AINEWS
Nvidia's Spectrum X Photonics is the first Ethernet switch to ship at volume with optics built into the same chip package, carrying 200 gigabits per second on each data lane.
AI data centers can only pull so much electricity from the grid, and the chips inside them can only move data so fast before the network, not the processor, becomes the limit. At a million GPUs, that limit is no longer theoretical. It is the line that decides whether a new cluster gets built or not.
Nvidia's Spectrum-X Ethernet Photonics platform has entered full mass production, becoming the first co-packaged-optics (CPO) Ethernet switch built for 200G-per-lane volume shipping. Each data lane carries 200 gigabits per second. Instead of plugging in separate optical transceivers at the faceplate, the switch puts the optics onto the same package as the switch ASIC. Compared with pluggable transceivers, the Nvidia silicon photonics design uses 4x fewer lasers, draws 5x less network power, holds 10x better resiliency at scale, and improves signal integrity 64x; optical loss drops from roughly 22 decibels to about 4.
That is the physical fix. The business question is who builds it.
The Nvidia press release names CoreWeave (CRWV), Lambda, and Oracle (ORCL) as the first customers, with production ramping from May through July before reaching full volume. That customer list is short, but each one is a hyperscale AI buildout that lives or dies on network throughput per watt. HPCwire's coverage frames the switch as Nvidia's path to "millions of GPUs" inside a single AI factory, Nvidia's term for a full training-and-inference site.
The supply chain behind that path is narrower than the customer list.
TSMC fabricates Nvidia's silicon photonics and packages the optics alongside the switch ASIC. That step, advanced packaging, is heading toward 20% of TSMC's planned 2026 capital budget of $52 billion to $56 billion, up from roughly 8% of TSMC's revenue in 2025 and on track to top 10% this year. The packaging capex is now large enough to move the company's capex mix, and TSMC trades at a trailing P/E near 30 against a 10-year median closer to 20. The market is paying for the packaging buildout to land on schedule.
Lumentum Holdings (LITE) sits one step further back, supplying the lasers and optics that feed the CPO line at a scale the 247 Wall St. synthesis puts in the multi-hundred-million-dollar range. Nvidia does not publicly break out Lumentum's share of the Spectrum-X bill of materials, so the dollar weight is approximate rather than audited.
The bottleneck this switch targets is real, the customer list is real, and the supply chain is concentrated enough that a packaging miss at TSMC or a laser yield problem at Lumentum would reach the customer before Nvidia would. Nvidia already runs a separate Spectrum-X Ethernet program with Meta and Oracle as deployed customers, and the Photonics SKU operates alongside that existing Ethernet line rather than replacing it. The first quarter that tested the ramp just closed with Nvidia reporting $81.7 billion in revenue, up 85% year-over-year, with $75.2 billion from data center alone. Whether the next quarter's print reflects the new switch line, or the new line's packaging constraints, will be visible in TSMC's packaging yield commentary by the next earnings call.