The Chip Nobody Designed by Hand

The “AI Super Chip” event promised a wafer-scale chip “50X faster than Nvidia” and “144X more powerful than Qualcomm,” numbers that point to Cerebras and its giant single-wafer processor. Lost in the speed figures is a simple fact: that chip, like every other, had to be designed in software before it could be fabricated. No one lays out billions of transistors by hand.

That is where Synopsys (SNPS) enters the story. Synopsys makes the electronic design automation, or EDA, tools that every AI chip designer uses, including Cerebras, Nvidia, and every startup chasing them. The free pick at the event was not the flashy chip. It was the software company upstream of that chip. There is more on this in our Brownstone Research’s ‘AI Super Chip Event’ — The Synopsy… article.

We broke down the event itself in the AI Super Chip teardown. The design layer is the quiet part of that story.

Why AI Chips Are Hard to Design

An AI accelerator is not a simple part. It packs thousands of small compute cores, huge banks of memory, and the wiring to move data between them without bottlenecking. The wafer-scale approach takes this further, printing an entire network of cores across one silicon wafer. A single defect can kill a design, so verification becomes enormous.

Each generation adds more. More cores, more memory, more custom blocks tuned to specific math. The design team has to simulate the whole thing before committing hundreds of millions of dollars to fabrication. One mistake at the design stage means a chip that does not work, and there is no software patch for that.

That is why EDA tools have become the toll booth of the AI buildout. A harder design problem means more tool usage, more simulation runs, and more revenue for the duopoly that sells the software. The math is direct.

Chiplets Make It Worse

The newest accelerator designs break the chip into chiplets, smaller pieces assembled in one package. That approach helps yields but makes the design problem more complicated, not less. Each chiplet has to be designed, verified, and then checked again for how it talks to its neighbors.

Custom accelerators are multiplying too. A company that wants a chip tuned to its own AI model will design something unique rather than buying an off-the-shelf GPU. Every one of those custom parts goes through the same EDA flow.

For the fundamentals of what that tool flow looks like from start to finish, the chip design software explainer walks through the designer’s workflow.

Upstream of Every Accelerator

The investment logic is the picks-and-shovels argument applied to chips. Whether the winning accelerator is a wafer-scale engine from Cerebras, a GPU from Nvidia, or a custom part from a startup, it is designed in EDA software. Synopsys collects revenue from all of them without betting on which one wins.

That is a different risk profile from owning the chipmaker itself. A chip company can lose a design cycle or a customer and see revenue swing sharply. The tool vendor sees the whole industry’s design spending, smoothed across every participant.

It also explains why the event’s free pick was more conservative than its marketing. A “50X faster” chip is a claim about one product. The design software behind it is a claim about the entire category.

Verification Is the Bottleneck

When a chip is small, engineers can test its behavior exhaustively. When it holds billions of transistors, they cannot. The number of possible states a chip can enter grows far faster than any team can enumerate, which is why verification now consumes more engineering hours than the design itself. Teams spend months proving a chip will behave correctly across cases they will never actually run in the lab.

Synopsys sells both the simulation tools and the formal verification methods that make that proof possible, and every wafer-scale engine and GPU passes through the same gauntlet. The gauntlet also gets longer each generation, because each new node adds states faster than it adds engineers.

That is the quiet reason design complexity turns directly into tool revenue: the hardest part of a chip project, measured in hours and headcount, is precisely the part the software is paid to automate. It also explains the pricing. The verification budget is a small slice of a program’s total cost, but it is the slice that decides whether the program ships at all.

The Verdict

AI chip design is getting harder, not easier, and harder design work is the raw material for EDA revenue. Synopsys sits at the top of that food chain, collecting a toll on every accelerator that reaches the fab.

The connection to the teaser is real but narrower than the hype suggested. The wafer-scale chip was the showpiece; the software used to design it was the pick. That is a sensible place to look for durable exposure to AI without betting on a single chip architecture.

Ready to see the research? Click here to access the Brownstone Research report.

NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.