GlobalFoundries (GF) recently confirmed that its Silicon Photonics Co-packaged Advanced Light Engine (SCALE) platform is commercially mature and that customer tape-outs for Optical Compute Interconnect (OCI) compliant silicon are occurring throughout 2026, with volume production expected to start in 2027.
Significance of OCI MSA
The OCI Multi-Supply Agreement (MSA) was set up at OFC26 to standardize optics as a replacement for copper in scale-up networks. Led by Nvidia, AMD, Broadcom, Meta and OpenAI, the OCI standard defines the entire optical physical layer, with precise specifications about wavelengths, multiplexing technologies, modulation, packaging as well as remote laser specifications. This enables hyperscalers, for example, to couple an NVIDIA GPU to a Google TPU or a Broadcom switch across the exact same optical physical layer:
- Dense Wavelength Division Multiplexing (DWDM) – OCI standardizes DWDM channels specifically for O-band and splits the wavelengths into two groups: Group A (1308-135nm) for one direction and Group B (1328-1335nm) in the opposite direction on the same single fibre. Using Bi-Directional Multiplexing in this way reduces the number of fibre cables in half, which is critical for very dense racks.
- Low Latency Signalling – OCI mandates Non-Return-To-Zero (NRZ) rather than the more complex, power-hungry PAM4 modulation thus keeping latency to a minimum.
Silicon-Level Modulation Hardware – instead of leaving this to the vendor, OCI specifies Micro-Ring Modulators (MRMs) on the silicon photonics die. - Packaging and Pin-Out Details – outlines the physical footprint of the optical chiplet and defines the electrical interface pin-outs, packaging tolerances and how optical chiplets map to D2D standards such as UCIe.
- Remote Laser Specs – OCI enables external laser sources. However, the OCI standard mandates extremely strict performance metrics for any third-party laser due to the incredible sensitivity of micro-ring modulators.
As explained above, OCI is strictly a physical-layer transmission standard. It does not dictate specifications about the logic layer. This means that OCI can accommodate any scale-up standard, including Nvidia’s proprietary NVLink, open UALink or standard PCIe.
GF’s SCALE platform
SCALE is the first custom-engineered platform to meet the OCI Multi-Source Agreement (MSA) specifications for AI scale-up. It offers an advanced portfolio of fully-qualified photonic devices, such as 50Gbps and 100Gbps micro-ring modulators, coupled ring resonators and integrated photodiodes. Additional features include through silicon vias (TSVs) for high-speed signalling and power delivery and copper pad pitches ranging from 110μm down to sub‑45μm for 2.5D/3D stacking from organic substrates to silicon interposers, enabling customers to move quickly from design to volume production.
While GF offers multiple fiber-attach approaches, the SCALE solution leverages broadband detachable fibres. As part of its SCALE roll-out, GF announced a manufacturing collaboration with Corning to use detachable glass-waveguide fibre connectors. This means that GF can perform full wafer-level automated optical testing to ensure an optical chiplet is a “Known-Good-Die” before it is packaged next to an expensive GPU or switch ASIC, drastically increasing manufacturing yields.
Right now, GF has not only met current OCI Gen-1 specifications, it has exceeded them. For example, in addition to demonstrating 4ƛ bi-directional DWDM, customers have also successfully demonstrated 8ƛ and 16ƛ bi-directional DWDM on GF’s production lines. This means that the SCALE platform is ready to move to next-generation bandwidth densities – without needing to upgrade the manufacturing platform. GF also believes that it can extend its silicon-based Micro Ring modulator technology to 200G (currently 100G) thus delaying the need to introduce more complex Thin-Film Lithium Niobate (TFLN) modulation technologies.
Analyst viewpoint
GF’s announcement that it is the first and currently the only foundry with an active OCI platform makes OCI a commercial reality today effectively derisking the supply chain for fabless chip vendors. By standardizing how chips communicate, OCI ensures that the optical supply chain can operate on a multi-vendor basis., i.e. it removes the threat of a single-supplier bottleneck, de-risks billions of dollars in hardware deployment and gives the AI industry a clear, multi-generational roadmap to scale clusters to Gigawatt levels and beyond.
SOURCE is effectively an open, multi-vendor alternative to TSMC’s COUPE that aligns perfectly with what hyperscalers such as Meta or Microsoft want. However, TSMC’s biggest customers (NVIDIA, AMD, Broadcom, etc.) are also founding members of OCI. Counterpoint Research therefore believes that it is only a matter of time before TSMC – and other foundries – will be forced to open up their advanced packaging lines to support OCI compliance in order to avoid losing lucrative AI accelerator contracts. Inevitably, TSMC will continue to market COUPE’s high-performance PAM4 capabilities (much favoured by Nvidia) for cutting-edge, proprietary systems but it is likely that it will also offer OCI-based silicon to meet the demand from hyperscalers that want to build open, multi-vendor clusters. Counterpoint Research expects that TSMC and other foundries, such as Samsung, will start supporting OCI silicon from 2028 onwards.
Cloud AI server analysis service
The above 2-page report is part of Counterpoint Research’s Cloud AI Server Analysis Service, which tracks, forecasts and analyses key components of AI servers and racks as follows:
- ASIC & Custom Silicon Tracker – tracks the shift away from standard merchant GPUs (like Nvidia’s) towards custom cloud chips. Includes shipments of Google’s TPUs, Amazon’s Trainium/Inferentia, Meta’s MTIA and Microsoft’s Maia chips.
- High Bandwidth Memory (HBM) Forecasts – tracks memory density and generational transitions, such as the shift to HBM3E and HBM4.
- Networking Components – provides qualitative and quantitative information, including forecasts, of key networking components such as switches, pluggable transceivers, CPO, NPO LPO, etc. and other data center networking components.
- Hyperscaler & ODM Dynamics – market share by CSPs such as AWS, Microsoft, Google, Meta and ODMs such as Foxconn, Quanta and Wistron building direct-to-cloud servers.
- Foundry and Node Tracking – maps out which chips built on advanced nodes (e.g., TSMC’s 3nm/5nm) and tracks packaging constraints like CoWoS (Chip-on-Wafer-on-Substrate).
Counterpoint Research
Cadence, HPE expand partnership to advance AI data centre modernisation
Tata Communications expands India-Singapore subsea cable capacity amid AI boom
You must be logged in to post a comment Login
You must be logged in to post a comment.
Magazine Subscription
We create rich business content, reach targeted business audiences, and provide valuable business information to our readers.
We create rich business
content, reach targeted business
audiences, and provide valuable business information to our readers.
Copyright © 2026 Communications Today maintained by Algocept
Leave a Reply