An editorial review of the most consequential infrastructure transition in a generation
For sixty years, copper was the default medium of computing. It carried signals between chips, between servers, between racks. It was cheap, compatible with everything, and good enough. The entire architecture of the modern data center — from the patch cables snaking through the floor to the PCIe traces etched into motherboards — was designed around the assumption that electrons could do the job.
That assumption has now broken down. Not because someone decided to change it. Because physics did.
The transition from copper to optical interconnects is underway across every major AI data center on earth. It is not a gradual upgrade. It is a structural replacement — the kind that happens once or twice in a generation, reshaping supply chains, rewarding new winners, and quietly making obsolete infrastructure that seemed perfectly adequate just five years ago.
Understanding why it is happening, where it stands today, and where it leads requires understanding what copper can no longer do.
Part One: The Copper Wall
The problem is not that copper is bad. The problem is that AI is different.
For most of computing history, data centers were optimized for latency and throughput on north-south traffic — requests coming in from users and responses going out. The compute inside the data center was relatively isolated. Servers processed jobs largely independently. The cable connecting a server to a switch carried modest loads at modest speeds, and copper handled it without complaint.
AI training changed everything about this picture. Training a large language model is not an isolated workload. It is a massively parallel operation that requires thousands — now tens of thousands — of GPUs to communicate with each other constantly, passing gradients, activations, and model weights back and forth at speeds that would have been inconceivable in a standard enterprise environment five years ago. The traffic is not north-south. It is east-west, all-to-all, relentless.
At the data rates this requires — 224 Gbps per lane and beyond — copper hits three walls simultaneously.
The first is bandwidth. At high signal speeds, copper wires suffer from what engineers call skin effect and dielectric loss: the signal degrades as it travels, attenuating faster the higher the frequency. Beyond 200 Gbps per lane, traditional passive copper cannot reliably span even a single server rack. In Nvidia's latest NVLink 6 architecture, the effective reach of copper wires has been compressed to less than one meter. A cable technology that cannot cross a rack is not a data center interconnect. It is a chip-to-chip bridge.
The second wall is power. The signal conversion circuits — SerDes, for serializer/deserializer — required to push copper signals at AI speeds consume enormous amounts of power. In a 10,000-GPU cluster, SerDes circuits alone account for approximately 50% of interconnect power consumption. As clusters scale toward the million-GPU configurations being discussed for next-generation AI factories, the power overhead of copper-based signal processing becomes not just inefficient but structurally prohibitive. Data centers are already fighting grid constraints in Northern Virginia, Ireland, Singapore, and Tokyo. Adding more copper doesn't just cost more money. In many jurisdictions, it is no longer physically possible to get the power.
The third wall is heat. Unlike light, electrical signals generate heat as they travel through resistance. At AI scale, the thermal load from copper interconnects becomes a secondary cooling problem layered on top of an already severe GPU cooling problem. Optical fiber generates no heat in transmission. Light travels through glass without resistance. The thermal arithmetic is simply different.
These three constraints — bandwidth, power, and heat — are not engineering problems that better copper can solve. They are expressions of physics. The copper wall is not a metaphor. It is a law.
Part Two: The Optical Transition — Where It Stands in 2026
The industry's response has not been a single decision but a layered migration, moving from copper to optical at different speeds depending on distance.
The first distances to go optical were the long ones. Between data centers, across campuses, down the spine of the network — fiber has dominated for decades. This was easy. The cost and complexity of optical components was justified by the distances involved, and the bandwidth advantages were unambiguous.
The transition is now moving inward — into the data center itself, and increasingly into the rack.
At the scale-out layer — connecting racks to each other and servers to top-of-rack switches — optical has become the clear winner for any connection beyond three meters. The pluggable optical transceiver market, which sits at this layer, grew over 60% to exceed $16 billion in 2025, driven by hyperscale deployment of 400G and 800G transceivers. By 2027, the majority of AI data center switch ports are expected to operate at 1.6 Tbps. By 2030, 3.2 Tbps becomes the mainstream standard.
But the truly consequential transition is the one happening at shorter distances — inside the rack, chip-to-chip, at the boundary between electronics and photonics. This is where co-packaged optics enters the picture, and where the real structural change is being made.
Co-packaged optics, or CPO, integrates optical engines directly onto the same package as switching chips and network ASICs, eliminating the pluggable transceiver entirely. Instead of converting an electrical signal to light and back again at each end of a cable — each conversion burning power and adding latency — CPO performs the conversion once, at the chip boundary, and carries the signal as light from there.
