NexTech Frontier: Welcome Silicon Photonics to the Quantum Game
- Danielle Franklin
- Oct 9
- 5 min read
The data demands of AI factories are officially too big for copper. NVIDIA has made it clear: the future of high-performance computing depends on light.
Instead of relying on long electrical traces and pluggable transceivers, NVIDIA is embracing silicon photonics and co-packaged optics to move data with radically lower power and higher bandwidth. This isn’t theory—it’s shipping product strategy.
Moving Beyond CMOS
CMOS fabs were designed for electrons—not photons. But the industry is now stretching and evolving beyond traditional constraints:
Heterogeneous integration of lasers and modulators
Back-end photonic layers on CMOS
Non-silicon materials for nonlinear optics and Pockels effects
3D chip stacking (electronics + photonics)
This is why terms like post-CMOS integration and photon-electronic convergence are now everywhere.
📈 Why This Matters Now
Silicon photonics is no longer “emerging tech”—it’s the backbone of next-generation:
AI factories
Defense communications
Cloud networks
Exascale computing
Spaceborne data links
Quantum networkin
🔹 NVIDIA’s Co-Packaged Optics Rollout
Their new Spectrum-X and Quantum-X switches integrate photonics directly with networking silicon—eliminating the energy and latency penalties of traditional electrical interconnects.
Source:
🔹 Why It Matters
AI data centers can no longer scale with copper and CMOS-only designs. Optical interconnects slash energy per bit, remove PCB bottlenecks, and enable GPU clusters at planetary scale.
Tom’s Hardware reports NVIDIA plans to deploy silicon photonics in AI systems as early as 2026.
Source:
🔹 Beyond CMOS — The Materials Shift
Silicon photonics started by “piggybacking” on CMOS manufacturing—but next-gen systems require heterogenous integration: III-V lasers, Pockels materials, back-end photonics, and 3D hybrid stacks.
A recent Nature Communications roadmap outlines precisely how the industry is moving past CMOS’s optical limits.
Source:
For deeper context on post-CMOS integration challenges and strategies, Optica’s June 2025 review is worth a read:
🔹 Supply Chain Momentum
Even the laser providers are being pulled in. Lumentum announced its role in NVIDIA’s silicon photonics ecosystem to support AI scaling.
Source:
⚡ The Takeaway
We’re not watching an experiment—we’re watching the copper era quietly exit the data center. Silicon photonics is no longer “emerging tech.” It’s infrastructure. And moving beyond CMOS isn’t about replacing it—it’s about extending it with light.
What is Silicon Photonics — Basics & Motivation
Silicon photonics refers to photonic (i.e. optical) devices and circuits built using silicon and silicon-compatible processes (often silicon-on-insulator, or SOI). The idea is to leverage the mature CMOS / semiconductor fabrication ecosystem (e.g. 300 mm silicon foundries, lithography, cleanroom infrastructure) to build optical waveguides, modulators, photodetectors, filters, etc., integrated on chip or in hybrid form.
Some of the principal advantages and motivations include:
Economies of scale / cost leverage: Using silicon fabs and CMOS-compatible processes lowers marginal cost vs exotic optical platforms.
Compactness & integration: On-chip or tightly packaged optical links reduce footprint, reduce coupling losses, and lower signal path lengths.
Lower power per bit: Compared to long copper interconnects or discrete optical modules, silicon photonics can reduce energy per transmitted bit, especially when integrated closely.
Bandwidth scaling: Optical links scale more gracefully with distance and frequency than electrical ones, which are limited by resistive, capacitive, and signal-integrity issues over longer traces.
However, silicon is not a perfect material for all optical functions: it is an indirect bandgap semiconductor, making on-chip light emission (lasers) difficult; it has non-ideal electro-optic coefficients, and there are thermal, optical loss, and integration constraints. Hybrid strategies (bonding III–V materials, heterogeneous integration) are often used to address those limitations.
The review “Roadmapping the next generation of silicon photonics” is particularly useful for understanding where the field is and where it’s headed.
Also, a perspective article on post-CMOS photonic integration discusses how to tightly couple photonics and electronics beyond the traditional CMOS limits.
NVIDIA’s Use of Silicon Photonics & Their Strategy
In recent years, NVIDIA has pushed into using silicon photonics (and co-packaged optics) in its networking / interconnect domain, especially for scaling AI “factories” (massive GPU clusters) with high bandwidth and low power overhead.
Co-packaged optics (CPO) and photonic switching
One of the cornerstones of NVIDIA’s approach is co-packaged optics (CPO). The idea: instead of having separate pluggable optical transceivers connected via PCBs and trace lengths, integrate the optical engine (photonics) very close (or even on) the switch ASIC/package. This drastically shortens the electrical trace lengths, reduces losses and parasitics, and eliminates separate DSPs and long PCB paths.
In 2025, NVIDIA announced Spectrum-X (Ethernet) and Quantum-X (InfiniBand) silicon photonics networking switches. These are built with co-packaged optics to scale AI data center networks.
The announcements claim about 3.5× energy savings, 10× resilience / network uptime, lower latency, and faster time to deploy compared to conventional pluggable-transceiver based networks.
The photonic systems use fewer lasers, more integrated modulators, and tighter integration of electronics and optics.
NVIDIA reports partnering with TSMC (for photonic process / integration), and multiple optical / photonics supply chain partners (e.g. Coherent, Lumentum) to build the ecosystem.
For example, Lumentum was selected as part of the NVIDIA silicon photonics supply chain, providing high-efficiency lasers.
A TweakTown article describes that at Hot Chips 2025 NVIDIA presented Spectrum-X as a “game changer,” reporting that AI data centers consume far more optics power. It mentions that in an AI factory, optics costs might account for ~10% of compute power, motivating tighter integration.
The Tom’s Hardware coverage states NVIDIA plans to roll out these photonic interconnects by 2026.
Another article by NextPlatform says NVIDIA is weaving silicon photonics into its Quantum (InfiniBand) and Spectrum (Ethernet) families to reduce network power.
The optics.org news covers NVIDIA’s reveal and plan to scale AI “factories” using co-packaged optics.
IEEE Spectrum also covered NV’s optical network switch announcement in depth.
NVIDIA is aggressively pushing to embed photonics closely with its switching / interconnect fabric to reduce power, latency, and complexity, especially for scaling AI compute.
Broader industry turning to silicon photonics and co-packaged optics
AI-scale GPU clusters require enormous bandwidth between nodes, across racks, even across pods. Electrical interconnects (copper traces, cables) struggle at the very high speeds and distances required (signal integrity, latency, power).
Long PCB or cable traces, parasitic capacitance/inductance, and the need for DSP/serdes overhead, all degrade performance at high rates.
Traditional pluggable transceivers (e.g. SFP, QSFP) have overhead, insertion losses, and power consumption.
As NVIDIA scales up to multi-million GPU “factories,” interconnect power overhead becomes a significant fraction of overall system power. Tight integration of photonics can reduce power per bit, reduce overhead, and make scaling more efficient.
Thus, by embedding photonics close to or on chip (or package), NVIDIA can bypass many of the inefficiencies of conventional optical transceivers and electrical interconnects.
Approach | Description | Maturity |
Co-Packaged Optics (CPO) | Photonic engines integrated with switch ASICs | Shipping now |
On-Package Photonics | Silicon photonics dies bonded near compute | Ramping |
Monolithic Integration | Electronics + photonics on the same wafer | Emerging |
Hybrid III-V on Silicon | Laser materials bonded to silicon | Active research |
Comments