At the OCP APAC Summit that wrapped up in Taipei on August 12, Nvidia confirmed a milestone the industry has been waiting years for. Gilad Shainer, the company’s SVP of Networking, said its Spectrum-6 co-packaged optics (CPO) switches are now in mass production and shipping.
The summit itself was notable: TSMC, Applied Materials, and ASE Group shared the keynote stage alongside Microsoft, Nvidia, and Google — a first for a major OCP event, and a sign of just how far up the supply chain the CPO problem now reaches.
The problem CPO is meant to solve comes down to physics. At the 448G PAM4 signaling that today’s AI switches use, copper cable can reliably carry a signal only about 25 centimeters. Compare that to 2 meters back in the 112G era, and the collapse looks steep — but it isn’t a manufacturing shortfall, it’s the fundamental way high-frequency electrical signals attenuate as they travel through a conductor.
The resulting losses inside data centers are substantial. According to data Lightmatter’s Bijan Nowroozi presented at the summit, today’s AI training clusters are only hitting 38-43% Model FLOPS Utilization (MFU) — meaning more than half of the compute a buyer pays for sits idle, starved by network bottlenecks rather than actually crunching numbers.
Translate that into dollars and a $500 million data center is leaving $280-310 million worth of compute capacity idle at any given moment. CPO targets exactly this gap: by packaging the optical engine — which converts electrical signals to light — right next to the switch chip, at millimeter-scale distance, it eliminates a separate signal-conditioning stage and cuts power consumption by up to 5x.
Nvidia’s Spectrum-6 CPO was built on TSMC’s “COUPE” platform, the product of years of joint development between the two companies. It uses SoIC stacking to place the chip that handles electrical signals directly on top of the chip that handles optical signals, physically minimizing the electrical-to-optical conversion path. TSMC’s VP of advanced packaging, Jun He, laid out a roadmap where COUPE’s bandwidth density grows from 0.5 terabits per millimeter in 2026 to 4 terabits per millimeter by 2030 — an eightfold jump in four years.
He also acknowledged, in the same breath, that three challenges haven’t yet become industry-wide standards: wafer-level test protocols, fiber array assembly, and high-speed optical packaging. CoreWeave, Lambda, Meta, Microsoft, and Oracle Cloud have signed on as early adopters — but that adoption happened inside a supply chain tightly co-engineered between Nvidia and TSMC alone.
Nicole Tien of ASE Group, one of the world’s largest semiconductor packaging and test companies, offered a more sober read at the same summit. CPO has become the inflection point for AI infrastructure, she said, but the shared test standards and simulation methods needed for a genuinely multi-vendor market don’t exist yet — and closing that gap will take another 12 to 24 months.
Broadcom is pushing back with its own “Bailly” CPO platform, positioned as an open ecosystem not tied to a single foundry. In parallel, more than 100 companies have lined up behind “XPO,” an extra-density pluggable optics spec. Arista Networks co-founder Andy Bechtolsheim has pegged the point where today’s pluggable optical modules hit their density ceiling at roughly 18 months out, casting XPO as a bridge to get the industry there.
An open standardization effort Lightmatter launched inside OCP back in March has already drawn Celestica, Corning, Dell, Qualcomm, and others. The goal: push MFU from today’s 38-43% up to 55-65%, and bring energy use per bit under 5 picojoules. How quickly that work bears fruit will decide whether Nvidia’s currently closed ecosystem opens up — or stays that way.
Applied Materials VP Subi Kengeri zeroed in on an even more fundamental constraint in his keynote. With rack power density now pushing past 100 kilowatts toward 200-plus, he argued, the real bottleneck isn’t compute or bandwidth — it’s energy efficiency. The CPO race, in other words, has expanded beyond chip design into a problem that materials, packaging, and system design all have to solve together.
Market scale is adding weight to the competition too. According to Dell’Oro Group, the data center infrastructure market grew 28% year-over-year in Q1 2026 to $12 billion — its fifth straight quarter of 20%-plus growth. Now that data has hit a physical wall it can no longer outrun as an electrical signal, how fast the industry can switch to light is starting to set the pace of the entire AI infrastructure race.
Sources
– Tech Times, “Nvidia Ships Co-Packaged Optics Switch as Ecosystem Testing Gap Widens” (Aug 12, 2026)
– OCP APAC Summit 2026, Taipei (Aug 11-12, 2026)
