Samsung Silicon Photonics Foundry Platform at OFC 2026

Samsung's foundry division launched its silicon photonics platform at OFC 2026 last week, targeting 2028 mass production of photonic integrated circuits paired with AI semiconductors and turnkey co-packaged optics (CPO) services by 2029. The pitch is not the photonics itself; several foundries are already further along. Samsung's argument is that it is the only company that can bundle HBM memory, logic fabrication, advanced packaging, and silicon photonics under one roof. Whether that vertical integration story holds up against TSMC's production lead is the real question.

What Samsung announced

At OFC 2026 in Los Angeles, Samsung unveiled a silicon photonics foundry service built on 300mm wafer production with a completed process design kit (PDK). The company demonstrated a 224 Gbps-per-lane modulator validated by imec in Belgium, targeting aggregate link speeds of 800 Gbps (four lanes) and 1.6 Tbps (eight lanes). The roadmap is aggressive:

Year - Milestone

2026 - PIC platform launch, PDK complete

2027 - Optical engines via thermo-compression bonding

2028 - Hybrid copper bonding, mass production with AI chips

2029 - Turnkey CPO services

2030 - Next-generation CPO

Samsung's Singapore R&D center, led by VP King-Jien Chui, a former TSMC executive, is working with Broadcom on commercialization. Broadcom designs both optical engines and networking ASICs, so this is a real validation partner. The open question is whether Samsung can convert collaboration into production tape-outs.

The vertical integration argument

Samsung's real differentiator is not any single component. It's the stack. No other foundry manufactures HBM memory in-house while also offering logic fabrication, 2.5D/3D advanced packaging, and now silicon photonics. TSMC dominates logic and packaging but buys memory from third parties. Intel has a broad technology portfolio but its foundry business is in restructuring mode. GlobalFoundries is building toward production-grade datacom silicon photonics but lacks the memory vertical.

For AI accelerator designers, the promise is tantalizing: a single foundry partner that can co-integrate your compute die, your HBM stacks, your packaging, and your optical I/O. Samsung pitched exactly this in a June 2024 Digitimes report, laying out a vision of 1.4nm process plus backside power delivery network (BSPDN) plus silicon photonics as a unified AI offering.

Does vertical integration at Samsung's current foundry scale create actual technical advantages? Or is it a marketing story to attract customers already drifting toward TSMC?

TSMC is already shipping

TSMC's COUPE (Co-packaged optics Universal Photonic Engine) platform is in mass production for 1.6 Tbps pluggable form factors, with CoWoS-integrated CPO following this year. The roadmap extends to 6.4 Tbps in Gen2 and 12.8 Tbps in Gen3 (still in pathfinding). TSMC builds its photonic ICs on a 65nm-class process using SoIC-X packaging - details Samsung has not matched in specificity.

NVIDIA already has TSMC-built CPO products commercially available, claiming 3.5x power savings over pluggable transceivers (9W per port versus roughly 30W) and a 10x improvement in connector resiliency.

Samsung's 2028 date is for PIC integration with AI chips. Full CPO does not arrive until 2029. That puts Samsung roughly two to three years behind TSMC on production timelines.

What's missing from Samsung's announcement

Samsung did not specify which fabrication node its silicon photonics platform uses. Both TSMC and GlobalFoundries publicly identify their SiPh process nodes. The omission suggests the platform may not yet be locked to a specific process, which raises questions about how far along the manufacturing qualification really is.

Then there is the customer question. No hyperscaler, AI chip designer, or optical module company has publicly committed to Samsung's silicon photonics. Compare that with NVIDIA's shipped CPO products on TSMC, or Meta's 90 million hours of CPO reliability validation, work that took years to complete on competing platforms.

The imec validation, while credible, is partial. A 224 Gbps modulator validated in a research lab is not the same as production yield at 300mm scale or system-level reliability. That gap between component demonstration and volume manufacturing is massive.

And Samsung's foundry division held just 7.3% market share in Q2 2025 versus TSMC's 70.2%, a gap that has widened over the past year. Silicon photonics is partly a customer acquisition play - a way to offer something TSMC cannot yet bundle, and pull AI chip designers into Samsung's ecosystem before they commit elsewhere.

Why this matters for AI infrastructure

The bandwidth wall in AI data centers is real and getting worse. Trillion-parameter-scale training runs demand interconnect bandwidth that copper traces cannot sustain without unacceptable power draw. Co-packaged optics eliminate pluggable transceivers by integrating optical engines directly onto switch ASIC packages.

OpenAI has said each new generation of infrastructure must deliver roughly 2x bandwidth density improvements while reducing I/O power. CPO is the clearest path to meeting that compound scaling requirement.

OFC 2026 removed any remaining doubt about silicon photonics as a production technology. NVIDIA is shipping CPO. TSMC is in volume production. Meta has validated reliability at scale. Samsung joining confirms the industry direction, even if it does not yet change the competitive order.

The bottom line

Samsung has the right thesis. Vertical integration of memory, logic, packaging, and photonics under one roof is a genuinely differentiated offering, if it works at production scale. But thesis and execution are different things. TSMC has two to three years of production lead time, NVIDIA is already shipping CPO systems, and Samsung has not named a single customer for its silicon photonics platform.

The Chui hire from TSMC and the Broadcom partnership suggest Samsung is serious. The 2028-2029 timeline will tell us whether this is a real competitive entry or a marketing exercise. For now, it is the latter until proven otherwise.