The power efficiency gains are not marginal. Nvidia's CPO-based Spectrum-X switches deliver five times better power efficiency than pluggable transceiver alternatives. At the scale of a million-GPU AI factory, five times better power efficiency is not a feature. It is a prerequisite.
CPO is no longer theoretical. Nvidia's Quantum-X InfiniBand CPO switches entered commercial availability in the first half of 2026, and Spectrum-X Ethernet CPO switches are shipping in the second half. Meta, Microsoft, CoreWeave, Lambda, and Oracle Cloud Infrastructure are among the first adopters. Coherent Corp., Nvidia's primary CPO optical component supplier, confirmed that scale-out CPO revenue began in the second half of 2026. Scale-up CPO — the harder problem of bringing optics inside the rack for GPU-to-GPU communication — is targeted for the second half of 2027.
The important nuance here is that CPO does not yet dominate. Penetration sits at roughly 0.5% of total interconnect ports in 2026. Pluggable transceivers remain the overwhelming majority of deployed optical connections, and they will remain sizeable — estimated near $26 billion in market value — even by 2030. The transition is not a cliff. It is a ramp. But the direction is unambiguous, and the investment flows have confirmed the destination.
Part Three: The Supply Chain Scramble
No story about infrastructure transitions is complete without a supply chain story, and this one is striking.
Fiber optic cable lead times have stretched to roughly a year. At least one major manufacturer has sold its entire inventory through 2026. The companies that make the physical glass — primarily Corning, Prysmian, Sumitomo Electric, and Fujikura — find themselves in a position they have not occupied in decades: as strategic assets to the world's most powerful technology companies.
Corning's trajectory is the clearest illustration of what is happening. In January 2026, Meta signed a supply agreement worth up to $6 billion, anchoring a new optical cable plant in Hickory, North Carolina. In May, Nvidia invested $3.2 billion in Corning, funding three new advanced manufacturing plants in North Carolina and Texas, with a commitment to increase US-based optical connectivity manufacturing capacity tenfold. In June, Amazon signed a separate multibillion-dollar, multiyear agreement for optical fiber, cable, and connectivity for its US data centers.
In a single year, three of the world's largest technology companies — collectively spending hundreds of billions on AI infrastructure — decided that securing dedicated fiber supply was as strategically critical as securing GPU supply. Corning's stock has more than doubled in 2026 and is up roughly sixfold from the end of 2023. The company's CEO has outlined a path from approximately $20 billion in annual revenue today to $40 billion by the end of the decade, with optical photonics as the primary growth driver.
Corning estimates that optical content per GPU will increase 1.3x to 1.5x by 2028. As Nvidia scales from 72 GPUs per rack to 576 with Vera Rubin Ultra NVL576, the fiber density per facility grows not linearly but exponentially. An AI factory designed for a million GPUs requires billions of fiber connectors and hundreds of thousands of optical backplanes. The supply chain for this does not exist at scale yet. Building it is the work of the next five years.
Part Four: The Players and the Race
The CPO transition is not a single company's story. It is a competitive race with distinct camps.
Nvidia is pursuing what analysts call a holistic, system-integrated approach — embedding silicon photonics deeply within its networking architecture, tightly coupled to its own switching ASICs and GPU ecosystems. The Quantum-X and Spectrum-X CPO platforms are designed to function as vertically integrated optical fabrics, optimized for Nvidia's NVLink and InfiniBand protocols. The longer-term roadmap points toward the Feynman generation, expected around 2028, where CPO extends into scale-up networking — the dense, sub-meter GPU-to-GPU connections that copper has held as its last stronghold.
Broadcom has taken a different path: modular and standards-based, leveraging its dominant position in Ethernet switching ASICs and building CPO solutions designed to integrate with the broader ecosystem rather than a proprietary stack. Where Nvidia competes on vertical integration, Broadcom competes on interoperability and Ethernet ubiquity.
The optical component supply chain runs through a handful of specialized manufacturers. Coherent Corp. and Lumentum together received $4 billion in Nvidia investment in March 2026, establishing them as the primary suppliers for CPO optical engines. Both are expanding indium phosphide manufacturing capacity — the laser material at the heart of CPO — doubling capacity in 2026 with plans to double again thereafter. The silicon photonics market, which provides the integrated photonic circuits that make CPO possible at scale, is projected to grow from $3.1 billion in 2025 to $10.4 billion by 2030.
