⚡ The semiconductor industry expected GPU-to-GPU optical interconnects to arrive around 2033. NVIDIA just yanked that future five years closer, announcing optical fiber will replace copper in its 2028 "Feynman" GPU generation. The pivot is so abrupt that even longtime industry watchers were caught off guard. And quietly, in factories tucked into Japan's countryside, a handful of suppliers are positioning to feed the boom — even as NVIDIA's $4 billion buying spree reshuffles who gets to be in the inner circle.

When "2033" Suddenly Becomes "2028"

Co-Packaged Optics (CPO) — the technology that fuses optical engines directly onto a chip's package — has so far lived mostly inside data center switches. The first wave puts CPO into the ASICs that route traffic between server racks. Useful, but contained.

The real prize is somewhere else: scale-up interconnects between GPUs themselves. A single NVIDIA B200 GPU already burns over 1,000 watts. Pack 72 of them into a "GB200 NVL72" rack and you're at 120 kW per rack. The NVLink fabric currently stitching these GPUs together is copper, and copper hits a wall around 1 meter — bandwidth flatlines and the heat from signal-correction circuits piles up.

Industry forecasts had penciled in 2033 as the realistic year for GPU-to-GPU optical links. Then at GTC 2026 in March, CEO Jensen Huang revealed that NVIDIA's 2028 "Feynman" generation will integrate CPO directly into the NVLink switch, enabling a single "NVL1152" system that links 1,152 GPUs across 16 racks via silicon photonics. That's roughly a five-year acceleration of the entire industry roadmap.

NVIDIA isn't just talking. In early March 2026, the company put $2 billion into Coherent and another $2 billion into Lumentum — about 600 billion yen total — paired with multi-year purchase agreements. There's also a separate stake in Marvell. Huang's framing was blunt: "Everyone in our ecosystem needs a lot more capacity." The optics supply chain is now being secured the way nation-states secure rare earths.

Why Optics Became Inevitable: The Power Crisis

Why the rush? Because AI data centers are running into an electricity wall.

US data centers consumed roughly 4.4% of all national electricity in 2024 — about 176 TWh annually — and projections put that figure near 9% (around 580 TWh) by 2030. NVIDIA's newest "NVL576" rack hits 1.4 MW. Hyperscalers are openly reviewing data center build-outs because of how much power they would draw.

A surprising chunk of that power isn't compute — it's wiring. A high-end switch ASIC consumes over 500W, and roughly 30% goes to the high-speed electrical I/O (SerDes) circuits driving the copper traces. The longer the trace, the worse the signal degrades, and the more power the correction circuitry burns. Copper has a hard physics ceiling, and we've hit it.

CPO sidesteps that ceiling by sitting the optical engine on the same package as the switch ASIC (or GPU), shrinking the copper run from over 100 mm to about 10 mm. NVIDIA claims its CPO designs improve power efficiency up to 5x and resiliency up to 10x versus pluggable optics.

In March 2026, AMD, Broadcom, NVIDIA, Meta, Microsoft, and OpenAI jointly launched the OCI-MSA (Optical Compute Interconnect Multi-Source Agreement), a standards body targeting 3.2 Tb/s and beyond. When six rivals form a standards group together, the message is clear: copper's days as the default scale-up interconnect are over.

Furukawa Electric: World-Leading DFB Lasers and a 38 Billion Yen Bet

Now for the part less covered in the international press: a Japanese supplier holds a global lead in one of the most critical CPO components.

Here's the trick: silicon photonic chips can modulate light, but they can't generate it. They need an external laser source feeding clean, high-power coherent light. The workhorse component is the DFB (Distributed Feedback) laser diode chip — a tiny semiconductor that emits a stable, single-wavelength beam. NVIDIA's Quantum-X800 CPO switch alone uses 18 external laser source modules to drive 144 channels.

Furukawa Electric — known internationally mostly as a cable manufacturer — has been shipping DFB laser chips since 2000 and holds an industry-leading 100 mW output spec. It's a quiet world leader in the category. In December 2025, the company announced it would invest about 38 billion yen (roughly $240 million) to quintuple its DFB laser production capacity by 2028.

The execution is concrete. A subsidiary, Furukawa FITEL Optical Devices, will set up a new factory inside the former Toshiba "Japan Semiconductor (JSC)" Iwate plant in Kitakami, Iwate Prefecture — leasing existing semiconductor manufacturing space rather than building from scratch. Mass production starts April 2028. A second-stage facility in Thailand (completing February 2026) will handle inspection and assembly. Together, these moves push 2028 capacity to 5x the FY2025 baseline.

That's not all. In March 2026, Furukawa Electric began mass-producing 13,824-fiber ultra-high-density optical cables — among the densest in the world. The company has also developed a compact 12-fiber connector specifically for CPO devices, capable of surviving the brutal 260°C reflow soldering step that CPO assembly requires.

The market has noticed. Furukawa Electric's stock has climbed roughly 3x over the past year, hitting around 28,000 yen by late February 2026. A century-old cable maker is being repriced as an "AI infrastructure core stock" — the picks-and-shovels play that's becoming a real shovel.

Sumitomo Electric: Inside Quantum-X, Outside the $4 Billion Round

The other Japanese protagonist is more complicated.

