Nvidia puts $4 billion into data-center optics

Four billion dollars: that is how much Nvidia is prepared to pay to make light move faster inside data centers. The company has announced investments in two firms developing optical technology for data centers, and the move says far more than just another deal in Jensen Huang's portfolio.

At first glance, data-center optics sounds like a niche topic, far from the loud announcements of new GPUs and record quarterly reports. But optical interconnects, the technology that moves data between thousands of graphics processors at speeds measured in terabits per second, are becoming one of the main bottlenecks in the race to scale AI systems. When a cluster of tens of thousands of chips trains a model on the level of GPT or Gemini, the speed and bandwidth of the links between those chips determine how efficiently the computing power is used.

Copper cables, the standard for decades, are hitting physical limits: beyond a few meters they lose signal and consume too much power. Optical fiber solves both problems, but optical transceiver technology for ultra-dense computing environments is still under active development.

That is why Nvidia is directing $4 billion not into another generation of chips, but into companies that make the connections between them possible. It is a strategically calculated move that reflects the evolution of the company's business model. Nvidia stopped being just a GPU maker long ago. In recent years it has methodically built a vertically integrated ecosystem: its own networking (after acquiring Mellanox in 2020 for $7 billion), AI software platforms, cooling systems, and now optical infrastructure. The logic is simple: the more of the stack Nvidia controls, the harder it is for customers to switch to competitors and the higher the margin on each solution sold.

The context for this investment becomes clearer when you look at the scale of data-center construction worldwide. By various estimates, in 2025-2026 the largest tech companies (Microsoft, Google, Amazon, Meta* and Oracle) will together invest more than $300 billion in AI infrastructure. Each of these giant clusters needs not only tens of thousands of GPUs but a comparable number of high-speed optical connections. According to analysts' forecasts, the market for data-center optical components will grow from $15 billion in 2024 to more than $40 billion by 2028. By investing now, Nvidia is in effect securing privileged access to technologies that will be in short supply within the next few quarters.

There is one more important aspect. Co-packaged optics, which integrates optical modules directly into processor packages, is considered the next big breakthrough in data-center architecture. It would sharply cut latency and power consumption, which is critical for real-time systems and for inference on large models. By investing in optical companies, Nvidia may gain early access to these technologies and integrate them into its future platforms, whether Blackwell Ultra, Rubin or whatever comes after. This is not merely a financial bet but a technological one, and it may shape AI infrastructure architecture for a decade to come.

For Nvidia's competitors, the move adds pressure. AMD and Intel, which are trying to win back share of the AI accelerator market, do not yet show comparable depth of vertical integration. Startups such as Cerebras and Groq focus on chips but do not control the networking layer. Nvidia, by contrast, is building a system in which GPUs, networking, optics and software work as one, and each new link in that chain strengthens the competitive moat around the company.

Four billion dollars in optics is not just an investment in cables and transceivers. It is a bid for control of the physical layer of AI infrastructure. In a world where computing power is becoming the new oil, Nvidia is building not only the drilling rigs but the pipelines. And whoever controls the pipelines ultimately sets the rules of the game.

*Meta is designated an extremist organization and banned in Russia.