The artificial intelligence (AI) playbook is familiar by now: Build a bigger GPU cluster. Add more Blackwell chips. Throw more electricity at the problem. If the chips get hot, build the data center next to a river. If the bandwidth runs out, lay more copper.
That is how Amazon, Alphabet, Microsoft, and Meta Platforms are solving AI in 2026. And it works — until it runs into physics.
One company looked at that same problem and arrived at a different answer. GlobalFoundries (GFS 0.82%) is betting that the real bottleneck in AI infrastructure is not compute power. It is the wire connecting the chips and replacing that wire with light.
Today’s Change
(-0.82%) $-0.66
Current Price
$79.97
Key Data Points
Market Cap
$44B
Day’s Range
$79.27 – $82.79
52wk Range
$31.51 – $92.55
Volume
5.9M
Avg Vol
4.9M
Gross Margin
26.11%
The copper wall nobody talks about
Inside every AI data center, thousands of chips must share information at enormous speeds. Right now, most of that communication travels through copper, which is running out of room. It generates heat, loses signal over distance, and consumes power in ways that become painful at scale. Every time an AI model gets bigger, the copper problem gets worse.
The industry has known this for years. The solution has a name: co-packaged optics (CPO). The idea is to move optical transceivers, components that transmit data through light rather than electricity, directly alongside the chip, shrinking the distance data has to travel through copper to almost nothing. The result is faster, cooler, more power-efficient AI infrastructure.
In May 2026, GlobalFoundries announced SCALE — Silicon photonics Co-packaged Advanced Light Engine solution — the industry’s first platform to meet the Optical Compute Interconnect Multi-Source Agreement specifications for AI scale-up architectures. The platform uses both coarse and dense wavelength-division multiplexing (DWDM) over each optical fiber to push bandwidth density and scalability past what copper can do, and GlobalFoundries has already demonstrated 8ฮป and 16ฮป bi-directional DWDM natively on its platform — a milestone the company describes as fundamental to everything that follows.
Image source: Getty Images.
The part of the stack everyone is chasing
Here’s the thing about silicon photonics that gets lost in the GPU coverage: It is a manufacturing problem as much as a physics problem. Designing a silicon photonic chip is hard. Building it at scale, with the precision required for optical fiber alignment, in volume, for hyperscale data centers is harder.
GlobalFoundries has spent years developing the process technology to do exactly that. Its silicon photonics platform supports 50 Gbps and 100 Gbps micro-ring modulators, broadband detachable fiber interfaces, and flat insertion loss characteristics that future-proof the platform as wavelength counts scale. In November 2025, the company acquired Advanced Micro Foundry in Singapore, a specialized silicon photonics manufacturer, adding manufacturing assets, intellectual property, and engineering depth that would take years to build from scratch.
That acquisition gave GlobalFoundries production capacity for silicon photonics in Singapore, a geography that matters for supply chain diversification amid elevated U.S.-China semiconductor tensions. The company is building a platform that the hyperscalers need and very few manufacturers can actually deliver.
GlobalFoundries dropped 10% in one day, but I’m not worried
GlobalFoundries fell nearly 10% on May 27, dragged down by Mubadala’s alleged stock sale. Mubadala is Abu Dhabi’s sovereign wealth fund and was formerly the controlling shareholder of GlobalFoundries. Despite the stock drop, the long-term Motley Fool framing wins here.
GlobalFoundries’ story is intact. The SCALE announcement sent the stock up 12% in a single session just weeks ago. A Q1 2026 earnings beat followed. Silicon photonics revenue is expected to nearly double again in 2026, with over 500 design wins logged in 2025 and momentum building. A sell-off does not change those facts.
Also, the U.S. government is backing GlobalFoundries with a proposed $375 million award to help build out domestic quantum manufacturing infrastructure.
Every major hyperscaler is asking how to train bigger models faster. GlobalFoundries is asking a different question: How do you move data between chips without the infrastructure melting?
Co-packaged optics is the answer. The company building the manufacturing platform to deliver it at scale is still, on most days, filed under “semiconductor foundry.” On days it drops 10% for reasons unrelated to its most important business, that drop becomes an opportunity.
