ECOC 2025 Tech Spotlight: Terahop on Optical Solutions for AI Data Centers

Introduction

AI's rapid growth is pushing data centers beyond traditional networking toward a new model: the "AI factory." That takes not only massive GPU/XPU compute, but also multiple network tiers to tie it together — from Scale-Up (GPU interconnect within the system), Scale-Out (expansion across nodes), to Scale-Cross (across data center campuses).

At ECOC 2025, Terahop presented its perspective on optical interconnect for AI data centers, emphasizing that reliability, power efficiency, and an open ecosystem will be the keys to the industry's progress.

Key Points

1. The Multi-Tier Network Demands of AI Data Centers

• Scale-Up: hundreds of GPUs must be tightly interconnected into a single compute unit, extending to >200 GPUs.

• Scale-Out: clusters of tens of thousands of GPUs, sometimes spanning multiple buildings.

• Front-End: the traditional customer-facing data center network.

• Scale-Cross: links between data center campuses over 2 km, 10 km, or even 80 km, requiring ZR-class modules.

👉 Takeaway: an AI data center is not a single network but a combination of multi-tier optical solutions.

2. The 800G → 1.6T → 3.2T Transition

• 800G: shipped at scale over the past two years, mostly based on silicon photonics.

• 1.6T: currently being standardized, with the industry already deploying multiple form factors:

  • AOC (≤10 m)

  • DR4/DR8 (100 m–500 m)

  • FR4 (2 km class)

  • ZR (80 km class)

• 3.2T / 3.5T: expected to emerge around 2027–2028. The industry is gradually converging on a 400G/lane × 8 channels design, though host-side and optics-side packaging approaches are still debated.

3. Linear vs. Fully Retimed Optical Modules

• Fully Retimed (FRO): mature and reliable, but more power-hungry.

• Linear Optics (LRO / RTLR): the new OIF standardized naming, combining a retimed TX with a linear RX. Cuts power significantly.

• LPO: significant challenges, but still under research.

  👉 Industry consensus: multiple implementations will coexist in the near term to address different application scenarios.

4. The Role of Silicon Photonics and New Technologies

• Silicon photonics is now mainstream: volume production of 800G and 1.6T has proven its reliability.

• Advantage: it brings large-scale shipment data, and the accumulated field reliability data becomes a key foundation for future generations.

• Exploring new materials: the 400G/lane era may require hybrid technologies (e.g., DML, hybrid integration), but SiPh remains the primary option.

5. A New Opportunity in Scale-Up: Optics Replacing Copper

• Today: companies such as Google still rely on copper, with as many as 5,000 cables per rack — complex and inefficient.

• The optical opportunity: optics can deliver higher bandwidth and lower latency, and let GPU counts scale beyond 200.

• Key challenges: it must balance high reliability, low power (an acceptable 5–10% of GPU power), and reasonable cost.

6. The Trade-Off Between Pluggables and CPO

• Pluggable modules: easy to maintain and highly open — customers' first choice.

• CPO: the decision must be made at the ASIC design stage, which is high-risk, and its refresh pace can't keep up with a new generation every two years.

• Middle-ground solutions: Near-Packaged Optics (NPO) or Co-Packaged Copper, balancing low loss with an open ecosystem.

7. Why an Open Ecosystem Matters

• The vertical integration problem: closed solutions pushed by major players such as NVIDIA limit industry participation.

• Terahop's view: the industry should push for an Open Socket and standardized connection interfaces, allowing multiple technologies to coexist (DML, SiPh, and slower wide-lane approaches can all plug in).

• Vision: build a shared ecosystem through international bodies such as OIF, so more vendors can take part in AI optical solutions.

Summary

Terahop's ECOC 2025 talk made the following points:

1. AI data center networking is multi-tiered: from AOC to ZR, and from 800G to 3.2T, it requires a comprehensive portfolio of optical solutions.

2. Reliability comes first: the heavy investment in GPUs pushes customers to demand modules with proven field reliability.

3. Silicon photonics has established its footing: accumulated shipment volume and field data have cemented its mainstream position, though future generations will need to blend in new technologies.

4. Pluggables remain the first choice, but transitional solutions are needed: CPO is too closed; NPO and Co-Packaged Copper may become pragmatic alternatives.

5. An open ecosystem is key to the industry's long-term development: only broad participation can sustain the rapid evolution of the AI era.

The talk highlighted that AI demand is pushing optical interconnect into a new era, and open, reliable, and flexible solutions are what can truly support the industry going forward.