The semiconductor industry has a consensus about what happens at AI rack scale: copper wiring runs out of breath somewhere north of 200 gigabits per second, and optical interconnects take over. Kandou AI just raised $225 million to argue that consensus is wrong.
Kandou closed a Series A in April 2026 led by Maverick Silicon, with SoftBank, Synopsys, Cadence Design Systems, and Alchip Technologies chipping in. The thesis is built on a signaling technique called Copper MIMO, or chord signaling, which treats the crosstalk between adjacent wires as an asset rather than noise. In a conventional interconnect design, coupling between neighboring wires degrades signal quality and has to be compensated with additional power and equalization circuits. Kandou inverts the problem: it applies a mathematical transformation at the transmitter that turns crosstalk into usable signal energy, and undoes it at the receiver. The result, according to the company, is higher data density without increasing physical complexity.
The technical claim is not trivial. The industry is moving to 224 gigabits per second per lane because generative AI workloads demand it. At those speeds, power consumption climbs and signal integrity degrades. The conventional answer is optical interconnects, which avoid the copper limitation entirely but introduce electro-optical conversion overhead and latency. Kandou's bet is that copper still has unused capacity, and that information theory shows you how to get it.
"The industry is operating far below the Shannon limit," said Srujan Linga, Kandou's CEO, referring to the theoretical maximum data capacity of a communication channel. "We have developed techniques, backed by about 500 patents, to increase the effective capacity of copper links."
The investor list is worth examining. SoftBank has broad AI infrastructure interests. Synopsys and Cadence sell the electronic design automation tools that every chip company uses, which means they have visibility into which interconnect approaches their customers are actually specifying for next-generation designs. Alchip makes ASICs and interposer-based packaging solutions. None of these are passive financial investors; they are strategic actors reading the same roadmap.
The performance numbers the company cites are large. Compared to existing interconnect approaches, Kandou claims it can reduce cost by roughly 12 times in certain cases, improve reach and scalability by up to 10 times, and cut power consumption by about 3 times. Those figures come from the company and have not been independently validated. Kandou has shipped more than 20 million units across non-AI applications, giving it some operational credibility, but the AI datacenter deployments are not yet wide.
The longer reach claim is potentially the most significant. Longer reach means fewer retimers, which means fewer chips, lower power, and lower system cost. It also means standard PCBs and simpler packaging can substitute for silicon interposers, which are expensive and supply-constrained. If that claim holds at 224 Gbps and above, it changes the bill of materials for AI racks in ways that matter at hyperscale.
Taher Madraswala, Kandou's COO, was direct about the talent constraint. "Analog circuit design is the hardest to find," he said. "A lot of our analog talent currently is in Europe." The company opened a design center in Hyderabad, India, to expand its engineering bench.
The optical interconnect transition is not theoretical. Companies like Credo, Juniper Networks, and the optical businesses of larger players are building businesses around it. If Kandou is right, they are building on an assumption that will be disrupted. If Kandou is wrong, or if its technology proves harder to manufacture or integrate than the company predicts, the optical road remains open. The $225 million is a bet placed at the frontier where physics meets supply chain economics.
† Performance figures (12x cost reduction, 10x reach improvement, 3x power reduction) and patent count (500 patents) are sourced from Kandou AI and have not been independently verified.
