Nvidia's $6.5 Billion Bet on Light: How Computex 2026 Marked the End of Copper in AI Data Centers

Taipei’s Computex 2026 became the stage for a definitive industry pivot. Nvidia didn’t just launch products—it laid out the financial and ecosystem blueprint for replacing copper electrons with photons. The logic is simple physics: as AI clusters scale toward a million GPUs, copper interconnects fail under prohibitive power and bandwidth limits . The solution is light.

Lightmatter Joins the NVLink Fusion Ecosystem

On June 3, 2026, Lightmatter officially joined the Nvidia NVLink Fusion ecosystem, integrating its Passage photonic interconnects and Guide lasers into the platform . The partnership aims to deliver Co-Packaged Optics (CPO) and Near-Packaged Optics (NPO) that align with Nvidia’s optical and SerDes specifications. Crucially, it allows custom third-party XPUs to connect directly to Nvidia switch silicon .

Lightmatter claims its bidirectional architecture can reduce fiber and connector requirements by up to 50% compared to conventional approaches . The company’s blog framed the moment bluntly: “Bandwidth is decoupled from the chip shoreline limitations for the first time. Reach extends from a meter to a kilometer without requiring any architectural change for the host XPU” .

The $6.5+ Billion Supply Chain Sprint

Since early March 2026, Nvidia has deployed at least $6.5 billion across five strategic photonics and optical connectivity companies, according to CNBC reporting . The spending breaks down as follows:

• $2 billion each in Lumentum, Coherent, and Marvell (announced in March), securing non-exclusive purchase agreements and R&D commitments for advanced lasers and silicon photonics .
• Up to $3.2 billion in Corning (announced May 6). Nvidia paid $500 million upfront for warrants—with an option for up to $2.7 billion more in equity over three years—to build three new U.S. optical fiber factories dedicated to AI data centers .
• Participation in Ayar Labs’ $500M Series E funding round, signaling a deeper push into chip-to-chip optical interconnects .

Production Switches and the Copper Wall

Beyond investments, Nvidia announced its Spectrum-X Ethernet Photonics switches, built entirely on CPO, are now in full production. The new generation targets scale-out and scale-across AI fabric deployments for the Vera Rubin platform .

At the Marvell Computex keynote, CEO Matt Murphy crystallized the urgency: for trillion-parameter AI workloads, connectivity—not compute density—is the dominant performance bottleneck . He described a “copper wall” where speed-of-light latency in electrical traces becomes the hard limit.

What Comes Next

TrendForce forecasts CPO penetration in AI data center optical modules could reach 35% by 2030, eventually displacing copper for all rack-scale interconnects. Copper will likely hold its ground in ultra-short-reach links until at least 2028 .

The strategy spans the entire stack: CPO at the switch level, silicon photonics for chip-to-chip links, and optical cabling throughout the facility . As Nvidia’s senior leadership has stated publicly, the goal isn’t to abandon copper immediately—it’s to build enough optical capacity to break through the limits electricity imposes on the next generation of AI .

At Computex 2026, the industry got the signal: the future of AI infrastructure runs on light.