Silicon Photonics: How Light Helps AI Data Centers and What Its Limits Are

When it comes to headlines about “light chips,” one thing needs to be clarified right from the start: in data centers today, light isn’t used for computing—it’s used to transmit data. Silicon photonics replaces electrical connections between chips with optical ones, which can reduce power consumption and increase bandwidth. However, this is not a solution to the “physical limit” of artificial intelligence, but rather a solution to one specific obstacle.

Why Electrical Connections Are an Obstacle

Large AI systems interconnect thousands of accelerators that constantly send data back and forth. At higher speeds, the signal in copper wires degrades more rapidly with distance, so it must be regenerated using additional circuits that consume energy (this is a general explanation). According to a review article published on nature.com, network interconnects account for nearly 10% of total computing power consumption, and this share increases as speeds rise.

The IEA provides further context: electricity consumption in data centers is expected to double by 2030. Savings from interconnections can therefore help, but they will not solve the energy demand problem on their own.

How Silicon Photonics Works

Here is a general explanation: an electrical signal is converted into a light signal in a modulator; this light travels through an optical waveguide formed in silicon; and on the other side, a detector converts it back into an electrical signal. The advantage is that the components can be manufactured using processes similar to those for conventional chips. The laser itself, however, is usually made from materials other than silicon and is added to the system as a separate component.

What Does “Co-Packaged Optics” Mean?

In traditional solutions, optical modules are plug-in types and are separated from the chip by a longer electrical path. With so-called co-packaged optics (CPO), the optical module is placed directly into a single package with the network chip. The electrical path is shortened to millimeters, which should reduce both power consumption and signal loss. NVIDIA states that its solution offers 3.5 times higher energy efficiency, 10 times greater network resilience, and 1.3 times faster deployment compared to plug-in modules. These are the manufacturer’s claims, not independent measurements.

What’s Already in Operation and What Isn’t Yet

According to a July 27, 2026, press release from the analytics firm TrendForce, NVIDIA has begun shipping its Spectrum-X switch with CPO to select partners, and Broadcom continues limited shipments of the Bailly switch (51.2 Tb/s). TrendForce describes this as a transition to mass production, though it also notes that volumes remain small. Broadcom has been shipping CPO solutions to customers since 2024.

Challenges and Limitations

• Optical module production. TrendForce cites the yield of optical modules, silicon photonic chips, and advanced chip packaging capacity as the main bottlenecks.
• Competition for Capacity. According to the same report, CPO packaging shares manufacturing capacity with AI chips, which limits the growth of shipments.
• Uneven Adoption. The authors of an article on nature.com note that the adoption of photonic technologies in data centers is proceeding unevenly and that system reliability remains an important goal.
• Service (my explanation). When optics are integrated into the chip package, repairing a damaged component is more complicated than replacing a plug-in module. The sources I found do not discuss this issue in detail.

What to Expect Next

According to a comment from the OFC 2026 conference, NVIDIA was expected to unveil plans to use CPO for interconnects within systems (so-called “scale-up”) in 2027–2028. This is the opinion of the comment’s author and the manufacturer’s outlook, not a confirmed product.

Conclusion

Silicon photonics addresses a real-world problem: data transmission between chips is energy-intensive, and copper interconnects are reaching their limits. Optical interconnects are already making their way into mass-produced switches, though currently in small volumes and with manufacturing constraints. The reduction in power consumption applies to the network, not the computation itself; therefore, it is not the answer to the overall growth in energy consumption driven by artificial intelligence.