ChipAgents Accelerates Chip Design With Agentic Physics-Aware Models

Using AI to speed chip design has become a hot topic of late! AI models can be used to do for chips what Anthropic has done for software: speed up the work so you can do more work. After providing some perspective on three startups, I recently covered what Cognichip and Cadence design has announced, and now another company has announced new chip-design AI, targeting both digital and analog chips: ChipAgents.

ChipAgents Is Targeting the Vast Physical AI Chip Design Market

While data center growth is currently consuming everyone’s attention, Physical AI is touted as the next big thing in AI, bringing robotic intelligence and automated vehicles to market. Unfortunately, this vision has a rather large problem: the current industry cannot possibly provide the number of chip designs needed to populate the vast array of physical devices envisioned. ChipAgents, a startup I covered in my overview of three startups driving Agentic AI solutions to accelerate chip development, has estimated that each physical AI device could consist of upwards of 3000 chips. Sounds high to me, but when you do the math at a projected 145 million devices, that implies some half trillion chips could be required from the semiconductor industry, an impossibility given then current capacity for chip design and production. Even if the 3000 number is too high, the industry will undoubtedly need a lot of chips; a good market to target for automation.

Chip Complexity Demands a 10X Design Improvement

I always thought new chip designs were increasingly being produced with “A0” silicon, but the data provided by ChipAgents show that engineering excellence is being overwhelmed by increasing design complexity, with only some 10% of designs working on the first tapeout. Even OpenAI’s famous Jalepeño AI chips is being spun into a B0 for production. And with the cost of a tapeout increasing exponentially, the situation has become untenable.