The unreal intelligence (AI) playbook is acquainted by now: Construct a much bigger GPU cluster. Add extra Blackwell chips. Throw extra electrical energy on the drawback. If the chips get sizzling, construct the info heart subsequent to a river. If the bandwidth runs out, lay extra copper.
That’s how Amazon, Alphabet, Microsoftand Meta Platforms are fixing AI in 2026. And it really works — till it runs into physics.
Will AI create the world’s first trillionaire? Our workforce simply launched a report on the one little-known firm, known as an “Indispensable Monopoly” offering the vital know-how Nvidia and Intel each want. Continue »
One firm checked out that very same drawback and arrived at a unique reply. GlobalFoundries (NASDAQ: GFS) is betting that the actual bottleneck in AI infrastructure shouldn’t be compute energy. It’s the wire connecting the chips and changing that wire with gentle.
The copper wall no one talks about
Inside each AI knowledge heart, 1000’s of chips should share data at monumental speeds. Proper now, most of that communication travels by way of copper, which is working out of room. It generates warmth, loses sign over distance, and consumes energy in ways in which develop into painful at scale. Each time an AI mannequin will get greater, the copper drawback will get worse.
The trade has recognized this for years. The answer has a reputation: co-packaged optics (CPO). The concept is to maneuver optical transceivers, parts that transmit knowledge by way of gentle reasonably than electrical energy, straight alongside the chip, shrinking the gap knowledge has to journey by way of copper to nearly nothing. The result’s sooner, cooler, extra power-efficient AI infrastructure.
In Might 2026, GlobalFoundries introduced SCALE — Silicon photonics Co-packaged Superior Mild Engine answer — the trade’s first platform to fulfill the Optical Compute Interconnect Multi-Supply Settlement specs for AI scale-up architectures. The platform makes use of each coarse and dense wavelength-division multiplexing (DWDM) over every optical fiber to push bandwidth density and scalability previous what copper can do, and GlobalFoundries has already demonstrated 8λ and 16λ bi-directional DWDM natively on its platform — a milestone the corporate describes as basic to every part that follows.
The a part of the stack everyone seems to be chasing
This is the factor about silicon photonics that will get misplaced within the GPU protection: It’s a manufacturing drawback as a lot as a physics drawback. Designing a silicon photonic chip is difficult. Constructing it at scale, with the precision required for optical fiber alignment, in quantity, for hyperscale knowledge facilities is more durable.
