[OFC 2026] Part 2 of 5: CPO and the AI Interconnect Challenge

Key Takeaways

• AI racks now exceed 1 MW power per rack.
• Copper interconnects hit bandwidth and power limits.
• Meta demands optical solutions match cost, power, reliability.
• Five OFC papers address lasers, margins, transceivers, reliability, detectors.
• Operational viability remains the biggest hurdle for CPO adoption.

Summary

AI training racks are scaling rapidly, with Meta's upcoming ORW rack doubling node count and pushing single‑rack power beyond one megawatt. Copper backplanes now face insurmountable limits in bandwidth, power density, and routing complexity. Meta used OFC 2026 to set a strict baseline for Co‑Packaged Optics, requiring optical solutions to match electrical options in reliability, performance, cost, and power efficiency. Five papers presented at the conference explore lasers, link margin, transceivers, reliability data, and photodetectors, highlighting both progress and remaining viability gaps.

Pulse Analysis

The relentless growth of AI training workloads is forcing data‑center racks to double in size each year. Meta’s latest ORv3 rack already houses 72 GPU nodes on a copper backplane, while the upcoming ORW version will host 144, and future designs aim beyond 256 nodes, pushing single‑rack power consumption past one megawatt. At these densities, traditional copper interconnects encounter simultaneous bottlenecks: excessive heat, limited bandwidth per pin, and complex routing that inflates latency and cost. Engineers therefore look to optical alternatives that can carry terabits per second without the thermal penalties of copper.

Co‑Packaged Optics (CPO) promises to integrate photonic transceivers directly with ASICs, collapsing the electrical‑to‑optical conversion distance. Meta used OFC 2026 to articulate a strict baseline: any optical technology must be reliable, performant, and cost‑ and power‑efficient relative to its electrical counterpart. This three‑pronged test reflects the economics of hyperscale operators, where even marginal increases in power draw or component price can erode profit margins. Consequently, CPO vendors are compelled to demonstrate not just peak data rates but sustained, low‑error operation under real‑world power budgets.

The five papers highlighted at OFC illuminate the full CPO value chain. Furukawa’s work on external laser sources tackles power‑efficient light generation, while Lightmatter’s quantum‑dot semiconductor optical amplifiers aim to compensate link margin losses and enable bidirectional (BiDi) transceiver modules. Meta and Broadcom present field‑reliability data spanning 36 million operating hours, exposing failure modes that must be mitigated before mass deployment. Finally, Marvell and McGill explore next‑generation photodetectors capable of handling higher optical powers with reduced noise. Collectively, these studies reveal progress yet underscore a critical gap: proving operational viability at scale remains the decisive hurdle for CPO adoption.