The numbers look triumphant at first glance. February semiconductor sales: $88.8 billion, up 61.8% year-over-year according to SIA. The industry is closing in on a trillion dollars in annual revenue, perhaps as soon as this year. But walk into the materials labs at any major chipmaker, and you'll find a different mood—something closer to controlled panic.
The problem isn't demand. It's that the fundamental building blocks of chips—the copper wiring, the insulating layers, the thermal interfaces that keep everything from melting—are running out of headroom. Copper interconnects below 10 nanometers hit resistivity walls. Dielectrics crack under stress. Photoresists can't keep up with the next wave of lithography tools. And the traditional playbook for solving these problems, the decade-long march from university lab to production fab, feels suddenly, acutely too slow.
Which is why a handful of startups, and some very large research institutions, are now betting on a shortcut: AI systems that are being developed to assist and potentially streamline materials science workflows. Call them AI scientists if you like. The pitch is seductive—compress years of trial-and-error into months of algorithmic brute force. Whether it actually works in the messy reality of semiconductor manufacturing is another question.
When Volume Meets Physics
Yes, global semiconductor materials revenue hit a record $73.2 billion in 2025, up 6.8% according to SEMI's reporting. First-quarter chip sales for 2026 were $298.5 billion, a 25% jump from Q4 2025, according to SIA. The money is real, the momentum unmistakable.
But here's the wrinkle: advanced packaging shortages are already pinching supply. TSMC's CoWoS—chip-on-wafer-on-substrate—technology is a case study. Analyst projections suggest strong demand for TSMC's CoWoS technology, estimated around 1 million wafers this year, and capacity isn't keeping up. Samsung and SK hynix sounded alarms in April about AI-driven memory shortages, particularly in high-bandwidth memory, with constraints likely stretching into 2027 or beyond.
Then there's power. The International Energy Agency noted that datacenter electricity consumption rose 17% last year. Over $400 billion was invested by major tech companies in datacenter capital expenditures in 2025, with anticipated significant increases according to IEA. Wood Mackenzie warned in May that AI datacenter power demands are beginning to stress electrical grids in certain regions.
Volume, in other words, isn't the bottleneck anymore. The materials themselves are.
Copper, the workhorse metal for chip interconnects since the late 1990s, doesn't scale gracefully to the smallest feature sizes. Below 10 nanometers, its resistivity shoots up due to grain boundary and surface scattering—physics asserting itself in inconvenient ways. Low-k dielectrics, the insulators meant to keep signals from bleeding into each other, struggle with mechanical reliability. And photoresists optimized for current extreme ultraviolet lithography tools aren't quite ready for the higher numerical aperture machines coming online.
It's a materials crunch dressed up in trillion-dollar sales figures.
Enter the Algorithm
Matforge, a San Francisco startup that emerged from Y Combinator's Spring batch this year, is among the companies attempting to innovate in materials discovery using AI technologies. Their pitch: "AI scientists" that discover new materials for semiconductors, explicitly targeting datacenters and fabs. The timeline they're proposing—months instead of the typical decade-plus—sounds implausible until you remember what's already happened in this space.
CEO Akash Ramdas has some credibility here. He spent his PhD and postdoc years at Stanford working on interconnect materials, publishing research that tackled exactly this problem. A 2024 paper in the journal Small detailed a multi-objective optimization framework that screened more than 15,000 candidates and validated cobalt-platinum as a promising post-copper option for local interconnects. Y Combinator's listing claims materials Ramdas discovered have found their way into roadmaps at Intel and TSMC, though we couldn't independently verify that in public disclosures.
His co-founder, Advaith Sridhar, comes from the AI side—previously at Luma Labs and Persona AI, with a master's from Carnegie Mellon. According to a CB Insights listing, the company has raised $500,000 via convertible note, though details remain scarce.
Matforge isn't operating in a vacuum. Google DeepMind's GNoME project, unveiled back in November 2023, predicted 2.2 million crystal structures—about 380,000 of them deemed likely stable—and framed it as equivalent to "800 years' worth" of discovery at historical rates. Microsoft's MatterGen, published in Nature in January last year, demonstrated a generative model for inorganic materials with improved novelty and stability metrics. One of the materials it predicted, TaCr₂O₆, was actually synthesized in collaboration with external labs. Not just a computational curiosity—real atoms, arranged in real space.
The question is whether that leap from prediction to synthesis can happen fast enough, and reliably enough, to matter.
