Overview: OCP Puts Full AI Systems in the Spotlight

The Open Compute Project (OCP) Global Summit, held October 12–15 in the United States, gathered AMD, Meta, Broadcom and NVIDIA, among others. As a leading annual venue for the global data center open ecosystem, OCP showcased a shift from single-chip performance to system-level co-optimization across compute, memory, network, power and software orchestration. The agenda condensed into three themes: optics, factories and power. Scaling AI infrastructure efficiently is becoming a core competitive differentiator and an investment guide.

Optical Interconnect: Copper Hits Limits, NPO and CPO Take Over

Network and interconnect were among the most discussed topics. Trillion-parameter models are pushing data movement inside AI clusters against physical limits. Copper has short reach and limited bandwidth density, and its ceiling arrives earlier as clusters grow. Pluggable optics reach farther but cost more power. The OCP consensus: scale-up interconnect inside the rack is moving from copper to near-package optics (NPO) and ultimately co-packaged optics (CPO). High-speed modules of 400G and above—with 800G and 1.6T now the main specifications—are the key building blocks.

Vendor Milestones and Market Context

The show floor made the transition concrete. Credo announced a 1.6T ZeroFlap active optical cable with 12W per-end power, about 30% lower than comparable active electrical cables, a 50-meter target reach, a 100-million-hour MTBF target and planned 2027 shipments. NVIDIA said its Spectrum-X silicon photonics CPO is in full production: the world’s first 200G/lane CPO Ethernet switch, with 102.4Tb/s per chip, roughly three-quarters fewer lasers and about five times better power efficiency. Broadcom showed its third-generation TH6-Davisson CPO and, with Accton, Edgecore and Celestica, demonstrated a 102.4T CPO system with 64 1.6T optical ports and 200G SerDes, targeting mass production in the first quarter of 2027. LightCounting estimates the 2026 datacom optical module market at $22.8 billion, with a 20% CAGR to 2030; the telecom market is $5.3 billion with a 7% CAGR. China holds seven of the top ten optical module companies and about 70% of the global market, according to Goldman Sachs. By 2025 sales, one domestic leader ranked first and another overtook Coherent for second. Their 800G products have begun scaled deployment in AI training cluster interconnect, while 1.6T has entered top-customer validation and early small-batch delivery. The weak link is high-end optical chips above 100G and high-speed DSP chips, where domestic localization remains below 5%.

The Standards Race Moves to the Rack and the AI Factory

A standards race is emerging at the rack level. A traditional intelligent computing center rack uses 4U servers with eight cards each; even a 30–50kW high-power rack holds only three to four servers, or 24–32 cards. NVIDIA contributed its Vera Rubin rack and compute tray design to the open ecosystem through the MGX 1.2 specification, which supports racks up to 250kW and introduces a 5000A liquid-cooled busbar and blind-mate liquid cooling connections. With AMD and Google, it extended standardization to power, cooling and physical layout; with Microsoft, it introduced a grid resilience specification requiring near-full-load recovery within two seconds after a grid disturbance. Microsoft showed its Maia300 rack with 96 chips and 167kW, claiming 40% higher token output per megawatt than GPU solutions and positioning it as its main inference compute in 2027. Meta released its first liquid-cooled network rack, IcePack, and the 102.4T IceCube switching platform for gigawatt-scale AI clusters.

China Standards and Supply Chain

China’s ecosystem is also moving. In September, the Open Data Center Committee (ODCC) released AI Factory Technical Specification V1. A leading handset vendor unveiled the Ascend 960 super node, described as the world’s first commercially deployed compute cluster using NPO. At the September Yunqi conference, a cloud vendor disclosed the Panjiu AL144 ScaleUp super node, with up to 144 cards in a single rack ScaleUp domain, more than 650kW of power, and megawatt-class racks plus 800VDC as the next direction.

800VDC Power Architecture and China HVDC Momentum

Power was a central theme. Delta and NVIDIA jointly described the migration to an 800VDC architecture as rack power density rises from about 145kW today toward 300–500kW and beyond. The architecture aims to improve conversion efficiency, reduce losses and accelerate gigawatt-scale AI factory deployment. NVIDIA’s roadmap rises from about 145kW in the GB200 and GB300 era to 330kW with Vera Rubin NVL72, then 570kW in the next generation, with a native 800VDC rack planned for 2028. The change touches solid-state transformers, high-voltage DC power supplies, distribution and backup, liquid cooling and grid resilience. China has a base: domestic data centers established 240V/336V HVDC systems years ago, a Panama power supply from a leading internet platform has scaled, and a vendor introduced an 800V HVDC solution in early 2026. In September, seven departments including the Cyberspace Administration of China issued a 2026–2030 plan calling for exploration of 800V HVDC and promotion of racks above 100kW. Domestic vendors have more than a decade of HVDC experience, while solid-state transformers and medium-voltage rectification remain at prototype and demonstration stages globally.

Market Signals and Outlook

Optical interconnect and compute hardware have corrected since July, but the correction appears to reflect prior gains and crowded positioning rather than a broken trend. Optical component orders extend to early 2029, AI factory standards are landing, and the 800VDC roadmap is explicit. This is an infrastructure buildout cycle, not a single-product trade. Watch CPO and NPO commercialization, 1.6T ramp, rack power standards, the 800VDC ecosystem, China’s high-end optical chip and DSP localization, and solid-state transformer demonstrations.