Source: Compiled and adapted from techbuzz

The booming development of artificial intelligence (AI) has hit a network bottleneck, forcing a fundamental shift in the way chips communicate with each other. Traditional electrical connections can no longer meet the explosive growth of AI’s data demands, prompting billions of dollars in venture capital to flow into photonics startups—companies that use light, rather than electricity, to connect processors. Even before AI expansion hits its limit, a race to replace decades-old networking technologies has already begun.

The operation of the new Silicon Valley relies on a vastly different kind of “network”—and it’s not the LinkedIn type. As NVIDIA and its competitors pour billions of dollars into AI data centers, a quiet revolution is taking place in the wiring and connections that underpin modern computing.

The core issue is speed. AI workloads are growing so rapidly that traditional networking technologies are already overwhelmed. “The computing power required for AI is now doubling every three months,” noted Nick Harris, CEO of Lightmatter, recently. This pace far outstrips the growth rate predicted by Moore’s Law.

This bottleneck has created massive opportunities for startups betting on photonics—companies that use light instead of electricity to transmit data between chips. The results are evident: After raising over $500 million from investors including GV and T. Rowe Price, Lightmatter now has a valuation of $4.4 billion. PsiQuantum boasts an even higher valuation of $7 billion and has secured $1 billion in funding from BlackRock and NVIDIA’s venture capital arm.

“For 25 years, this technology was considered outdated, expensive, and of limited use—until the AI boom reignited interest in optics,” Pete Shadbolt, co-founder of PsiQuantum, stated at a recent Wired magazine panel. Today, it is entering its golden age.

From a data perspective, this shift makes perfect sense. As Ben Bajarin, CEO of Creative Strategies, put it, traditional networks work well for “swapping packets of bits.” “But now, with AI, it has to handle significantly larger workloads—and that’s why you’re seeing innovations in speed.”

NVIDIA anticipated this trend long ago. Its $7 billion acquisition of Israeli networking firm Mellanox Technologies in 2020 now appears highly prescient. Later, it went on to acquire Cumulus Networks to power its Linux-based networking systems. These acquisitions were no accident—NVIDIA is betting that GPUs can only reach their full potential when clustered in large-scale data centers.

Yet NVIDIA is not the only player taking action. Broadcom, a company with a $1.7 trillion market capitalization, has become the top partner for custom data center chips among firms like Google, Meta, and OpenAI. Last month, Reuters reported that Broadcom is preparing to launch a new networking chip called Thor Ultra, designed to be the “critical link between AI systems and the rest of the data center.”

The acquisition spree continues. ARM recently announced plans to acquire networking company DreamBig for $265 million. DreamBig collaborates with Samsung to produce AI chips—small modular circuits that can be packaged into larger systems. Rene Haas, CEO of ARM, described their intellectual property as “critical for both vertically and horizontally scaling networks.”

Startups are racing to seize this opportunity. Celestial AI raised $250 million earlier this year from Fidelity, BlackRock, and Tiger Global Management. Even Intel’s CEO joined the company’s board of directors. Celestial AI focuses on optical interconnect technology, which is expected to solve the speed issues plaguing current systems.

Lightmatter’s approach is particularly ambitious. The silicon photonics devices it builds use optical connections, rather than traditional wires, to link chips. Harris claims they have created “the world’s fastest photonic engine for AI chips”—essentially a 3D silicon stack fully connected by light.

“The future of computing is actually about light,” Harris explained. “Electronics are certainly necessary, and software is crucial, but at this level of computing, you need new ideas.”

Photonics, however, faces numerous challenges. The technology is costly and requires highly specialized equipment. It also must integrate with existing electrical systems—a task that is far from easy.

Bajarin points out that established players like Broadcom and Marvell have advantages beyond just technology. “Companies like Broadcom have the expertise and resources to work with hyperscale data center operators to meet their specific needs for AI data center chips and networking,” he noted. “These firms know how to scale.”

The industry is moving toward high customization—a trend that may favor large enterprises over startups. “Networking is the foundation of how computers operate, but it feels like the entire industry is moving toward high customization, which can be harder for smaller companies,” Bajarin said.

Nevertheless, these emerging companies hold valuable intellectual property. The demand for faster data transmission speeds is not only here to stay but also accelerating. Every new AI model requires more powerful computing capabilities—and that power needs to be transmitted between more chips at an unprecedented speed.

The question is not whether photonics will play a role in the future of computing, but which companies will profit from this transition, and whether startups can move fast enough to compete with deep-pocketed tech giants that already have existing customer relationships.

The boom in AI has exposed a fundamental limitation in how computers communicate—and the race to fix this limitation is reshaping the entire semiconductor industry. While photonics startups have attracted significant investment and promised revolutionary improvements, established chip giants are not standing still. The company that wins this networking arms race will control the infrastructure driving the next generation of AI—making it one of the most critical technology battles today. For investors and technologists, the message is clear: In the AI era, the companies that connect chips may be just as important as the chips themselves.

Disclaimer: This article is created by the original author. The content reflects the author’s personal views only. Our reprint is solely to convey an alternative perspective and does not imply our approval or support of the views expressed. For any objections, please contact our backend team.