Every AI cluster has a plumbing problem, and it is rarely the one that gets talked about on stage. Tower Semiconductor says it is scaling silicon photonics for 1.6T data center optical modules designed for Nvidia networking protocols, which drops a specialty analog foundry into the middle of how AI racks talk to one another.
Start with the unit. A 1.6T module shifts 1.6 terabits of data per second, and it lives at the edge of a switch or a server, turning electrical signals into light and back again. Copper runs out of road at those speeds over anything longer than a couple of metres. Light does not, which is why so much of the hard engineering has migrated into the transceiver.
Silicon photonics is the practice of building the optical furniture, the modulators, the waveguides, the detectors, on the same kind of silicon wafer that makes ordinary logic chips. Picture swapping a hand-soldered hi-fi for a printed circuit board. Parts shrink, cost falls, and a fab turns out thousands of identical copies where a workshop managed dozens.
The headline number is a doubling. Tower says its silicon photonics enables up to double the data rate of prior silicon photonics generations, giving more bandwidth and throughput for optical connectivity. No benchmark, no methodology and no third-party measurement arrived with that figure, so it stands for now as the foundry's own characterisation rather than a measured result.
Russell Ellwanger, Tower's chief executive, framed the work as a continuation rather than a leap. "Tower Semiconductor is proud to deliver advanced, high-speed technologies that support demanding data center and AI requirements," he said. Ellwanger also pointed at the spending behind it: "We continue to invest significantly across our SiGe and silicon photonics platforms to support the ecosystem with industry-leading performance, scalability, and manufacturability, enabling customers to advance next-generation data center architectures." SiGe there is silicon germanium, the high-frequency analog process that does the electrical heavy lifting alongside the optics.
Nvidia's side came from Gilad Shainer, its senior vice president for networking. "The exponential growth of AI is driving the need for a new class of high-speed, scalable networking to connect AI infrastructure," he said, adding that "NVIDIA is collaborating with Tower Semiconductor to advance the ecosystem, enabling more efficient AI infrastructure through next-generation silicon photonics and accelerating AI applications at scale."
What neither company said is at least as interesting. The announcement carries no volume commitment, no named module customer, no production timeline and no revenue figure. Tower sells wafers to the firms that build transceivers rather than shipping finished modules itself, so the money arrives second-hand, one step behind whoever wins the design socket.
Geography is part of the pitch. Tower runs one 200mm plant in Israel, two in the United States and two in Japan through a 51 percent holding in TPSCo, shares a 300mm facility in Agrate with STMicroelectronics, and has access to a capacity corridor inside Intel's New Mexico factory. For a buyer nervous about single-sourcing optics, that spread is the argument.
Whether it becomes a large business turns on a question the announcement does not touch, namely how many 1.6T ports the next wave of AI clusters actually buys. If optics keep taking a bigger share of the interconnect budget, a foundry able to print photonics and silicon germanium under one roof would sit in a comfortable spot. If copper hangs on inside the rack longer than the optics camp expects, the wait gets longer.
Leave a Reply