Sumitomo Electric supplies several core components inside NVIDIA's Quantum-X Photonics switch: a 400 mW-class 1.3 μm high-power semiconductor laser, and a "90-degree bent 2D fiber array" technology used to connect optical fibers into the silicon photonic engine. NVIDIA's official ecosystem list of 11 companies — TSMC, Browave, Coherent, Corning, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo Electric, TFC Communication — explicitly names Sumitomo. The technical contribution to Quantum-X is undeniable.

And yet, when NVIDIA wrote a $4 billion check in March 2026, Sumitomo Electric was not on the receiving end. Coherent and Lumentum got $2 billion each plus multi-year purchase guarantees. With two direct competitors now operating with that kind of strategic backing, the playing field has visibly tilted.

This isn't a technology problem. Sumitomo's InP (indium phosphide) lasers remain world-class and continue shipping inside NVIDIA's CPO architecture. The structural question, instead, is this: when a platform company starts selecting partners based on capital, purchase contracts, and the geography of manufacturing, can technical excellence alone keep you in the top tier?

Sumitomo isn't standing still. A new R&D building at its Yokohama plant comes online in July 2026, and the company plans to double its optical device production capacity from 2027 onward. It's also a participant in NTT's IOWN initiative and continues benefiting from the submarine cable boom as Meta, Google, and Microsoft accelerate their own intercontinental cable rollouts. As one of Japan's "fiber three" alongside Furukawa and Fujikura, the foundation is solid.

NTT IOWN: Japan's Own Long-Range Vision

A note on NTT's IOWN (Innovative Optical and Wireless Network) initiative, since this is what most Japanese coverage means by "photonic-electronic convergence."

IOWN aims to rebuild the entire information infrastructure stack — networks, computing, eventually inside chips — using photonics, with an ultimate target of cutting power consumption to 1/100th of today's levels. NTT President Akira Shimada declared 2026 "Year One of Photonic-Electronic Convergence," with commercial IOWN photonic-electronic switches launching in fiscal 2026. The plan calls for 30,000 units a month, built in an ecosystem with Broadcom and Shinko Electric.

Looking further out: PEC-3 (photonic chiplets, ~2028), and IOWN 4.0 (in-package die-to-die optical, ~2032). NTT's edge here is "membrane photonics" — compound-semiconductor structures just 0.3 micrometers thick, with around 100 patents filed.

But it's worth holding two things in mind at once. NVIDIA's CPO and NTT's IOWN are both "optical," but they operate on different timelines and different problems. IOWN is a long-horizon vision aimed at putting light inside the chip itself by the 2030s. NVIDIA's CPO is firefighting equipment for today's AI factory bandwidth bottleneck. The "Japan leads in photonics" framing common in domestic media tells a real story but not the whole one — NVIDIA's capital-driven supply chain integration is moving on its own coordinate system, on its own schedule.

The Japanese Position: Strong Components, Constrained Endgame

Step back and the photonic supply chain layer-cake looks like this:

At the finished-product layer — AI chip design (NVIDIA, Broadcom, Marvell), optical module assembly (Coherent, Lumentum, China's Innolight, AAOI) — Japan's share is genuinely thin.

At the components and materials layer, the picture flips. Japan holds strong-to-leading positions in semiconductor lasers (Sumitomo Electric is world-class), optical fiber (Furukawa, Sumitomo, Fujikura), optical connectors (Furukawa, Sumitomo, Senko), ceramic packages (Kyocera, NGK Ceratec), and InP substrates for optical semiconductors (JX Metals is expanding capacity ~50%).

A Mitsui & Co. Global Strategic Studies Institute analysis identifies Japan's competitive strengths in CPO bottlenecks: fiber alignment, high-temperature-resistant materials, and glass substrates. AGC and DNP are both developing package substrates with embedded optical waveguides. The depth of Japan's materials and component industry could meaningfully shape the CPO supply chain.

The constraint, as the Sumitomo case reveals, is that when a finished-product company restructures the supply chain via capital, technical excellence alone may not be enough to stay in the top tier. When Huang says "everyone in our ecosystem needs more capacity," the definition of "our" now carries equity strings. Whether Japanese suppliers can graduate from vendor to co-development partner — and ideally to NVIDIA-funded partner — is the strategic question for 2027 and beyond.

From "Emerging" to "Center Stage"

Three years ago, "photonic-electronic convergence" was niche jargon known mostly to semiconductor specialists. Today, NVIDIA's CEO talks about it in keynotes, six of the world's largest tech companies have just formed a standards body around it, and hyperscalers are revising data center plans to accommodate it.

The race to build the supply chain ahead of the 2028 Feynman GPU launch will play out across 2026 and 2027. Furukawa Electric's new Iwate factory comes online in April 2028 — almost exactly when Feynman ships.

In Japan, photonic-electronic convergence is often discussed as an NTT-led national initiative. But on the AI infrastructure front lines, it's parts suppliers like Furukawa Electric and Sumitomo Electric who are already in the trenches. How is the AI data center power crisis and the "shift to optics" framed in your country? Are specific domestic suppliers being spotlighted, or is the conversation centered on the finished-product layer? We'd love to hear how the story reads from where you are.

References