The Glass Factory: Why AI Is Abandoning Copper For The Speed Of Light

By Staff of Financial Sense
Wednesday, February 4, 2026 1:12 PM EST

Source


As the race for artificial intelligence reaches a fever pitch, the physical infrastructure of the datacenter has hit a trifecta of constraints. These include a compute bottleneck, a signal speed wall, and a severe power crisis. While the promise of small modular nuclear reactors offers a potential long-term energy solution, the timeline for their deployment remains far beyond the immediate needs of today’s training clusters. In this high-stakes environment, the industry is searching for an escape hatch that can shatter the physical limitations of legacy systems.

That escape hatch is light: specifically, fiber optics.

To break through “Moore’s Wall,” which is the new barrier where physics, rather than just economics, halts exponential progress, datacenters are pivoting from electrons to photons. No longer can copper wires keep pace with the relentless demands of scale and speed. Instead, the industry is embracing strands of pure silica and using light itself to move information at unprecedented speeds.

Gavin Baker, a prominent investor, perfectly captured this shift on X, stating: “During the internet, ‘Switch when you can, route when you must’ was an important principle. In the datacenter, ‘copper when you can, optics when you must’ is similar, and the ‘must’ is inexorably approaching for almost the entire datacenter.” This means that while copper was once the efficient and cheap default, the sheer scale of AI models has made the move to optics an unavoidable necessity. The shift toward this optical future is being cemented by massive capital commitments and strategic technological shifts from the industry’s largest players.

NVIDIA recently signaled a new era for networking by announcing its co-packaged optics (CPO) integration within the Spectrum-6 architecture and its next-generation "Rubin" platform. These innovations aim to unify high-speed links directly with the GPU. This technical pivot is mirrored by unprecedented procurement deals: Meta recently signed a landmark $6 billion agreement with Corning to secure a supply of high-density fiber through 2030, ensuring they have the glass needed to build their AI factories. Meanwhile, component giants like Lumentum and Coherent have solidified their roles as the arms dealers of this revolution, maintaining deep supply relationships with Google, NVIDIA, Microsoft, and AWS to provide the high-speed lasers and transceivers that act as the engines of modern compute.

To understand the current moment, it is vital to recognize that the "new discovery" is not fiber-optic technology itself, which has existed for decades, but the radical innovation of Co-Packaged Optics (CPO). We are currently transitioning from a world where optics were peripheral components to one where they are integrated into the very heart of the silicon. CPO involves the radical idea of soldering the fiber-optic "engine" directly onto the GPU or switch package. This fundamental change in architecture eliminates the need for the long electrical "flyover" copper cables that previously sat inside the server, legacy cables that cluttered racks and leaked energy. We are now witnessing a structural timeline where the "must" of optics moves closer and closer to the processor.

  • Pre-2022: Optics were used for "highways," or inter-building connections, while copper remained the standard for "local roads," or intra-rack connections.
  • 2022: Google deployed Optical Circuit Switches (OCS). The emergence of 800G physics revealed the "Copper Wall," where electrical signals could no longer travel more than a few meters.
  • 2023: Meta and Microsoft began massive procurement of high-density fiber for AI "scale-up" fabrics to link vast clusters of GPUs.
  • 2025–2026: The adoption of Co-Packaged Optics (CPO) begins in earnest, moving light from the hallways of the rack directly onto the doorstep of the chip.

corning fiber optic

Source


This transition is driven by the fact that copper is not only becoming prohibitively expensive but is also physically hitting a wall. Think of copper wires as local garden hoses: they work great for a few feet, but the longer the hose, the less pressure comes out the other end. In datacenters, once data rates hit 200G or 400G per lane, the electrical signal in copper degrades so rapidly that its reliable reach is limited to just two meters. Fiber optics, by contrast, are like a high-speed train traveling on a frictionless track, allowing data to move kilometers with virtually no loss.

Consequently, companies are actively switching to high-density fiber for both intra-datacenter connections (linking racks across the facility) and "inside-the-box" connections using flex-connect types and CPO. This allows clusters to grow far beyond the two-rack limit that copper imposes, enabling the massive parallel processing required for the next generation of LLMs.

copper vs fiber optic

Source


The ripple effects of this optical shift will fundamentally reconfigure datacenter economics. Specifically, it will improve training times and enable cost-effective inference. In a massive AI cluster, GPUs must constantly "synchronize" their data. If the network is slow, GPUs sit idle and waste expensive compute time. Optics eliminate this "communication tax" and ensure training is finished weeks or months faster.

Furthermore, by adopting CPO, companies can eliminate power-hungry Digital Signal Processors (DSPs). These chips account for nearly 50% of an 800G transceiver's total power consumption and up to 30% of its Bill of Materials (BoM). By adopting CPO, operators can disintermediate these power-hungry chips entirely, reducing transceiver power by a staggering 84%. This creates a massive trickle-down effect, where lower power consumption means less heat, reducing the burden on expensive HVAC and liquid-cooling infrastructure. For a company serving an AI model to millions of users, these power savings can make the difference between a profitable service and a bankrupting electricity bill.

Ultimately, we are only at the beginning of the optical decade. The evidence for the longevity of this trend lies in the structural nature of the problem. Until we find a way to circumvent the laws of physics regarding electrical resistance, light is the only medium capable of scaling with AI. With companies like Corning and Lumentum scaling capacity for 1.6T and 3.2T roadmaps, the transition to an optics-first architecture appears to be a permanent migration. As AI models grow in complexity, the "must" of optics will likely continue to drive demand for every laser engine and every strand of glass. In fact, global datacenter demand is projected to require 92 million miles of new fiber optic cable over the next decade. This forecasted demand does not seem to be a temporary bubble, but a structural baseline for the future of global compute. This means that the datacenter of the future is likely built not on wires, but on light.

Sources

More By This Author:

This Week's Market Wrap: Mega-Cap Earnings, Inflation Data, And AI-Driven Spending
This Week In Markets: Geopolitics, More Selective AI Investment, Steady Macro Backdrop
The 'Banana Republic' Portfolio: A Survival Guide For Spiking Commodities

Disclaimer: Content is for informational purposes only and does not constitute financial, investment, legal, or other advice.

The strategies and concepts discussed—such as portfolio ...

more
How did you like this article? Let us know so we can better customize your reading experience.

Comments

Leave a comment to automatically be entered into our contest to win a free Echo Show.