Much like the transition from ADSL to fiber optics in everyday use, artificial intelligence is currently undergoing a similar transformation, but within its very infrastructure. It is no longer just the chips that matter, but rather the interconnects that link them.
When the network becomes the bottleneck
With the surge in AI models, the volumes of data to be processed and exchanged are exploding. These flows must circulate ever faster, with minimal loss and controlled energy consumption. However, in many cases, these exchanges still rely on electrical signals traveling through copper cables.
At ultra-high speeds, typically between 400 and 800 Gbits/s, these signals reach their physical limits. They degrade, generate heat, and ultimately slow down the entire system. The problem is now clearly identified: the network has become a bottleneck, sometimes even more restrictive than the GPUs themselves.
Silicon photonics: A key technological response
Faced with this constraint, a solution is gradually gaining ground: replacing electrical signals with light signals. This is the principle of silicon photonics, often referred to by the acronym SiPh.
The concept is simple on paper: using light instead of electricity to transport information. In practice, the benefits are substantial. Data transfers become significantly faster, signal losses are reduced, and energy efficiency improves markedly. This is a crucial point at a time when data centers are showing titanic levels of consumption.
Integration ever closer to the processors
Beyond the nature of the signal, the challenge also lies in how these technologies are integrated around the processor, a process known as packaging. Several approaches coexist, notably CPO (Co-Packaged Optics) and NPO or OBO (Near-Packaged Optics/On-Board Optics) architectures.
All pursue a common goal: to bring optical components as close as possible to the processors. By reducing the distance between computation and transmission, these solutions limit losses, decrease heat, and improve the overall performance of the systems.
These innovations are shaping the future of ultra-high-speed computing, making it more efficient and better suited to the demands of modern artificial intelligence.
An industrial supply chain already in place
This transformation is accompanied by a massive mobilization of technology players. Nvidia, at the forefront of AI, is investing heavily in photonics, both internally and through strategic partnerships with specialists such as Lumentum and Coherent, both recognized for their laser technologies.
Marvell is following the same momentum with the acquisition of Celestial AI, a player entirely dedicated to optics for artificial intelligence. Meanwhile, Intel is already mass-producing 400G optical chips for data centers, while Cisco provides a large portion of the 800G optical modules used by hyperscalers.
Broadcom integrates these technologies into its network switches, notably the Tomahawk range, which is widely deployed in infrastructure. Other players play a key role in the value chain: GlobalFoundries manufactures photonic chips for several companies, while Ayar Labs develops optical chiplets like TeraPHY, which can be directly integrated into processors.
IBM is also contributing with components dedicated to short optical links, essential in CPO architectures, while STMicroelectronics is exploring applications ranging to optical sensors for autonomous vehicles.
Europe also in the race
While American giants dominate the landscape, several European companies are holding their own. STMicroelectronics develops and industrializes silicon photonics platforms for high-speed interconnections. Soitec provides the substrates essential for these circuits, while Nokia designs coherent optical solutions integrating these technologies.
Toward a new technological race
After years focused on computing power, the industry is entering a new phase. Competition is no longer played out solely on chip performance, but on the speed and efficiency of the connections that link them.