At the switch and networking layer, Cisco, Arista, and Nokia are each pursuing their own optical strategies. Arista has developed a 12.8 Tbps liquid-cooled optics module. Cisco's OFC 2026 showcase featured an open transport system integrating multiple fiber rails into a single line card. Nokia has absorbed Infinera and is repositioning as an optical networking systems company for the AI era.
Beyond CPO, a parallel track of alternative interconnect technologies is emerging for niche applications. Silicon photonics covers most distances well, but for the shortest in-rack connections — sub-meter, ultra-high density — companies like Point2 Technology and AttoTude are exploring radio-frequency and THz-over-fiber approaches that offer narrower cable profiles and lower power than even short-reach copper. These are early-stage, but they underscore that the race to replace copper is not a single-technology story. It is an ecosystem competition across the full spectrum of interconnect distances and use cases.
Part Five: The Future — What the Data Center Looks Like in 2030
Extrapolating from current trajectories, the AI data center of 2030 will be a fundamentally different physical environment from what exists today.
At the scale-out layer, optical fiber will be universal. Every rack-to-rack and server-to-switch connection beyond a few meters will carry light, not electrons, at speeds standardizing around 1.6 Tbps per lane, with 3.2 Tbps ports entering production. High-speed Ethernet port shipments are projected to climb from 70 million in 2023 to over 240 million in 2026, mirroring the ramp in 1.6T interfaces.
At the scale-up layer — inside the rack, GPU to GPU — copper will persist for the very shortest distances through the late 2020s, but CPO will take over the medium-distance connections within and between nodes, with the Nvidia Feynman generation (2028) serving as the likely inflection point for broad adoption. TrendForce forecasts CPO penetration exceeding 35% of optical interconnect ports by 2030, up from essentially zero in 2025.
By 2030, 70% of all new data center cabling systems are projected to be optical-first. 90% of AI compute nodes will rely on multi-fiber MTP connectivity. The ratio of fiber connectors to GPU units in a hyperscale AI facility will be measured in thousands per GPU.
The economic implications flow downward through the supply chain. Corning, Coherent, Lumentum, and the handful of companies that manufacture precision glass and indium phosphide lasers will occupy strategic positions in the AI infrastructure stack that they do not occupy today. Silicon photonics foundry capacity — currently concentrated at TSMC and Intel Foundry — will become a constraint as meaningful as GPU capacity is today. The companies that locked in supply agreements in 2025 and 2026 will have structural advantages measured in years.
The power implications are equally significant. A data center running optical interconnects at scale consumes dramatically less power per bit moved than one running copper. As global data center electricity usage approaches 1,050 terawatt-hours by 2026 — nearly the total power consumption of Japan — the pressure to squeeze efficiency from every layer of the stack becomes existential. Optical interconnects are not a nice-to-have efficiency gain. In a world of grid nationalism and power permit denials, they are the difference between a data center that can expand and one that cannot.
Conclusion: The Nervous System Rebuilds Itself
The shift from copper to optical is not, in the end, a story about cables. It is a story about the physical limits of computation at scale, and about what happens when an industry's default assumptions — cheap, compatible, good enough — stop being true all at once.
The data center was built around copper because copper worked. It kept working, generation after generation, as clock speeds increased and bandwidth demands grew, because the engineering community found ways to push the material further. Equalization. Retiming. Active electrical cables. Each technique bought a few more years.
The million-GPU AI factory has exhausted those techniques. The electrons have run out of road.
What replaces them is light — in fiber, in silicon photonic integrated circuits, in co-packaged optical engines embedded directly in the switching ASIC package. The transition is messy, multi-layered, and will not be complete for years. Pluggable transceivers and copper will coexist with CPO and silicon photonics through most of this decade. The supply chain is undersized and under strain. The engineering challenges of manufacturing optical systems at GPU scale are genuinely hard.
But the direction is settled. Physics settled it. The copper wall is real, the optical alternative is ready, and the largest technology companies on earth are now signing multi-billion-dollar, multi-year agreements to lock in the glass they need.
The nervous system of AI is being rebuilt, fiber strand by fiber strand, in factories in North Carolina, Texas, and Japan. The world will run on light.
Research sourced from OFC 2026, Nvidia GTC 2026, IEEE Spectrum, Mordor Intelligence, TrendForce, SemiAnalysis, Corning investor disclosures, and reporting from Network World, CNBC, Light Reading, and EE Times. June 2026.