NVIDIA does not provide stitches without thread. At the end of August 2025, the company led by Jensen Huang announced that in 2026 their platforms artificial intelligence next generation (AI) will use photonic interconnections to achieve higher transfer speeds between GPU clusters. This announcement came during the conference specializing in semiconductor engineering and high-performance computing ‘Hot Chips’, which was held in Palo Alto (California), and was just the prelude to what was to come.
And this same week NVIDIA has revealed that is going to invest 2,000 million dollars in Lumentum, and the same amount in Coherent. These two companies have something very important in common: they are specialized in developing photonic technologies. Shortly after NVIDIA confirmed its interest in them, the shares of these two companies rose 5 and 9% respectively. And the company led by Jensen Huang has committed to purchasing products from Lumentum and Coherent for several billion dollars, and also to use their advanced laser solutions and optical networking technologies.
Photonics is the support that cutting-edge semiconductors need
Most IC designers and manufacturers are working on the development of silicon photonics. Douglas Yu, a TSMC executive with responsibility for systems integration, explained in September 2023 very clearly what disruptive capacity this technology has: “If we manage to implement a good integration system for silicon photonics, we will unleash a new paradigm. We will probably place ourselves at the beginning of a new era.”
Silicon photonics is a discipline that in the field in question seeks to develop the technology of this chemical element to optimize the transformation of electrical signals into light pulses. The most obvious field of application of this innovation is implementing high performance links which, on paper, can be used both to resolve communications between several chips and to optimize the transfer of information between several machines.
In AI clusters, thousands of GPUs must work in unison, so it is essential to connect them using high-performance links
The advanced packaging technologies used by leading semiconductor manufacturers, such as TSMC, Intel or Samsung, can greatly benefit from a very high-performance inter-chip communication mechanism. And large data centers where it is necessary to connect a large number of machines, too. However, there is one discipline in particular that has an overwhelming future projection and that would benefit greatly from building on the advantages offered by silicon photonics: AI.
This is precisely NVIDIA’s bet. In AI clusters, thousands of GPUs must work in unison, so it is essential to connect them using high-performance links. It is possible to solve this challenge using traditional copper cables or optical modules, but both of these solutions introduce into the infrastructure very important inefficiencies. The most problematic are energy loss and bottlenecks. Data transfer can consume up to 30 watts per port, which increases energy dissipation as heat and increases the likelihood of failure.
Additionally, latency limits the scalability of clusters as the number of GPUs in data centers increases. To resolve these inefficiencies, NVIDIA will integrate the optical components required for photonic interconnections into the same switching chip package. This technology is known as CPO (Co-Packaged Optics) and manages to reduce power consumption to only 9 watts per port. Additionally, it minimizes signal loss and improves data integrity. Looks really good.
NVIDIA has confirmed that it will integrate CPO technology into its Quantum-X InfiniBand and Spectrum-X Ethernet interconnect platforms during 2026. However, there is something important that is worth not overlooking: CPO is not going to be an extra. When it arrives, it will be established as a structural requirement of the next generation of AI data centers in a clear attempt to increase the competitiveness of NVIDIA’s AI hardware platforms.
Image | Generated by Xataka with Gemini
More information | Reuters
GIPHY App Key not set. Please check settings