Currently, most data centers use both electronic and optical communication protocols, for different tasks. Scale-out networks connect thousands of AI computers across a data center, making optics the obvious choice for long distances. Scale-up networks connect several GPUs inside a single mega-computer or rack, where latency is critical, and dense copper interconnects, such as Nvidia’s NVLink, have long been the engineering solution of choice.

That distinction is beginning to blur, and Nvidia may be quietly dipping its toes in the water of optics for a scale-up. In 2025, the company introduced NVLink Fusion, a program that allows hyperscalers and cloud providers to build custom AI systems around Nvidia’s scale-up fabric. This summer, the company’s list of partners has grown to include several photonics players such as Ayar Labs, Marvell Technologies, and Lightmatter.

As AI needs grow, the number of GPUs and the bandwidth of connections between them continues to grow. Electrical links are being pushed toward terabit-per-second signaling. But higher frequencies increase attenuation, power consumption, and heat. To cope, copper cables must become shorter and thicker, making it more difficult to route them through crowded server racks. At the same time, companies like Nvidia are planning to add even more processors to racks, going from 72 GPUs today to as many as 576 by 2027.

“The physics of copper just changes as you increase the frequency of the signals going across that copper,” says Nvidia senior manager Jesse Clayton. “You can mitigate that by limiting the length of the cable, and right now all of our copper cabling is within that single NVL72 rack, so the distances aren’t that long, but we are getting close to the limits of what we can push.”

Engineers call this the “copper wall.” Many now believe that keeping up with AI will eventually require moving optical interconnects closer to the processors themselves.

Bringing optics to the GPU

To move data as light, engineers must figure out how to convert electrical signals into optical ones, then integrate lasers, photonic devices, and electronic chips into a single package without blowing up cost, power consumption, or manufacturing complexity.

Those hurdles are no longer insurmountable.

“More than any other time that I recall, I think it’s concluded that the co-packaged optics will happen,” says Keren Bergman, a professor of electrical engineering at Columbia University.

Ayar Labs, one of the photonics companies participating in the NVLink Fusion ecosystem, has developed optical chiplets meant to sit alongside GPUs and other processors, converting electrical signals into light only inches from the compute silicon. “The most optimal way to do that is having a photonic chiplet with an electronic chiplet and having them hybrid bonded together,” says Ayar’s director of product management, Vishal Chandrasekar.

He argues that optical scale-up has become practical because the semiconductor manufacturing ecosystem has matured around co-packaged optics. Advances in hybrid bonding now allow electronic and photonic chiplets to be manufactured separately, then integrated into a single optical engine that sits beside GPUs or switches.

“What has really happened in the last few years is that the process maturity coming out of the fabs has really improved,” Chandrasekar says. He believes the industry is now on a path toward high-volume optical scale-up systems within the next couple of years.

Lightmatter has designed a 3D photonic connector with input and output ports across the entire chip area instead of only along the edges. Lightmatter

Lightmatter has a different solution. Rather than placing optical chiplets beside processors, the company is building a photonic interposer that serves as the packaging substrate itself. The idea is that future processors could be stacked directly on top of a silicon photonics engine. Vice president of product Roy Kim describes these interposers and optical chiplets as “complementary steps in the photonics road map.”

According to Kim, packaging is no longer the main obstacle. With standard foundries and assembly partners, optical interconnects can now be manufactured, tested, and integrated in ways that look a lot more like conventional chip production, he says.

The remaining challenge is laser integration. Today’s pluggable laser modules take up valuable rack space and are hard to scale. Instead, Lightmatter is putting large numbers of lasers directly onto silicon, which Kim believes could soon support much denser optical scale-up fabrics.

The future of optical scale-up

Nvidia, for its part, is taking a gradual approach to adopting these technologies. Clayton says the company expects optics to move into scale-up networking eventually, but only when the technology is mature enough to justify the transition.

“I think we’ve taken an approach of migrating to optical when it makes the most sense for our platform,” he says. “If you migrate the entire platform at once, you take a tremendous amount of risk on new designs. So, starting at the scale-out space, and then in the future moving to scale-up is kind of a sensible, measured approach from our perspective.”

That slow-and-steady attitude is one reason why some researchers view NVLink Fusion as more than an interoperability play. “The Fusion is sort of this umbrella—you can put copper in it, you can put photonics in it. It’s very photonics friendly,” Bergman says. Rather than committing Nvidia to a specific interconnect technology, she says, Fusion creates an ecosystem in which electrical and optical approaches can evolve side by side.

Not everyone expects photonics to become the only answer. Researchers continue to improve electrical interconnects through advances in signaling, packaging, and transceiver design. And other groups are pursuing alternative technologies.

Underneath all these advances is a bigger issue: Can optical scale-up become something the industry at large can do, rather than a proprietary feature of Nvidia’s ecosystem?

“Absolutely,” Chandrasekar says. “There are going to be multiple implementations in the 2028 time frame in very high volume.”

If that happens, future AI systems could span multiple racks while behaving as a single computing domain, connected via a mix of electrical, optical, and perhaps other emerging technologies.