The wires are getting too small. That's the blunt truth facing anyone trying to cram more computing power into silicon these days.
Every generation of chips pushes transistors closer, stacks silicon dies higher, demands more. But the materials holding it all together—the microscopic copper veins threading through billions of transistors, the thermal pastes pulling heat away, the glass and ceramic substrates bearing the weight—haven't evolved at anything close to the same pace. What used to take a decade of painstaking lab work to discover and validate? The industry now needs it done in months. Because waiting isn't an option anymore.
A two-person startup called Matforge thinks it has an answer: AI agents purpose-built to hunt for new materials. Emerging from Y Combinator's Spring 2026 cohort, the company claims its software can compress more than ten years of materials R&D into a matter of months—a claim the startup makes on its promotional materials, though independent validation remains pending. The founders are Akash Ramdas, a Stanford PhD whose interconnect research is documented in his academic profile, and Advaith Sridhar, a founding applied scientist at Persona AI. Their target? The critical bottlenecks choking semiconductor fabs and datacenters—precisely where demand has shifted from predictable to something closer to panic.
And the timing, well, it's no accident.
The International Energy Agency reported in 2026 that datacenter electricity demand jumped 17% in 2025, with big tech capex surpassing $400 billion that same year. For 2026, the agency is projecting substantial increases in capital expenditures. IDC's estimate for AI infrastructure spending alone: $487 billion this year. This isn't a gradual build. It's a scramble, plain and simple.
The Limits Are Physical Now
No amount of software cleverness can solve the problems AI training clusters are hitting. High-bandwidth memory shortages could stretch through 2027—that's according to warnings from Samsung and SK hynix reported back in May 2026. Tom's Hardware noted something telling: customers are now reserving multi-year HBM capacity, behavior previously seen only during the most severe supply crunches.
Advanced packaging has become another choke point. Epoch AI flagged CoWoS capacity as a critical constraint in 2025. S&P Global reported in January 2026 that HBM average selling prices are climbing—up 8% for Samsung, 22% for Micron. Direct liquid cooling revenues, per Dell'Oro data cited in Data Center Frontier, are on pace to exceed $2 billion, expanding 85% year-over-year as facilities wrestle with rack densities climbing from 10 to 30 kilowatts.
These are material problems in both senses of the word.
Copper interconnects—the nanoscale wiring threading through chips—face fundamental physics constraints below 5 nanometers. Thermal interface materials need higher conductivity as power densities keep climbing. Glass substrates and interposers promise better signal integrity for AI accelerators, but the supply chain is still taking shape. The industry is running out of room and running out of materials that work at the required scale. Simultaneously.
Intel's December 2024 IEDM presentation highlighted "subtractive ruthenium" as an alternative interconnect metallization—a signal that copper's decades-long dominance faces genuine challenges. Applied Materials announced advances in low-resistance copper wiring and ruthenium CVD for 2-nanometer nodes and beyond. These aren't theoretical exercises. They're urgent engineering needs with compressed timelines.
Discovery Is Too Slow
Historically, finding a new semiconductor material meant years of hypothesis generation, synthesis attempts, characterization, failure analysis, iteration. A materials scientist might test dozens of candidates over a decade to find one meeting the multi-objective requirements: low resistivity, acceptable electromigration resistance, thermal stability, manufacturability at scale, cost viability.
This sequential, human-driven process worked when Moore's Law ticked forward every 18 to 24 months and manufacturing nodes lasted years.
It doesn't work when TSMC is laying out roadmaps for kilowatt-class chips and customers are reserving HBM capacity years ahead of need.
The semiconductor materials market hit $67.5 billion globally in 2024, according to SEMI data released in April 2025. Wafer fab equipment, test, and assembly/packaging is projected to reach $156 billion by 2027, per SEMI figures cited by Tom's Hardware. But size doesn't solve the materials gap. The supply chain needs new compounds, alloys, structures—preferably yesterday.
AI Scientists, Meet the Lab
Using artificial intelligence to accelerate materials discovery isn't exactly new. But the execution? That's maturing fast.
DeepMind's GNoME project, published in Nature in 2023, predicted 2.2 million crystal structures, releasing around 381,000 of the most stable candidates. A companion effort at Lawrence Berkeley National Lab's A-Lab reported synthesizing 41 novel compounds in 17 days using autonomous robotics—though that claim sparked verification debates that continue today.
