Kandou: Redefining High-Speed Connections in the Age of AI

Celebrating 15 years of the Top100 Swiss Startup Award, we spotlight the pioneers who shaped their industries. Among them is Kandou, recognized early in the ranking and now a global leader in high-speed, energy-efficient chip connectivity. Born as an EPFL spin-off and supported by Venture Kick, Kandou has grown into a technology powerhouse with over 350 patents and a footprint across three continents — redefining how electronic systems communicate.

As Top100 marks its 15th anniversary, Kandou stands as a prime example of Swiss deeptech success. From academic research at EPFL to Venture Kick support and global industry adoption, the company has rewritten the rules of chip-to-chip communication. We spoke with founder Amin Shokrollahi about Kandou’s journey from lab innovation to worldwide recognition, competing with global giants, and what’s next for wired connectivity in the era of AI.

Amin Shokrollahi, Co-Founder and CEO of Kandou AI

What was the original problem in chip-to-chip communication that you set out to solve? 
Compute kept producing more data, but the number of pins and traces between chips couldn’t grow — packages, interposers, and boards capped wire count. The industry fix — adding lanes and increasing SerDes speeds — was running into hard limits: power, complexity, and cost. We set out to invert the constraint: instead of adding wires, make each wire carry far more data, reliably and efficiently. That drive led us to a new signaling paradigm, Chord Signaling, which boosts bandwidth per pin while cutting energy per bit. This allows systems to scale data movement without exploding pins, heat, or budgets.

What was your very first prototype or early test like?
Chasing the speeds we envisioned was prohibitively expensive, so we started in simulation — pushing our signaling ideas through advanced EDA tools and corner cases to de-risk the physics. With that confidence, we spun out Kandou and taped out a proof-of-concept in a mature semiconductor node to keep costs and risk under control. That first chip was modest by today’s standards, but it proved the core idea worked on real silicon — more bits per pin, reliably. It marks our birth, and we’re incredibly proud of it.

"We design for the next decade."
 

You secured Venture Kick support while founding Kandou, ultimately totaling CHF 130,000. How did this help you at the time?
Venture Kick was instrumental in those early days. It gave us not only the funding to move from simulation to first silicon, but also the structure and visibility to refine our business case. The staged support forced us to focus on milestones, sharpen our story, and engage with a network of mentors and investors. That combination — capital plus guidance — helped transform a research idea into a company ready to compete in the semiconductor industry.

Which milestones took Kandou from EPFL lab research to industry adoption?
Three milestones shaped that journey:

• Peer validation: After three years of R&D, our work was accepted at ISSCC, earning Best European Paper and our first major external recognition.

• Industry visibility: The ISSCC presentation put Kandou on the map and opened doors with tier-one partners.

• Commercial proof: Collaboration with Marvell integrated our signaling into their switch chips, showing we could combine novel ideas with robust, production-ready implementation.

What has been the toughest challenge in the semiconductor industry?
The toughest challenge has been the industry’s extreme risk aversion. With mask sets costing millions and a single misstep stalling a product line, the default is “don’t be first.” Design cycles are long, qualification is strict, and backward compatibility often outweighs step-change ideas. Moving the needle requires overwhelming proof: silicon that works under all conditions, clean interoperability, measurable gains in bandwidth per pin and energy per bit, and a tier-one customer willing to vouch for it. In short, you don’t just invent — you de-risk every angle until adoption feels safer than staying put.

How has Kandou changed the way chips communicate with each other?
We made chiplets real. Before there was a playbook, we built and shipped production-grade chiplet links with Marvell, proving systems could scale by boosting bandwidth per bump and cutting energy per bit without adding pins or power. That early silicon set benchmarks still used today, and many modern chiplet standards reflect design choices we pioneered. In short, we helped move chiplet connectivity from lab research to industry practice.

Silicon in production: PCIe retimers for hyperscale data centers (top) and AMD-powered laptops (bottom)
 
What role will Kandou play in the semiconductor industry of the next decade?
AI has turned semiconductors into an interconnect problem — and that’s our home turf. Over the next decade, Kandou will continue driving:

• Power-lean bandwidth: high-throughput, low-power links for scalable AI systems.

• Chiplets at scale: turnkey fabrics for plug-and-play integration.

• System cost breakthrough: AI components cut total cost ~10×, boost bandwidth-per-pin, reduce energy/bit.

• Proven at volume: USB and PCIe chips show reliable high-volume production since 2021.

Which early engineering decision still defines Kandou today?
At the start of our Marvell collaboration, we could have shipped a simpler link. Instead, we chose to build a future-proof interconnect — scaling bandwidth per pin, cutting energy per bit, and performing consistently across use cases. That decision still defines us today.