced packaging capacity remains critically constrained through at least 2027, according to an April analysis from TrendForce. The transition from 4nm and 5nm AI chips to 3nm nodes late last year and into this one has consumed more wafer and packaging resources per device. High-bandwidth memory stacks, 2.5D and 3D packaging, the substrates holding them together—every piece of the puzzle is running into supply constraints.
The materials side is even tighter, perhaps more than the industry anticipated. SK hynix, for instance, unveiled a new cooling architecture in late May called "iHBM," embedding cooling elements directly into high-bandwidth memory interfaces to slash thermal resistance by over 30 percent. That's the kind of engineering required when AI chips approach power densities that make traditional cooling methods essentially untenable. Microsoft Azure researchers projected in a mid-May preprint that some datacenter deployments could hit one megawatt per unit by next year.
Meanwhile, U.S. energy regulators are reportedly fast-tracking AI datacenter interconnections—but only if projects "bring their own power or curtail at peaks," as Tom's Hardware reported in mid-June. In other words, if your chips can't run cooler and more efficiently, you might not get to plug them in at all.
Why This Time Is Different
The semiconductor industry has faced materials challenges before. Every node shrink brings them. But the AI era differs in two critical respects: speed and scale.
Traditional materials discovery follows a rhythm that's almost ritualistic. Researchers identify a need, screen candidates computationally, synthesize the most promising options, characterize their properties, iterate—and eventually, if everything breaks right, hand off to manufacturing. The Materials Genome Initiative, launched back in 2011, aimed to cut discovery-to-deployment time and cost in half. Even that ambitious goal assumed years, not months.
Today's AI infrastructure doesn't have years. Chip designers at NVIDIA, AMD, Groq, Cerebras, and others are pushing the limits of interconnect metals, thermal interface materials, and power delivery systems right now. Ruthenium, cobalt, molybdenum—exotic conductors are replacing copper in nanoscale wiring because traditional materials hit fundamental physics limits below 16 nanometers. Research groups like Belgium's imec demonstrated 16-nanometer-pitch ruthenium semi-damascene lines with record-low resistance at a conference last June. Intel showcased subtractive ruthenium interconnects with airgaps at IEDM in late 2024.
These aren't lab curiosities. They're manufacturing roadmap items for 2-nanometer-class production and beyond.
The materials science community has responded by turning to AI itself. Google DeepMind's GNoME, announced back in November 2023, predicted 2.2 million crystal structures, around 380,000 of which were flagged as potentially stable. Berkeley Lab's A-Lab used those predictions—plus literature text-mining and computational tools—to autonomously synthesize 41 novel inorganic compounds from 58 targets in just 17 days. The Materials Project now counts more than 650,000 registered users and sees roughly 5,000 daily active users, according to a Berkeley Lab announcement in January.
But there's a yawning gap between discovering a material in silico and actually putting it into a fab. Nature ran a feature last October with a pointed title: "AI is dreaming up millions of new materials. Are they any good?" The piece highlighted validation challenges and the replicability gap between AI predictions and experimentally realized materials. Scientific debate continues over which claims hold up under rigorous testing.
This is the terrain Matforge is attempting to navigate.
The Founders and Their Gamble
Matforge's two founders bring complementary technical chops to the problem. Advaith Sridhar, previously a research engineer at Luma Labs and Persona AI, holds a master's from Carnegie Mellon and an undergraduate degree from IIT Madras. His co-founder, Akash Ramdas, completed his PhD and postdoc at Stanford, focusing on materials for nanoscale electronics and interconnects. Ramdas published a paper in the journal Small in 2024 on multi-objective optimization for novel compound metals in interconnect applications. Matforge's Y Combinator profile notes that Ramdas' interconnect materials "have been adopted into the roadmaps of Intel and TSMC"—though that remains a self-reported claim without independent public confirmation as of late June.
The company's core pitch revolves around autonomous AI agents that search the materials design space, propose candidates, and iterate rapidly. In their YC launch post from roughly April or May, they framed the problem starkly: "10+ years of lab work." They asked for introductions to packaging and thermal engineers at major chip companies. Their website promises "10x better alternatives in months."
How far along they actually are remains unclear. No public funding amounts beyond the YC backing. No disclosed customer names or quantitative performance metrics. But the timing is worth noting—Matforge is positioning itself squarely at the intersection of two acute industry pain points: materials bottlenecks in advanced packaging and thermal management, and the long cycle times of traditional R&D.
They're hardly alone in applying AI to materials discovery, though most competitors focus on broader domains or different verticals.
