The AI boom has hit a networking bottleneck that's forcing a fundamental shift in how chips talk to each other. Traditional electrical connections can't keep up with AI's explosive data demands, pushing billions in venture capital toward photonics startups that use light instead of electricity to connect processors. The race is on to replace decades-old networking technology before AI scaling hits a wall.
The new Silicon Valley runs on a different kind of networking - and it's not the LinkedIn variety. As Nvidia and competitors funnel billions into AI data centers, a quiet revolution is happening in the wires and connections that make modern computing possible.
The problem is speed. AI workloads are growing so fast they're breaking traditional networking technology. "The amount of computing power that AI requires now doubles every three months," Lightmatter CEO Nick Harris recently pointed out. That's way faster than Moore's Law ever predicted.
This bottleneck is creating massive opportunities for startups betting on photonics - using light instead of electricity to move data between chips. The results speak for themselves: Lightmatter just hit a $4.4 billion valuation after raising over $500 million from investors like GV and T. Rowe Price. PsiQuantum topped that with a $7 billion valuation and $1 billion funding round from BlackRock and Nvidia's venture arm.
"Optical technology was considered lame, expensive, and marginally useful for 25 years until the AI boom reignited interest," PsiQuantum cofounder Pete Shadbolt told a recent WIRED panel. Now it's having its coming-of-age moment.
The shift makes sense when you look at the numbers. Traditional networking was fine when it was "switching packets of bits," as Creative Strategies CEO Ben Bajarin puts it. "Now, because of AI, it's having to move fairly robust workloads, and that's why you're seeing innovation around speed."
Nvidia saw this coming. The company's $7 billion acquisition of Israeli networking firm Mellanox Technologies in 2020 now looks prescient. They followed up by buying Cumulus Networks to power their Linux-based networking systems. These weren't random purchases - Nvidia bet that GPUs would only reach their potential when clustered together in massive data centers.
But Nvidia isn't the only one making moves. Broadcom, now worth $1.7 trillion, has become the go-to partner for Google, Meta, and OpenAI on custom data center chips. Last month, Reuters reported they're readying a new networking chip called Thor Ultra, designed as the "critical link between an AI system and the rest of the data center."
The acquisition spree continues. ARM just announced plans to buy networking company DreamBig for $265 million. DreamBig makes AI chiplets - small modular circuits that get packaged into larger systems - with Samsung. ARM CEO Rene Haas called their IP "very key for scale-up and scale-out networking."
Startups are racing to capitalize on this moment. Celestial AI raised $250 million earlier this year from Fidelity, BlackRock, and Tiger Global. Intel's CEO even joined their board. The company focuses on optical interconnect technology that promises to solve the speed problem plaguing current systems.
Lightmatter's approach is particularly ambitious. They build silicon photonics that link chips using light-based connections instead of traditional electrical wires. Harris claims they've created the "world's fastest photonic engine for AI chips" - essentially a 3D stack of silicon connected entirely by light.
"The future of computing is really about light," Harris explains. "You're obviously going to have electronics, and software is critical too, but at this level of computing you need new ideas."
But photonics faces real challenges. The technology is expensive to build and requires highly specialized equipment. It also has to integrate with existing electrical systems - no small technical feat.
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 hyperscalers and cater to their specific needs in both AI datacenter chips and networking," he notes. "These companies know how to scale."
The industry is moving toward much more customization, which might favor larger players over startups. "Networking is the thing that makes computers function, but it just feels like the industry is moving towards much more customization, which might be harder for the small guys," Bajarin says.
Still, the upstarts have valuable intellectual property. The demand for faster data speeds isn't going away - it's accelerating. Every new AI model requires more computational power, and that power needs to flow between more chips faster than ever before.
The question isn't whether photonics will play a role in the future of computing. The question is which companies will capture the value as this transition happens, and whether startups can execute fast enough to compete with tech giants who have deeper pockets and existing customer relationships.
The AI boom has exposed a fundamental limit in how computers talk to each other, and the race to solve it is reshaping the entire semiconductor industry. While photonics startups are attracting massive investment and promising revolutionary improvements, established chip giants aren't standing still. The companies that win this networking arms race will control the infrastructure that powers the next generation of AI - making this one of the most important technology battles happening right now. For investors and technologists, the message is clear: in the age of AI, the companies connecting the chips might be just as valuable as the chips themselves.
