The rapid development of artificial intelligence has encountered network bottlenecks, forcing a fundamental shift in how chips communicate. Traditional electrical connections cannot meet the explosive growth in data demands from AI, prompting billions of dollars in venture capital to flow into photonics startups that use light instead of electricity to connect processors. Before AI expansion reaches its limits, a race to replace decades-old network technologies has already begun.

The new Silicon Valley operates on a completely different network—and not like LinkedIn. As Nvidia and its competitors invest billions of dollars in AI data centers, a quiet revolution is underway in the wires and connections that underpin modern computing.

The problem is speed

Artificial intelligence workloads are growing so rapidly that traditional networking technologies are struggling to keep up. "The computing power required for AI is now doubling every three months," Lightmatter CEO Nick Harris recently noted. This far exceeds the predicted rate of Moore's Law.

This bottleneck has created a huge opportunity for startups betting on photonics—which uses light, rather than electricity, to transfer data between chips. The results are evident: Lightmatter is valued at $4.4 billion after raising over $500 million from investors like GV and T. Rowe Price. PsiQuantum is valued even higher at $7 billion and has secured $1 billion in funding from BlackRock and Nvidia's venture capital arm.

"For 25 years, before the AI boom reignited interest in optics, the technology was considered outdated, expensive, and with limited uses," said PsiQuantum co-founder Pete Shadbolt at a recent Wired magazine panel. Now, it is entering its golden age.

Swap bit data packets

From a data perspective, this shift makes perfect sense. As Ben Bajarin, CEO of Creative Strategies, put it, traditional networks functioned well when "exchanging bit packets." "But now, because of the advent of artificial intelligence, it has to handle a fairly massive workload, which is why you're seeing innovations in speed." Nvidia anticipated this long ago. Its $7 billion acquisition of Israeli networking company Mellanox Technologies in 2020 now appears remarkably prescient. Subsequently, they acquired Cumulus Networks to power their Linux-based networking systems. These acquisitions are no coincidence—Nvidia is betting that GPUs can only reach their full potential in large data center clusters.

But Nvidia isn't the only company making moves. Broadcom, with a market capitalization of $1.7 trillion, has become the preferred partner for custom data center chips from companies like Google, Meta, and OpenAI. Last month, Reuters reported that they were preparing to launch a new networking chip called Thor Ultra, designed to be "a critical link between AI systems and the rest of the data center."

The acquisition frenzy continues. ARM just announced plans to acquire networking company DreamBig for $265 million. DreamBig is partnering with Samsung to produce AI chips—small, modular circuits that can be packaged into larger systems. ARM CEO Rene Haas called their intellectual property "critical to both scale-up and scale-down networks."

Startups are scrambling to capitalize on this opportunity. Celestial AI raised $250 million earlier this year from Fidelity, BlackRock, and Tiger Global Management. Intel's CEO has even joined the company's board. The company focuses on optical interconnect technology, which promises to solve the speed problems plaguing current systems.

Lightmatter's approach is particularly ambitious. Their silicon photonic devices use optical connections instead of traditional wires to connect chips. Harris claims they've created "the world's fastest AI chip photonic engine"—essentially a three-dimensional silicon stack structure entirely connected by light.

"The future of computing is really about light," Harris explains. "Electronics are certainly essential, and software is crucial, but at this level of computing, you need new ideas."

But photonics faces numerous challenges. The technology is expensive and requires highly specialized equipment. It must also integrate with existing power systems—no easy feat.

Bajarin points out that established companies like Broadcom and Marvell have advantages beyond just technology. "Companies like Broadcom have the expertise and resources to work with hyperscale data center operators and meet their specific needs for AI data center chips and networks," he notes. "These companies know how to scale." The industry is moving towards high customization, which may benefit large enterprises more than startups. "Networks are the foundation of computing, but it feels like the entire industry is moving towards high customization, which may be more difficult for smaller businesses," Bajarin says. Despite this, these emerging companies possess valuable intellectual property. The need for faster data transfer speeds hasn't disappeared; in fact, it's accelerating. Every new AI model requires more powerful computing capabilities, and this capability needs to be transferred between more chips at unprecedented speeds.

The question isn't whether photonics will play a role in the future of computing, but which companies will benefit from this transformation, and whether startups can move fast enough to compete with well-funded tech giants with existing customer relationships.

The rapid development of AI has exposed a fundamental limitation in how computers communicate, and the race to overcome this limitation is reshaping the entire semiconductor industry. While photonics startups have attracted significant investment and promised revolutionary improvements, established chip giants haven't stood still. The company that wins this cyber arms race will control the infrastructure that powers the next generation of AI—making it one of the most important technological battles of our time. For investors and technologists alike, the message is clear: in the age of AI, the companies that connect chips may be just as important as the chips themselves.

Source: Content compiled from techbuzz

View more at EASELINK