Observations By: Souma Basu, Senior Director, Insurance Practice, Sutherland Global Services
Based on the July 2026 IEEE Spectrum cover story "AI Designs Radio Chips That Humans Couldn't Even Imagine" by Kaushik Sengupta. Research published in Nature Communications (December 2024) and demonstrated at ISSCC 2025 by the Sengupta group, Princeton University.
Imagine handing an engineer a challenge: design something better than anything in the textbooks, without using the textbooks.
No templates. No inherited assumptions. Just the laws of physics and a blank page.
Most engineers, even brilliant ones, would struggle. Human design is built on accumulated intuition. We start from what we know and refine what came before.
Now imagine an AI doing exactly that. Starting from physics itself, exploring millions of possibilities no human would think to try, and producing designs that look nothing like what we have built for decades, yet perform at the state of the art.
That is not a thought experiment. It is happening now, in one of the most stubbornly human corners of engineering.
The July 2026 IEEE Spectrum cover story by Prof. Kaushik Sengupta of Princeton crystallized something I have been thinking about for a long time. His team is using AI to design radio-frequency chips, the silicon behind everything wireless in your life: your phone, your car's radar, satellite links, the coming wave of 6G.
The last dark art in chip design
Most of the chip industry has been automated for years. RF design has resisted. Practitioners literally call it a dark art, mastered only through years of apprenticeship.
Here is why. Designing an RF chip is like solving millions of interconnected physics puzzles at the same time. Electromagnetic fields, heat, mechanical stress, and signal behavior all interact. Adjust one element and ten others shift.
A single new chip can take months or years and enormous investment. So for generations, engineers reused proven templates and refined them incrementally. Safe and reliable, but almost certainly nowhere near optimal.
What the Princeton team did differently
Sengupta's group asked a braver question. What if we discard the templates entirely?
Instead of teaching AI to imitate human designs, they had it optimize directly against physics-based simulation. Propose a structure, test it against real electromagnetic behavior, learn, try again. Millions of times.
The approach echoes how game-playing AI reached superhuman performance: not by studying human matches, but by learning from the rules of the game itself.
The results, published in Nature Communications and demonstrated at ISSCC 2025, are striking. The AI generated RF structures that differ substantially from conventional human-designed templates. Some look like lace or abstract art. And in the demonstrated cases described in the article, they achieved state-of-the-art performance.
Designs that once took weeks or months emerged in hours. In one demonstration, minutes.
One finding stopped me cold. The research found no evidence that the templates the industry has relied on for decades are anywhere close to optimal for modern design goals.
We were not standing on the shoulders of giants. We were standing inside a comfortable local maximum, unaware of the terrain beyond it.
Why business leaders should pay attention
First, speed changes economics. When design cycles compress from months to hours, companies can explore more options, fail faster, and bring differentiated products to market sooner. Think next-generation smartphones, connected vehicles, and network infrastructure, where wireless performance is a competitive battleground.
Second, better designs mean better efficiency. The article describes AI-designed amplifiers achieving record efficiency at wide bandwidth. Across billions of wireless devices, small efficiency gains in silicon become real energy and battery-life outcomes.
Third, and most important: this creates a compounding innovation loop.
Today's AI runs largely on chips designed by humans. Tomorrow's AI will increasingly run on chips designed with AI assistance. Better AI helps design better hardware. Better hardware enables better AI. Each turn of the wheel accelerates the next.
We have watched compounding loops transform software. We have rarely seen one take hold in physical engineering. That is what makes this moment different.
The honest caveats
The article is refreshingly candid about the limits, and we should be too.
AI can still hallucinate designs that do not work, so human verification remains essential. These systems are powerful design partners, not replacements for engineering judgment.
The field's biggest constraint is not algorithms but data. Engineers worldwide run nearly identical simulations daily, yet most of that knowledge sits locked behind nondisclosure agreements. An open data ecosystem could do for chip design what shared image datasets once did for computer vision.
Anyone who has led large engineering programs will recognize the pattern. The technology is rarely the hardest part. The ecosystem is.
The bigger picture
We often say AI is transforming software. That may already be yesterday's story.
The next chapter appears to be AI helping redesign the physical world itself: the chips, the antennas, the materials, the machinery underneath everything digital.
The article notes that similar AI-driven approaches are already emerging in biology, materials science, automotive, and aerospace engineering. Chip design may simply be the most visible early example of a broader shift: less reliance on inherited templates, more direct conversation with the laws of physics.
Full credit to Prof. Kaushik Sengupta and his Princeton team for the research, and to IEEE Spectrum for telling the story so well. It is worth your time.
A question for my network
If AI can help design the hardware that powers future AI, which field experiences this compounding effect next?
Materials science? Drug discovery? Robotics? Aerospace?
I would genuinely love to hear your perspective, especially from those working at the intersection of AI and physical engineering.