Materials Nexus, based in the UK, reported last June that it used AI to discover a rare-earth-free magnet dubbed "MagNex" in just three months from design to test—a claimed 200-times speedup. Orbital Materials (sometimes called Orbital Industries) reportedly raised $50 million in a round covered in news reports this June, with backing from NVentures. Citrine Informatics focuses on materials informatics platforms for multiple industries, including semiconductors; a January application note described their work on photolithography formulation optimization.
Other players—Kebotix, Atinary (which reported 1,000-times acceleration claims in certain catalysis optimization loops in an ACS Catalysis paper last November), ExoMatter, CuspAI (which raised a $30 million seed round back in June 2024)—are all pursuing variations on the same theme: use AI to navigate vast materials design spaces faster than humans ever could.
On the enterprise and tooling side, established players are weaving AI into their existing workflows. Synopsys QuantumATK offers atomistic modeling integrated with TCAD, with performance boosts from NVIDIA GPUs noted in March release notes. Dassault Systèmes' BIOVIA Materials Studio has rolled out updates. Applied Materials has announced a series of materials engineering platforms for what it calls the "angstrom era," covering advanced patterning, wiring, and integrated packaging. Lam Research's Semiverse digital twin and process modeling solutions, introduced back in 2023, continue seeing adoption.
What distinguishes the startup wave from the incumbents is focus—and speed. The new entrants aren't trying to sell simulation software to fabs. They're trying to discover the actual materials, then license or partner on deployment.
The Validation Problem
The semiconductor industry has trust issues with AI-generated materials, and for good reason. The GNoME and A-Lab results from 2023 were scientifically impressive—on the surface. But subsequent scrutiny revealed cracks. Not all predicted materials were genuinely novel or viable. Some turned out to be known already. Others couldn't be synthesized. Still others showed properties that diverged from predictions once subjected to testing.
This validation gap matters more in semiconductors than in nearly any other domain. A bad battery electrolyte candidate might cost a few thousand dollars and some wasted lab time. A flawed interconnect material integrated into a leading-edge fab could derail a multi-billion-dollar production line.
So companies like Matforge, if they succeed, won't just need good predictions. They'll need materials that can pass the gauntlet: real-world testing, process integration, reliability qualification, supply chain scaling. Ramdas' academic track record—including that Small publication and the purported Intel and TSMC roadmap adoptions—suggests he's aware of this reality. But awareness and execution are very different things, particularly when you're a two-person team.
The industry tailwinds are undeniable, at least. Gartner projects semiconductor revenue exceeding $1.3 trillion in 2026. The Semiconductor Industry Association reported first-quarter global sales of $298.5 billion, putting the industry on track for roughly $1 trillion annually. AI infrastructure drives the primary growth engine across virtually every forecast this year, from Deloitte's February outlook to WSTS's spring projections.
But growth creates pressure, which creates bottlenecks. Power electronics for datacenters are shifting toward 800-volt DC architectures, driving demand for silicon carbide and gallium nitride modules. Yole Group's 2026 SiC report projects an $11 billion market for power SiC devices by 2031. Thermal interface materials are forecast to see double-digit compound annual growth rates through 2033, according to recent market research. Advanced packaging substrates using ABF films and hybrid bonding remain constrained.
Each of those bottlenecks represents a materials problem. And each materials problem is a potential customer for AI-driven discovery platforms—assuming they work.
Geopolitics and Other Complications
There are geopolitical wrinkles layered on top of the technical ones. China imposed export licensing on gallium and germanium back in August 2023, and those controls have persisted. Chipmaking material prices doubled in some cases as of March, according to reporting that attributed the spike to Middle East conflicts compounding China's restrictions. Applied Materials paid a $252 million penalty to the Bureau of Industry and Security earlier this year related to SMIC exports—a reminder that the compliance environment for equipment and materials remains fraught, to put it mildly.
On the policy side, the CHIPS Act continues funding materials and packaging R&D. NIST issued a notice of funding opportunity for the National Advanced Packaging Manufacturing Program's materials and substrates track in early 2024, with updates rolling through October and into this year. Some CHIPS awards have explicitly included materials-related manufacturing, such as ultra-high-purity chemicals and gallium nitride on silicon.
None of this guarantees Matforge—or any other AI materials startup—will succeed. The scientific challenges are formidable. The commercial risks are high. The timelines, even with AI acceleration, remain uncertain.
But the need is clear, perhaps clearer than it's ever been. The semiconductor industry is running out of time to solve materials bottlenecks the old-fashioned way. Whether a two-person team from Y Combinator can genuinely compress a decade of R&D into months remains very much to be seen.
What's certain is that someone has to try. The alternative—waiting another ten years for better materials—isn't really an alternative at all.
