By Ian Krietzberg — November 11, 2025
Over the past few years, the astronomical flow of capital into AI has been driven by a simple thesis: unlocking transformative superintelligence will require trillions of dollars, thousands of data centers, and massive new power infrastructure. Tech leaders fuel that vision with bold predictions—Sam Altman speculating about breakthroughs in high-energy physics and space colonization, Elon Musk insisting superhuman invention is “just a matter of time,” and Anthropic’s Dario Amodei suggesting AI may soon surpass the smartest PhDs.
The entire frontier-model narrative rests on one idea: bigger is better.
But while the Big Five chase theoretical breakthroughs, a quieter revolution is emerging. Smaller companies are building intricate systems of lighter-weight models—not to chase AGI, but to solve highly specific, real-world problems.
The Rise of Practical, Smaller-Scale AI
One domain especially hungry for practical AI progress is materials science—fields like advanced batteries, solar panels, semiconductors, aerospace materials, and next-generation engines. Progress is notoriously slow because breakthroughs require expensive, time-intensive experimentation.
Radical AI, founded in 2024, is aiming to change that. CEO Joseph Krause described their system as a “materials flywheel,” using diffusion models, neural networks, generative AI, and atomistic simulations to design and autonomously test new materials. Large language models play only an interface role; the real intelligence sits in specialized models built for the domain.
Radical’s autonomous lab—funded by a $55M seed round with participation from Nvidia—attempts to run 100 experiments a day, compared to the ~50 per year Krause ran during his PhD. Every failure feeds data back into the system, accelerating discovery. The company is already working with the U.S. Air Force to develop high-entropy alloys for hypersonic flight.
“We envision a world where discovery through light manufacturability takes a week, instead of fifteen years,” Krause said.
Geminus and the Power of Multi-Model AI
A systems-driven approach also sits at the core of Geminus, whose platform helps industrial operators optimize production processes while reducing environmental impact. Industrial data is often sparse, noisy, and high-stakes—making precise predictions with measurable confidence crucial.
Geminus blends AI architectures with physics-based models, guardrails, and constraints to deliver the accuracy required for real-world engineering decisions.
Dr. Karthik Duraisamy, founder and chief scientist at Geminus and a professor at the University of Michigan, argues that a single giant model is unlikely to ever robustly handle scientific or engineering environments on its own.
“The best frameworks involve a big reasoning model at the top,” he explains, “which talks to smaller, expert foundation models that know only molecules, or biology, or weather.” Large models still play a role, but “on their own, they’re very deficient.”
This multi-model philosophy is spreading across high-stakes sectors. Companies like xCures use the same approach in healthcare, constraining their systems with narrowly trained expert models to avoid hallucinations and ensure trustworthy clinical insights.
General-purpose LLMs alone simply aren’t safe enough.
Rethinking the Infrastructure Race
Despite these innovations, the tidal wave of data-center construction shows no signs of slowing. The economic incentives are enormous. But smaller companies are developing alternatives that challenge the narrative that gigawatt-hungry infrastructure is essential.
SiMa.ai, for example, focuses on powering AI at the edge—embedded in devices, vehicles, robotics, and medical equipment. CEO Krishna Rangasayee believes the industry has overlooked the hardware side of AI in favor of the cloud. SiMa has raised $355M to build specialized silicon that outperforms Nvidia in efficiency.
“We’ve gotten comfortable with megawatts and gigawatts going into data centers,” Rangasayee said. “We are proof that you don’t need to burn that much power.”