What Falls Between the Silos

Design, manufacturing and the data center are each adopting AI agents. The value shows up in the connections between them, and today almost no one is managing those connections.

At SEMICON West, a lot of the conversation will be about the true value and cost of agentic AI in SoC design and manufacturing. It's the right question. From the watchtower, the answer is becoming clear: agentic AI pays off most not inside any one tool, but in the spaces between design, fab and deployment, where the most expensive problems start.

Every complex program runs on human glue: the architect who remembers why a margin was set, the product engineer who notices a test result doesn't match the model, the program manager who realizes two email threads are about the same problem. For decades that glue held. It was never designed; it was people being good at their jobs. It's failing now because the systems it holds together have outgrown what any person can track.

One change, from design to data hall

Take one change and follow it.

A model team revises an AI workload. Utilization rises, and the accelerator will run hotter for longer. The chip's thermal envelope moves, and on paper it's just a revised number in a spec.

In design, that number reopens power and timing closure. It changes which blocks need attention and which margins still hold.

In manufacturing, it reaches further than most design teams see. Test limits and burn-in conditions need revisiting. Binning strategy shifts, because the parts that meet the new envelope may be a different slice of the distribution. The package may need a different thermal solution, which can mean a different substrate, assembly flow or OSAT qualification.

In the data center, heat per rack goes up. The rack may cross the line where air cooling stops being enough and direct liquid cooling becomes mandatory. That changes coolant distribution capacity, the facility water loop and the mechanical design of the data hall. Those are long-lead, capital-intensive decisions, often locked in many months before first silicon ships.

The ripple runs the other way too. Early yield and test data might show a bin running hotter than the model predicted. That's a fab signal with consequences for the design team's next revision and for the facility planner's cooling budget. Neither of them is on the distribution list.

In a data hall, "set concrete" is not a metaphor: it's poured, it has piping in it, and changing it is very expensive. Digging it up costs far more than getting the pour right at the start. The same holds for a mask set, a qualified package or a locked test program.

Nobody in that chain did anything wrong. Each team was correct inside its own boundary. What failed was the space between them.

Every silo is getting its own AI

Design tools are gaining agents for simulation, placement and verification (EDA Is Entering Its Third Era). Fabs are applying AI to process control and yield learning (Photonics Is Driving a New Approach to Yield Analytics). Data center operators are using it to plan power and cooling. These are real gains.

But every one of them stays inside its own boundary. A design agent doesn't know the test program changed. A yield model doesn't know the workload moved. A facility planner's model doesn't know either. The problems that break schedules and budgets sit at the handoffs. That's why we argued the engineering stack itself has to be re-architected (Re-Architecting the Engineering Stack), and why so many issues still surface late (The Validation Crisis in AI Silicon; Why Are We Still Finding So Many Bugs So Late?).

Making each silo smarter doesn't connect the silos. It just gets each one to the wrong answer faster.

What an orchestration layer actually needs

An intelligent orchestration layer that spans design, manufacturing and deployment is the obvious next step. It would be made of agents that watch across domains, trace a change to everything it touches, and raise the issue early. But agents are only as good as the context they're given. Three foundations have to be in place first.

Without these, an orchestration layer is guessing. With them, it can reason.

The right human, on the right problem, at the right time

With that context in place, the orchestration layer does something no single tool can. It notices that a requirement moved, or that a yield signal doesn't match the model. It walks the thread to find everything upstream and downstream. It traces the linked BOM to see which dies, packages, test programs and suppliers are affected, and re-runs the models that can be re-run quickly. Then it takes the result, with the evidence attached, to the one person who needs to see it: the design lead, the product engineer, the packaging lead or the facilities planner.

It doesn't replace judgment. It moves judgment earlier. Depending on the change, that's hours, days, weeks or months ahead of when conventional methods would have caught it, if they caught it at all. That's where the value of agentic AI shows up: in the difference between adjusting a drawing and breaking up concrete.

Beyond EDA 3.0

At AiT we've called this EDA 3.0: orchestrating the semiconductor lifecycle from intent to yield rather than speeding up individual steps. Our look at the memory wall in AI inference silicon (The Memory Wall Is Reshaping AI Inference Chip Design) pointed the same way.

The ripple above shows the idea doesn't stop at the chip, or even at the fab. Design, manufacturing and the data center are one industrial system. The pattern gets far bigger when it's paired with an industry leader that is bringing AI into every stage of the industrial process. One of the largest technology companies has already moved an agentic R&D platform into general availability, built to coordinate data analysis, hypothesis generation, experimentation and knowledge management through specialized AI agents. It has named semiconductors and advanced manufacturing among its target domains.

See it working, the same evening

This is no longer a whiteboard idea. Working demonstrations are being built on real designs, with real requirement chains and real change requests flowing through them.

If you're at the IEEE panel on the true value and cost of agentic AI in SoC design and manufacturing, Tuesday, October 13 at 12:15 PM in Moscone North Hall, Room 20AJ, you'll hear the question framed well. That evening we're showing one answer, live, at a small invitation-only gathering near Moscone. Request an invitation and we'll save you a spot.

Simon Bennett, CSO and Co-Founder, AiT