When NVIDIA's Vice President of Networking, Gilad Shainer, casually announced that the company's co-packaged optics (CPO) switches had entered mass production — and were already running inside NVIDIA's own AI factories — the optical networking world felt the ground shift. For years, CPO has been the industry's most hyped "next big thing," perpetually two years away. No longer. In the second half of 2026, NVIDIA's Ethernet-based photonics switches go commercially available, and TrendForce has confirmed the technology has officially crossed into volume production. This is the moment the "swap the cable box" era of data centers starts to die.
The Chinese article Steve shared calls this wind "a bit strong" (这次的风，吹得有点猛). It's not wrong. CPO isn't just a faster switch — it's a fundamental re-architecture of how AI data centers move data, one that trades repairability for a massive leap in power efficiency and density. Here's what's actually happening under the hood.
Today's data centers move data with pluggable optical transceivers — those little removable modules that snap into the front of a switch, converting electrical signals into light that travels down fiber to the next switch. It works, but there's a catch: the electrical signal has to travel all the way from the switch ASIC, across the PCB, through connectors, and into the transceiver before becoming light. That journey is lossy (up to 22 dB per channel at 200 Gb/s), power-hungry (often ~30W per interface), and full of failure points.
Co-packaged optics flips the script. Instead of dragging electrons across a board to a distant module, the optical engine — lasers, modulators, photodetectors — is placed directly on the same package as the switch ASIC. Fiber plugs straight into the engine sitting beside the chip. The electrical path collapses to millimeters, electrical loss drops to roughly ~4 dB, and power per link falls to as low as ~9W.
The frame NVIDIA uses: 3.5x power efficiency, 10x resiliency, and 1.3x faster time-to-operation. In real power terms, a transceiver that burned 20–30 picojoules per bit in traditional designs drops below 5 pJ/bit with CPO. When optics can eat more than 30% of a data center's total power on massive AI clusters, that's not a footnote — that's a line item.
The deep reason CPO matters in 2026 is simple: AI doesn't scale the way old data centers did. Modern AI factories pack tens of thousands of GPUs that must talk to each other at insane bandwidth with minimal latency. Traditional enterprise data centers kept Tier-1 switches inside each rack so copper cables stayed short. But AI topologies relocate switches to the end of the row, stretching distances and making optics mandatory for both NIC-to-switch and switch-to-switch links.
Copper has a hard wall: beyond roughly two meters, signal integrity degrades and power consumption explodes. Light doesn't care. The poster child is NVIDIA's Vera Rubin NVL72 — a single rack packing 72 GPUs and 36 CPUs connected by sixth-generation NVLink, delivering a staggering 260 TB/s of scale-up bandwidth (3.6 TB/s per GPU). That's more bandwidth than the entire internet. NVLink inside a rack, at this density, simply cannot be done with copper cabling. The answer is co-packaged optics that put the light emitter micrometers from the GPU.
One of the article's sharpest insights is about where this bandwidth is headed. There are two ways to grow AI clusters:
NVIDIA's bold claim: the biggest optical opportunity isn't scale-out, it's scale-up. Because trillion-parameter models like the next GPT generation need GPUs to share data with each other constantly, scale-up bandwidth demand is roughly 10x that of scale-out. It's exactly why NVLink and rack-scale supernodes like Vera Rubin NVL72 exist — and why the optics have to live right on the chip. In this world, whoever sits closest to the GPU wins. That's CPO's whole reason to exist.
But mass production doesn't mean it's easy — or cheap to maintain. Here's the tradeoff nobody likes to talk about. Because the optical engine is now fused with the ASIC, you can't just swap a failed transceiver module anymore. A broken link can mean replacing a far more expensive assembly; the Chinese article cites repair times jumping from hours to as long as 72 hours. The "swap a part" model becomes "swap the whole thing."
There are real manufacturing walls too. Silicon photonics wafer yields reportedly hover around 65%, and CPO optical coupling loss can vary by ±2 dB (vs ±0.5 dB for pluggables) — a few points of yield, at scale, become astronomical cost differences. Broadcom and Meta validated their Bailly CPO switch over one million link-hours of reliability testing just to feel comfortable. That's why credible analysts still say full-scale CPO deployment is 3–5 years out, and LightCounting points to 2027 for the real volume ramp.
NVIDIA isn't waiting. It's committed $4 billion into two laser champions — $2 billion each into Coherent and Lumentum — locking up indium phosphide (InP) capacity, the scarcest ingredient in the whole picture. Lumentum's CEO has warned the InP shortfall "will become worse than memory," with capacity booked through 2028. Coherent's InP line is described as the world's only large-scale CPO laser production line, and it's already doubled capacity. On the packaging front, NVIDIA and TSMC jointly developed the Spectrum-X CPO switch using TSMC's COUPE advanced packaging — a scarce resource now competing with AI GPUs for the same 2.5D/3D capacity.
Chinese module vendors aren't sitting out: companies like Tianfu Communication (天孚通信), Zhongji Innolight (中际旭创), and Eoptolink (新易盛) reported 2025 net-profit growth frequently exceeding 50%, even doubling. The article notes the phosphorus indium substrate is still dominated by U.S. supply, and Yunnan Germanium is racing to expand. But the highest-value layers — silicon photonics design, advanced packaging, InP substrates — remain the battleground where NVIDIA is spending billions to dominate.
Ultimately, the question CPO poses is philosophical as much as technical: would you accept a system that's harder to repair, for one that uses a fraction of the power and delivers 10x the density?
NVIDIA's answer is a clear "yes — ship it, and figure out the repairs later." The 2026 products tell the story: the SN6800 Spectrum switch pushes 409.6 Tb/s across 512 ports at 800 Gb/s, and the Quantum-X InfiniBand lines bring 115 Tb/s in a single box. Power efficiency per port is up 3.5x, resiliency up 10x. These aren't theoretical — they're shipping to partners now.
2026 is the year CPO stopped being a promise and became a shipping product. By 2027–2028, when the yield curves flatten and capacity catches up, the "replace the transceiver" data center will start to look very retro. There's real risk — repairability, yield, and a brutal supply-chain war over indium phosphide and packaging. But if NVIDIA is right about scale-up being the future, then light living on the chip isn't just a nice efficiency win. It's the only way to build the next generation of AI at all. And when Jensen bets $4 billion on a piece of physics, the rest of the industry usually ends up following.
For the people running the data centers — grab some popcorn, and maybe pre-order a few spare whole chassis.