The Specifics Are Brutal
The materials challenges aren't abstract. At IBM's IEDM presentation in late 2024, researchers showed off fully subtractive ruthenium top-via interconnects with airgap spacing around 9 nanometers. Ruthenium is now a leading candidate to take over from copper at advanced nodes. Applied Materials has commercialized ruthenium-cobalt liners under the Volta Ru brand, enabling barriers thinner than 20 angstroms—freeing up more cross-sectional area for the metal that actually carries current, reducing resistance in the process.
Backside power delivery networks, a key architecture for future logic chips, are pushing exploration of ruthenium, molybdenum, and other metals for through-silicon vias and backside metallization. IBM demonstrated BSPDN integration with nanosheet transistors at VLSI last year, and academic papers continue to wrestle with thermal management and reliability trade-offs.
Lithography materials are evolving just as fast, if not faster. Lam Research and JSR's Inpria unit announced a cross-licensing and collaboration deal in September 2025 targeting dry resist technologies for EUV and atomic layer processes. In March, IBM and Lam revealed a five-year partnership at Albany NanoTech focused on High-NA EUV and dry resist development, aiming to scale logic beyond 1 nanometer.
The old way—wet chemistry photoresists optimized over years of patient iteration—can't adapt quickly enough. And then there's regulation. The EPA finalized the first U.S. national drinking water standards for PFAS in April 2024, with maximum contaminant levels as low as 4 parts per trillion for certain compounds. The European Union is advancing staged restrictions on PFAS subgroups, some of which have already begun. Many fluorinated chemistries used in lithography and wet processes are now under scrutiny, adding friction to an already strained R&D cycle.
A Crowded, Messy Landscape
Matforge has company. Berkeley Lab announced FORUM-AI in February, an open-source, agentic AI platform developed by a multi-lab consortium focused on energy and materials discovery. The platform aims to close the loop between computational prediction, autonomous synthesis, and experimental validation—what researchers sometimes call self-driving labs, though that term carries a whiff of hype.
Recent preprints detail autonomous thin-film epitaxy systems controlled by real-time X-ray diffraction and agentic reasoning workflows for air-sensitive lithium halide spinels. The National Laboratory of the Rockies showcased examples in semiconductor thin films and catalytic nanocrystals in May.
Established players are moving, too. Citrine Informatics, an enterprise SaaS platform for materials and chemicals, raised $16 million in a Series C round a few years back and continues to serve large industrial customers. Mat3ra offers a materials R&D cloud emphasizing simulations and AI. Kebotix combines AI with robotic self-driving labs and has partnered with bp on molecular design testbeds. Schrödinger, publicly traded with 2025 revenue around $256 million, serves pharmaceutical and materials use cases with computational discovery tools.
Startups like MatCraft, Synthera (which markets an "AI copilot" called Raven for synthesis planning), and MolAgent (an on-premises agent for computational discovery) are all jockeying for position. Competitive differentiation will likely hinge on domain specificity—can these tools actually navigate the multi-objective constraints of semiconductor processes? Resistivity, scattering, thermal coefficients, integration chemistry. It's not just about predicting a stable structure. It's about predicting one that survives the gauntlet of a modern fab.
What Comes Next
The stakes aren't getting any smaller. SEMI forecasts that wafer fab equipment, assembly, packaging, and test sales will reach $145 billion this year and $156 billion in 2027. The IEA's Electricity 2026 report projects global electricity demand will grow 3.6% annually through the end of the decade, with AI datacenters as a primary driver. If materials discovery can't accelerate, everything downstream slows.
New interconnect metals. Ultra-low-k dielectrics that don't fail under stress. Thermal interface materials capable of handling higher power densities. Photoresists compatible with High-NA EUV and hybrid bonding. All needed. All on multi-year timelines under traditional methods.
AI-driven discovery offers a potential shortcut, but let's not pretend it's a sure thing. Predicted materials still require synthesis, validation, and integration into fiendishly complex process flows. The gap between a promising computational candidate and a material that survives the thermal budgets, etch chemistries, and reliability tests of a modern fab remains wide. Perhaps uncomfortably so.
The real test isn't whether AI can generate novel structures—DeepMind and Microsoft have already shown that's possible. It's whether these systems can navigate the unglamorous realities of process integration at scale. Edge cases. Failure modes. The thousand small ways a material can look great in silico and terrible in a cleanroom.
Matforge has the pedigree, the timing, and the positioning. Whether it has the partnerships, the validation infrastructure, and the patience to turn algorithmic promise into fab-ready materials—that's the next chapter. The trillion-dollar chip industry, for better or worse, is watching.