Microsoft's Azure Quantum Elements screened over 32 million candidates, narrowing them to roughly 500,000 stable materials in its battery electrolyte collaboration with Pacific Northwest National Laboratory. NVIDIA highlighted accelerated computing platforms for materials discovery at SC25 last November, including tools for nano-imaging and thermal research under its ALCHEMI initiative.
Then there's Citrine Informatics, which ran a controlled experiment in July 2024 for photolithography formulation—a direct semiconductor application. They pitted 200 AI-suggested experiments against 200 traditional approaches. The top nine results? All from the AI suggestions. That's not marginal improvement. That's a different success rate entirely.
Academic labs are publishing what they call agentic workflows—systems where language models propose candidates, orchestrate simulations, suggest next experiments. A preprint posted in April 2026 described an agentic LLM campaign targeting lithium-halide spinel conductors, showing improved hit rates over successive iterations. Self-driving labs, as benchmarked in recent reviews, reduce the experiments required to find viable candidates. Acceleration factors are becoming a standard metric.
How Matforge Works (Allegedly)
Matforge describes its platform as "AI scientists that discover new materials for the semiconductor industry—specifically datacenters and fabs." The company's stated goal: compress ten-plus years of lab work into months using what it calls "a swarm of AI agents." These agents, according to the YC directory, span the entire materials discovery process—generating candidates, synthesizing, testing in physical labs.
Ramdas brings domain expertise from Stanford, where according to his publicly available profile, his research focused on nanoscale interconnect materials using multi-objective optimization. His work explored candidates including cobalt-platinum alloys as potential alternatives to copper and ruthenium for local interconnects below 5 nanometers. The company's YC page includes a self-reported claim—unverified by third-party sources as of May 2026—that materials Ramdas discovered have been adopted into Intel and TSMC roadmaps.
The startup is seeking introductions to packaging and thermal engineers at NVIDIA, AMD, Groq, Cerebras, Google, Intel, Samsung, TSMC. Essentially the entire AI accelerator and foundry ecosystem. This isn't scattershot. It's targeting organizations facing the most acute materials constraints, ones willing to evaluate alternatives if they deliver performance gains.
What remains unclear: the composition of this "swarm," how it integrates with physics-based simulations like density functional theory, which lab partners handle synthesis, what validation pipeline bridges computational predictions to high-volume manufacturing. Materials that work in simulations sometimes fail electromigration tests or time-dependent dielectric breakdown under real-world stress.
The path from candidate to qualified part is littered with failures. No AI system has fully automated that journey yet.
The Competition Is Crowded
Matforge isn't operating in a vacuum. Orbital Materials raised funding from NVentures in October 2024 for AI-driven materials and climate technologies. Kebotix operates a self-driving lab platform. Mat3ra offers cloud-based materials simulations with AI models. Mater-AI focuses on thermoelectrics. MatCraft provides AI-powered discovery tooling, datasets, workflows.
Perhaps the most direct comparison is Citrine Informatics, which demonstrated closed-loop experimentation in semiconductor materials. Their photolithography case study showed quantifiable acceleration in a domain adjacent to Matforge's targets. Microsoft's Azure Quantum Elements represents the platform play—offering simulation and screening tools to enterprises rather than discovering materials themselves.
Differentiation likely comes down to semiconductor-specific focus.
General-purpose materials discovery platforms can screen millions of compounds, but translating those into back-end-of-line interconnects or advanced packaging substrates requires domain expertise. Knowing which properties matter—resistivity versus electromigration versus chemical mechanical polishing compatibility versus integration with existing tooling—narrows the search space in ways generic AI can't.
Federal Money, Geopolitical Urgency
Policy is amplifying the pressure. The CHIPS Act allocated $1.4 billion in final awards announced in January 2025 to support next-generation semiconductor advanced packaging, with $300 million specifically for advanced substrates and materials R&D. Recipients included Absolics, Applied Materials, Arizona State University. A separate $1.1 billion went to establish the Advanced Packaging Pilot Facility.
These programs aim to rebuild domestic capabilities in areas where the U.S. ceded ground over decades. Taiwan leads in advanced packaging through TSMC's CoWoS and SoIC. South Korea dominates HBM. China's export controls on gallium and germanium—first imposed in August 2023, subject to licensing changes and temporary suspensions through 2026—exposed supply-chain fragility for compound semiconductors.
According to reports from late 2025 and early 2026, China suspended certain bans but maintained licensing requirements, with gallium prices spiking amid geopolitical tensions and Middle East conflicts. The result: persistent material price volatility and renewed attention on alternatives. McKinsey's Global Materials Perspective from October 2025 warned that geopolitics is raising material supply concentration risks just as AI expansion drives demand for copper and power.
