Eight Chiplets Walk Into a Bot Build

Dr. Nasrullah, speaking at Chiplet Summit 2026, framed the problem in a way that stuck with me. AI’s energy trajectory is heading toward a 10GW challenge, and his answer was not a bigger chip or a better process node. It was four techniques for designing chiplets, starting with node mixing. That framing is a small act of honesty: we are not going to compute our way out of this with monolithic silicon.

I build bots. I write orchestration code, I wire up inference endpoints, I watch latency graphs. Silicon packaging is several layers below anything I touch directly. So why am I writing about it? Because the bill for every bot I ship eventually lands on a piece of hardware someone had to design, and the cost of that design shows up in what I pay per token.

What co-design actually means for the rest of us

Chiplets are the idea that instead of etching one enormous die, you build a package out of smaller specialized dies stitched together. Co-design is the part that matters more: designing the chiplets and the system that uses them at the same time, rather than throwing parts over a wall and hoping the interconnect sorts it out.

The University of Michigan work on co-design frameworks reports reductions in both energy and design cost for AI accelerators. Two numbers moving in the same direction is unusual. Normally in hardware you trade one for the other: spend more engineering effort to save power, or ship faster and eat the inefficiency. A framework that pulls both down suggests the old approach was leaving value on the table simply because the pieces were designed separately.

The Fengshui research puts sharper numbers on the payoff. A pool of just eight co-optimized chiplets, covering different dataflows, processing-in-memory, and network switches, produced a 48.5% reduction in energy and 88.1%, 93.0%, and 97.8% reductions across the other metrics they tracked, including energy-delay product.

Eight. Not eight hundred. That detail is the one I keep turning over.

Why a small parts bin is the interesting result

Anyone who has built software systems recognizes this shape. You do not need a custom component for every use case. You need a small set of well-chosen primitives that compose. Eight chiplets covering distinct roles, mixed and matched, beats a sprawling catalog of bespoke designs that each need their own verification cycle.

That is the same reasoning behind every decent bot architecture I have written. A handful of solid tools with clean interfaces, an orchestrator that routes between them, and you can cover an enormous range of tasks. Swap the router logic, get a different bot. No rewrite. The chiplet people arrived at the same conclusion from the opposite direction, and they had to invent advanced packaging to get there.

The design cost angle is underrated

Energy savings get the headlines because power bills and carbon are easy to talk about. Design cost reduction is quieter and possibly more consequential. When it gets cheaper to design an accelerator, more people design accelerators. You stop needing a hyperscaler budget to build silicon aimed at a specific workload.

For those of us building on top, that means the accelerator space could get less uniform. Instead of one general-purpose part that everyone tunes their models toward, you might see parts shaped around particular inference patterns. Retrieval-heavy workloads look nothing like long-context generation, which looks nothing like the small classifier models a lot of production bots quietly depend on.

The ecosystem question

None of this works if chiplets from different vendors cannot talk to each other. That is why the public programs matter. The U.S. CHIPS National Advanced Packaging Manufacturing Program explicitly includes chiplet ecosystems and co-design in its scope, and vendors like Cadence are building platform tooling for chiplet-based SoC design. There is also an open benchmark now for evaluating AI thermal models in 2.5D packaging, which tells you heat is a first-class concern rather than an afterthought.

Interoperability is where good hardware ideas usually go to die. A parts bin is only useful if the parts fit, and fitting means agreed-upon interconnect, agreed-upon thermal budgets, agreed-upon test methodology. The tooling and benchmark work suggests people understand this, but understanding a coordination problem and solving it are different projects.

What I am taking from this

Nothing here changes what I write tomorrow. I am not going to refactor a bot because of a packaging paper. But the direction is worth tracking for anyone who thinks about where their inference runs and what it costs.

The composability lesson generalizes, though. Eight well-chosen pieces outperforming a hundred specialized ones is a principle, not a coincidence of silicon. When you are designing your own agent stack, resist the pull toward a custom tool for every case. Build the small set that composes. Let the routing layer do the work.

The hardware people spent years and considerable money learning that. We can just read the paper.

🕒 Published: September 19, 2026