AI training & inference = massive east-west traffic
Traditional data centers were north-south (users ↔ servers).
AI clusters are east-west (GPU ↔ GPU ↔ GPU).
- Training a large model requires constant gradient synchronization
- Thousands of GPUs exchange data every few microseconds
- Network bandwidth, not compute, becomes the bottleneck
👉 800G doubles bandwidth per link, cutting congestion without doubling fibers.
2️⃣ GPUs are scaling faster than networks
Look at the mismatch:
| Generation | GPU compute growth | Network link growth |
|---|---|---|
| Pre-AI | ~2× every ~2 yrs | 100G → 200G |
| AI era | 3–5× per gen | 400G → 800G |
NVIDIA’s latest systems assume:
- Higher bandwidth per GPU
- Lower latency
- Fewer hops
👉 400G fabrics start choking before GPUs are fully utilized.
3️⃣ Fewer cables, ports, and power = real money
Data centers don’t just pay for optics — they pay for everything attached to them.
800G vs 400G lets operators:
- Cut port count in half
- Reduce switch radix
- Lower fiber density
- Save power per transported bit
Example:
- 2 × 400G links → 1 × 800G
- Same throughput, less heat, less space, less failure points
👉 Hyperscalers care about $ / bit / watt, not just raw speed.
4️⃣ AI racks are getting physically bigger
Modern AI racks:
- 8–72 GPUs per rack
- NVLink / copper inside the rack
- Optics between racks
As racks scale:
- Copper stops working beyond very short distances
- Optics take over sooner
- Higher speed per optical lane becomes mandatory
👉 800G is the sweet spot before 1.6T matures.
5️⃣ Switch silicon forced the transition
The optics follow the switches, not the other way around.
- 51.2T switches (Tomahawk 5 class)
- 64 ports × 800G = full utilization
- Using 400G would waste switch capacity
👉 Once switch silicon goes 51.2T+, 800G optics are inevitable.
6️⃣ Cost curves finally crossed
Earlier, 800G was:
- Too hot
- Too expensive
- Too complex
Now:
- DSP efficiency improved
- Optical engines more integrated
- Manufacturing yields rising
Invest Alpha 