In data centers that train the most advanced artificial intelligence models, the problem is no longer just to calculate, but to circulate increasingly massive volumes of data between GPUs, memories and internal networks. As models increase in size and complexity, this circulation of data becomes the determining factor in performance.
It is on this precise point that Kandou AI is positioned. The Lausanne-based company has just raised approximately 259 million euros (225 million pounds) in a strategic round bringing together Maverick Silicon, as well as SoftBank Group, Synopsys, Cadence Design Systems and Alchip Technologies.
The shift: from calculation to circulation
For more than a decade, the performance of computer systems has been thought of as an almost exclusive function of computing power. The rise of artificial intelligence has extended this logic, with a concentration of investments on GPUs and specialized accelerators, but this equation today shows its limits.
In a modern AI cluster, components must constantly exchange quantities of data that grow faster than the computing capabilities themselves. Memory is expanding, architectures are distributed, models use thousands of chips. In this context, the bandwidth and latency of interconnections become decisive.
Overall performance no longer depends solely on the available power, but on the ability to effectively orchestrate flows between components. This is what Kandou summarizes by evoking a paradigm shift: “GPUs and CPUs must be connected to exponentially growing volumes of memory, which makes memory bandwidth and interconnections major bottlenecks”.
A discreet but structuring market
These trends are bringing to the fore the high-speed interconnection market, which has long remained in the background. It brings together all the technologies allowing data to be transported between chips, racks and systems, SerDes (serializer/deserializer), retimers, copper or optical interfaces.
These bricks are rarely visible, and do not constitute finished products, but essential intermediate layers. However, their role becomes central as architectures become more complex.
The issue is twofold:
- increase bandwidthto absorb the growth in flows
- reduce energy consumptionas data centers reach physical and economic limits
In this context, Kandou claims an approach based on a reinvention of copper interconnections, with signaling technologies from information theory and semiconductor design. The objective is to achieve performance levels comparable to certain optical solutions, while maintaining more favorable cost and energy efficiency.
Copper’s bet, against the rise of optics
One of the key trade-offs in AI infrastructure today is between copper and optical.
Optics offers high performance over long distances, but remains expensive and energy-intensive at large scales. Copper, historically more limited, has the advantage of lower cost and simpler integration. Innovation therefore consists of pushing one’s physical limits.
Kandou fits into this logic. “Our goal is to transform AI hardware with products built on our copper interconnects and IP systems,” says Srujan Linga, co-founder and CEO of the company. “Our first products will target AI systems and high-speed rack-level connectivity, beyond 448G, with the ambition to gradually cover the entire infrastructure. »
A cap table revealing a systemic issue
The round brings together investors with varied profiles: alongside Maverick Silicon, Synopsys and Cadence Design Systems are at the heart of chip design tools. Alchip Technologies is involved in the design and manufacturing of custom circuits. SoftBank Group, for its part, is increasing its positions in critical AI infrastructures.
Physical AI infrastructure emerges as a new investment field for VCs
Beyond the Kandou case, this lifting reflects a broader trend with numerous investments in numerous infrastructure and industrial projects: energy, cooling, interconnections, memory.
Although this type of operation remains rare in Europe, several funds are nevertheless starting to position themselves in these infrastructure building blocks. Players like Atlantic Bridge, IQ Capital and Vsquared Ventures have gradually built a thesis around semiconductors, computing architectures and the physical constraints of AI. In France, investors are becoming rarer, notably with Supernova Invest. But at this stage, few European funds have the capacity, both financial and organizational, to carry out rounds worth several hundred million euros on such capital-intensive technologies on their own. The financing of AI infrastructure thus remains largely structured around international coalitions, where European investors still play a supporting role rather than leadership.
More than 500 million dollars raised to structure a disruptive technology
Since its founding in 2011 by Srujan Linga and Dr. Amin Shokrollahi, Kandou AI has followed a gradual funding trajectory, characteristic of capital-intensive deeptech companies. After a pre-seed supported by Venture Kick, the company raised around $10 million in 2012, before launching several growth rounds led in particular by Bessemer Venture Partners. The announced strategic round, worth 225 million dollars (around 209 million euros), marks a change of scale, both in terms of its volume and the nature of its investors. It brings the total funding raised by Kandou to more than $500 million.
