# The 800G Optical Transceiver Era: How AI/ML Is Reshaping Data Center Networking

The AI revolution isn't just about GPUs — it's driving an unprecedented upgrade cycle in data center optical networking. Training runs that used to take weeks now need days, and that requires optical infrastructure most data centers weren't designed for. Here's what's changing and how to prepare.

## The Numbers Behind the AI Bandwidth Explosion

A single NVIDIA H100 GPU has 900 GB/s of NVLink bandwidth. A cluster of 10,000 H100s needs 4.5 PB/s of interconnect bandwidth — and that's just GPU-to-GPU. Add storage, east-west traffic between nodes, and north-south to the internet, and the optical requirement multiplies.

Data center optics are being pulled from 100G to 400G, and from 400G to 800G, at a pace never seen before:

| Year | Typical DC Optics Speed | AI Cluster Optics |

| ---- | ----------------------- | ----------------- |

| 2020 | 25G/100G                | 100G              |

| 2022 | 100G                    | 200G/400G         |

| 2024 | 100G/200G               | 400G              |

| 2026 | 200G/400G               | 800G              |

| 2028 | 400G                    | 1.6T              |

## 800G Form Factors: What's Available Today

800G transceivers are shipping now, primarily in two form factors:

**QSFP-DD 800G** — The mainstream choice. Eight lanes of 100G PAM4. Available in:

- SR8 (MMF, 60m/100m) — shortest reach, for same-rack and adjacent rack

- DR8 (SMF, 500m) — broke-out to 2x400G or 8x100G

- 2xFR4 (SMF, 2km) — for campus DCI

**OSFP 800G** — Preferred by some hyperscalers (Meta, some Google deployments). Larger form factor, more thermal headroom. Not backward-compatible with QSFP-DD ports.

## The Real-World Challenge: Power and Cooling

An 800G transceiver consumes 14-16W. A 64-port switch fully populated with 800G optics draws over 1KW just for the transceivers. This isn't theoretical — data center operators are already hitting power limits in existing racks.

Practical implications:

- **Plan power per rack carefully** — 800G optics may require higher power density racks

- **Consider liquid cooling** — some 800G optics already benefit from cold-plate cooling

- **Check switch compatibility** — not all 800G ports support all optics types

## Breakout: The Bridge from 400G to 800G

The most practical migration strategy today is breakout:

- 1x 800G QSFP-DD → 2x 400G QSFP56 ports via breakout cable

- 1x 800G QSFP-DD → 8x 100G QSFP28 ports

This allows you to deploy 800G-capable switches today while connecting to existing 400G and 100G infrastructure. As you upgrade endpoints over time, the same switch ports can transition to native 800G.

## Silicon Photonics: The Next Frontier

Traditional transceivers use discrete lasers, modulators, and photodetectors assembled on a PCB. Silicon photonics integrates all optical functions onto a single silicon chip — dramatically reducing cost, size, and power.

Companies like Intel, Cisco (Acacia), and Ayar Labs are leading silicon photonics development. While still early for mainstream 800G deployment, silicon photonics will be essential for 1.6T and beyond.

## How to Prepare for 800G

1. **Upgrade your fiber plant** — MMF limits you to SR8 at 800G. SMF gives you DR8, FR4, and future options. If you're installing new fiber, go single-mode.

2. **Choose QSFP-DD switch ports** — they support 100G through 800G with the same form factor

3. **Budget for higher optic costs** — 800G transceivers currently cost 3-5x more than 400G equivalents

4. **Plan power and cooling** — a fully-utilized 800G switch draws significantly more power

5. **Consider break-out migration** — 800G to 2x400G today, native 800G tomorrow

## The Bottom Line

800G is here now, driven by AI/ML workloads that cannot wait. The data centers adopting 800G today are building competitive advantage for the next 3-5 years. If you're planning infrastructure upgrades, make sure 800G is in your roadmap.

---

**About the Author:** APEX GROUP LIMITED is a Hong Kong-based optical transceiver specialist, authorized distributor for Hisilicon, Accelink, Acacia, Coherent, and UAISPACE. We supply 25G to 800G transceivers for data center, AI/ML cluster, and telecom applications. Contact us at Optical@apexallinone.com for volume pricing and technical consultation.