HyperLight secures $80 million to expand TFLN photonics for artificial intelligence.

      The next hurdle for AI is not the chips themselves, but the connections between them. As clusters expand to encompass hundreds of thousands of GPUs, the copper connections that transmit data are becoming insufficient. Consequently, the industry is working swiftly to shift that data traffic to optical systems.

      HyperLight, a startup that originated from Harvard, has recently secured $80 million to advance its approach to this challenge. The Massachusetts-based company focuses on thin-film lithium niobate (TFLN), a material valued for its ability to convert electrical signals into optical signals rapidly, while consuming less power and minimizing loss. This capability is essential for densely populated AI networks.

      In contrast to most competitors who base their optics on silicon, HyperLight is wagering that TFLN can outperform these alternatives, especially as data transmission speeds increase. Its "Chiplet" platform aims to cover a range of applications, from short connections within data centers to extended telecommunications links, all within a single manufacturable design. They are already shipping products with 200G per lane, while parts with 400G per lane are currently in sampling.

      The broader competition mirrors the one that prompted Nvidia to invest billions to address AI’s copper limitations.

      The composition of the investor base provides more insight than the amount of funding itself. The round was spearheaded by MediaTek, the chip manufacturer. However, noteworthy participants include contract manufacturers Foxconn and Jabil, foundry group UMC, Singapore’s EDBI, Taiwan’s CDIB-TEN Capital, and the Qatar Investment Authority. This blend of investors reflects a segment of the AI-hardware supply chain, encompassing everything from silicon fabrication to assembly and investment. As Mian Zhang, the CEO, stated, “This financing is about more than capital. It is about ecosystem alignment.” In simpler terms, the companies likely to manufacture and purchase this technology are now financially involved.

      The significance of this development lies in the optics sector, which is becoming increasingly crucial for AI advancements. Copper cannot continue to supply data to ever-expanding GPU clusters without excessive energy consumption, while light can efficiently do so. This is why Nvidia partnered with Marvell in silicon photonics and why startups continue to assert substantial efficiency improvements from photonic networks.

      HyperLight argues that its material and its unified platform can be scaled for mass production where others may fall short. The newly acquired funds will be utilized to expand manufacturing capabilities, validate customer requirements, and strengthen relationships with foundry partners. Nevertheless, it is important to note that the company’s technical assertions are their own, and many of their backers could benefit if TFLN proves successful. Ultimately, the market—not just press releases—will determine whether lithium niobate is adopted as the standard optical solution for AI.