Materials discovery isn't just a technical challenge anymore. It's increasingly a national security concern.
What to Watch
The materials informatics market—admittedly hard to size with precision—is projected by some firms to grow from roughly $157-184 million in 2025 to over $700 million by 2035. FactMR reported in April 2026 that the U.S. segment alone could see 19.8% CAGR from 2025 to 2035. Treat these figures directionally, not definitively, but the trend is clear: more money flowing toward computational materials design.
For datacenter operators facing power and cooling constraints, the near-term path involves direct liquid cooling adoption, higher-voltage distribution, squeezing more performance per watt from existing hardware. But the medium-term gains—2027 and beyond—come from better thermal interface materials, advanced substrates, interconnects that don't bottleneck at 4-nanometer pitches. If AI agents can compress the discovery cycle, competitive advantage accrues to whoever deploys those materials first.
For chip packaging engineers, the HBM supply crunch through 2027 means exploring alternatives: better through-silicon vias, hybrid bonding techniques, glass interposers. Each requires materials innovations—dielectrics with lower loss, adhesives that withstand thermal cycling, metals that maintain conductivity at finer pitches. Teams that prototype faster, validate faster, qualify faster will win capacity in constrained supply chains.
For deep tech investors, the question is whether AI-driven materials discovery follows the software playbook—network effects, winner-take-most dynamics—or remains fragmented by domain and application. So far, evidence suggests domain specificity matters. Battery materials require different expertise than interconnect metals. But the tooling (agentic workflows, self-driving lab orchestration, open datasets like OMat24) is becoming modular, reusable.
Validation Remains the Hard Part
None of this happens without rigorous validation. The A-Lab controversy—41 novel compounds reportedly synthesized, then later scrutinized for methodology—underscores the challenge. Autonomous labs can generate data quickly. Whether that data translates to manufacturable, reliable materials? Different question.
Reliability testing for semiconductor materials is brutal.
Electromigration under current stress. Time-dependent dielectric breakdown. Mechanical shock. Thermal cycling. Chemical compatibility with etchants and slurries. A material that works in a small-scale synthesis run might fail when scaled to 300-millimeter wafers. A compound that meets performance specs might not survive the qualification gauntlet.
The industry has spent decades building these validation frameworks precisely because early promise doesn't guarantee production success. AI can accelerate the discovery phase, but compressing the validation timeline requires trusted partnerships with foundries, OSATs, equipment suppliers who control the characterization and integration steps.
Matforge's pitch—ten years to months—implicitly assumes it can navigate this valley of death. That may depend less on algorithmic sophistication and more on access: access to characterization tools, access to pilot lines, access to engineers willing to run experiments on unproven candidates. The Stanford pedigree and YC network provide some of that. The rest comes from early design wins.
The Next Decade
The International Roadmap for Devices and Systems, published in April 2026, projects that copper will remain the preferred interconnect metal at least through 2029, with ongoing research into alternative metals and barrier materials. Translation: even the most aggressive timelines assume incremental progress, not wholesale substitution.
But incremental progress at compressed timelines still matters when every percentage point of resistivity reduction or thermal conductivity gain translates to rack-level power savings.
TSMC's roadmap outlines SoIC 3D stacking migrating from 6-micron pitches today to 4.5-micron by 2029. Terabit optical links are moving from research to integration. Glass substrates are positioned as growth drivers for advanced packaging, though full transition timelines remain subject to industry development. Each of these requires materials that don't exist in commercial production today.
The question isn't whether AI will accelerate materials discovery—the evidence from controlled experiments and academic labs says it already is. The question is whether startups like Matforge can execute fast enough to matter in an industry where incumbents have deep pockets, established relationships, internal R&D pipelines.
Perhaps the more interesting scenario is one where AI-discovered materials become table stakes—where every major player deploys similar computational tools, and differentiation shifts to integration expertise, manufacturing agility, customer relationships. In that world, value doesn't accrue to the AI itself but to the organizations that wield it most effectively.
For now, the industry's materials bottleneck is real. Timelines are compressed. Capital is flowing.
Matforge is betting AI agents can navigate the search space faster than humans can. Whether that bet pays off depends not just on algorithmic prowess but on bridging the gap between computational prediction and production-qualified parts.
The datacenter operators installing next year's racks, the foundries qualifying 2-nanometer processes, the packaging engineers designing for kilowatt chips? They don't need perfect materials. They need better materials, faster.
That's the window.
