Why The AI Race Is Also A Chemistry Race
Dr. Ahmed Ismail, founder of Dunia Innovations and Aquature, builds AI-driven autonomous labs and technologies for energy and chemicals.
gettyHere is a trivia question for your next board meeting: Which company holds a near-monopoly on a material that sits inside the package of almost every advanced AI chip? If you said Nvidia or TSMC, you are in good company, but you are wrong. It is Ajinomoto, the Japanese company best known for MSG.
Ajinomoto Build-up Film, or ABF, insulates the dense wiring in the substrate that an AI processor is fixed on. Since a leading CPU maker adopted it in 1999, ABF has reportedly captured more than 95% of its market. In 2026, riding the AI boom, the electronic-materials business built around it reported operating margins above 50%.
I love this story because it exposes a blind spot. We talk about AI as models and parameters, yet every GPU is a stack of chemistry, engineered atom by atom for over 70 years, and that chemistry is fast becoming the hardest part of the race.
In the 1980s, chip technology was based on about 12 chemical elements. By the 2000s, it used more than 60 elements. A 2025 preprint estimates that the industry now relies on over 85% of all non-radioactive elements.
Some of AI’s material problems, like ABF supply, are about scale, and capital may fix them. The harder problems are physics walls, where the materials we need do not exist yet.
These physical walls include wires, heat and space.
As chips shrink, copper wiring starts to behave like a resistor. Below roughly 17 by 17 nanometers, ruthenium’s resistivity drops below copper’s.
In 2024, Nvidia’s thermal engineers reportedly saw two-phase direct-to-chip cooling as the preferred route for next-generation accelerators. Its classic fluids, however, belong to the PFAS family of forever chemicals. 3M completed its exit from PFAS manufacturing at the end of 2025. Someone has to invent the replacement.
On October 1, 2026, Google launched a prototype satellite carrying four of its AI chips to test how they handle radiation and the thermal extremes of orbit. With no air or water to carry heat away, space behaves like an insulator, and heat can only leave by radiation. Without better coatings to radiate that heat, an orbital data center will be a materials problem first before it solves a computing problem.
Nvidia has committed to what Jensen Huang calls a “one-year rhythm” for new AI chips. A new material, by contrast, commonly needs 10 to 20 years to go from discovery to deployment. So we may soon see a big mismatch between an industry moving at software speed that is building on a foundation that moves at chemistry speed.
The good news? The same technology hungry for new materials is finally good enough to help find them. Better materials make better chips, better chips train better models and better models can search the chemical space faster than any human team. That loop makes the AI race and the chemistry race one and the same.
In November 2025, the U.S. launched the Genesis Mission, a Department of Energy effort to link national-lab supercomputers and datasets with autonomous, AI-directed experiments. In July 2026, the White House said more than $5 billion in federal commitments were being made to expand it. The AI labs are moving, too. In December 2025, Google DeepMind said that it would open a UK lab in 2026, where robots could make and test hundreds of materials a day.
Frontier models have already read much of the internet’s text and code. The next great dataset has to come from the lab, failed experiments included. Governments and AI labs are betting that the most valuable data comes from touching the physical world.
In 2023, Google DeepMind’s GNoME model predicted 2.2 million new crystal structures. When two materials chemists at UC Santa Barbara examined a sample in 2024, they found “scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility.”
I know this simulation-to-reality gap firsthand. During my graduate research at synchrotron facilities in France, Germany, the U.S. and Japan, I watched catalysts at work under X-ray beams, and these materials rarely stayed the same during the chemical reaction. Inside the reactor, they kept reshaping themselves until they settled into a final form that was different from the original one. How could a computer simulation have predicted this?
A computer can propose a thousand candidates, but only a furnace, a reactor or an X-ray beam can tell you which candidate is real. Building an autonomous lab taught my team the same lesson: The scarcest and most valuable resource is the capacity to make and measure candidates quickly, under realistic conditions.
What does this mean for technology leaders today? Start with a few first steps:
• Map your AI roadmap down to its materials. For every plan, ask what it physically needs. If the answer is a material that does not exist yet, it goes on your risk register, right next to your supply chain risks.
• Treat experimental data as a strategic asset, failures included. The failed experiments in lab notebooks are useful data, too. In a 2016 Nature study, a model trained on a research group’s archived lab notebooks, including failed reactions, recommended experiments that worked 89% of the time. Experienced chemists got 78%.
• Judge AI claims by the physical proof in the lab. When someone announces a million new materials, ask how many were actually made and how many were stable under realistic conditions. One million predicted candidates is just a list of hypotheses. A novel material that survives weeks inside a real reactor is what we could call the start of a materials breakthrough.
The take-home message is simple: AI’s next big step depends on materials that don’t exist yet. The leaders who plan for that now will be the first to benefit when those materials arrive.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
