Optical Circuit Switch ( OCS )

An Optical Circuit Switch (OCS) is a specialized type of network switch, but it operates at a fundamentally different layer and mechanism compared to traditional Ethernet or InfiniBand network switches.

Key Differences: OCS vs. Traditional Network Switches

What Makes an Optical Circuit Switch Unique?

1. Pure Light Switching:Instead of parsing individual data packets, an OCS uses micro-mirrors (often MEMS – Micro-Electro-Mechanical Systems), liquid crystals, or beam-steering technology to physically redirect light beams from input fiber ports to output fiber ports.
2. Role in AI & High-Performance Data Centers:
   • In modern AI clusters (such as Google’s TPU clusters or Nvidia AI supercomputers), training massive AI models requires dynamic remapping of inter-GPU networks.
   • OCS units allow data centers to reconfigure high-bandwidth link topologies on demand without needing power-hungry electronic packet switches at every layer.
3. Complementary, Not Replacement:OCS devices are generally used alongside traditional packet switches. While traditional switches handle packet-level routing and traffic management inside network layers, OCS devices manage large-scale bandwidth trunking and topology management across racks and pods.

While Optical Circuit Switches (OCS) offer massive benefits in power efficiency, latency, and data-rate independence, they cannot completely replace Ethernet (or InfiniBand) packet switches because they operate on entirely different principles to solve different networking problems.

Here is why OCS cannot standalone as a total replacement:

1. Circuit Switching vs. Packet Switching

• Packet Switches (Ethernet): Can inspect data down to individual packets (bytes of data) and route them to different destinations on a nanosecond-by-nanosecond basis. They handle variable traffic, statistical multiplexing, and routing between millions of endpoints dynamically.
• OCS: Operates at the physical layer (Layer 1). It establishes a fixed, continuous light path (a “circuit”) between Port A and Port B. Changing that path requires physically moving micro-mirrors or adjusting optical components, which takes milliseconds—thousands of times slower than packet-level switching.

2. Lack of Buffering and Packet Management

Ethernet switches act as intelligent traffic controllers:

• Queueing & Buffering: When two nodes send data to the same destination simultaneously, an Ethernet switch buffers the packets to prevent data loss.
• Packet Inspection & Error Handling: Ethernet switches handle packet headers, access control lists (ACLs), drop policies, frame checking, and network address translation (NAT).
• Pure Light Transit: OCS devices are completely “dumb” regarding data contents—light enters one port and exits another. An OCS has no memory, cannot buffer data, and cannot inspect packet headers.

3. Granularity and Fan-Out Constraints

• Ethernet: Supports one-to-many, many-to-one, and multicast configurations seamlessly at the packet level.
• OCS: Is strictly point-to-point (one fiber input maps directly to one fiber output). You cannot split packet streams mid-fiber through a simple mirror switch to send individual packets to different GPUs or servers on demand.

4. Connection Setup Latency

• When a network topology change is required, an OCS takes several milliseconds to re-orient its mirrors or steer optical beams.
• If every single packet required setting up an OCS circuit, network throughput would drop to near zero. Packet switches handle routing decisions in nanoseconds.

How They Work Together (The Hybrid Approach)

Instead of replacing Ethernet switches, OCS is used in tandem with them, particularly in AI supercomputers and hyperscale data centers (like Google’s TPU clusters):

+-----------------------------------------------------------+
|                   Top-of-Rack Ethernet                    |
|             (Handles local packet-level routing)          |
+-----------------------------+-----------------------------+
                              |
                     Optical Transceivers
                              |
+-----------------------------v-----------------------------+
|                Optical Circuit Switch (OCS)               |
|     (Dynamically reconfigures rack-to-rack links/topologies)|
+-----------------------------------------------------------+

• Ethernet Switches handle short-distance, fine-grained, high-frequency packet routing within racks and local nodes.
• Optical Circuit Switches sit at the spine/inter-rack layer to dynamically reconfigure massive bandwidth “highways” between clusters, bypassing expensive and power-hungry electronic switches for long-haul trunking.

Indium Phosphide

What is Indium Phosphide (InP)?

Indium Phosphide (InP) is a semiconductor material (like silicon), but specifically optimized for high-speed photonics — i.e., generating and detecting light at telecom wavelengths (1310nm / 1550nm).

Why InP matters

Silicon → great for electrons (compute)
InP → great for photons (light)

If you need to generate laser light or modulate it at very high speeds, silicon struggles — InP is the gold standard.

Purpose of InP for companies like AAOI

It enables the laser itself

For optical transceivers:

• You need a laser source
• That laser must:
  • Be stable
  • Operate at telecom wavelengths
  • Modulate at very high speeds

InP is the material used to build these lasers (DFB, EML, etc.)

Without InP → no high-performance optical transceivers

It enables high-speed modulation (100G → 800G → 1.6T)

As speeds increase:

• Signal quality becomes harder
• Noise, dispersion, and heat increase

InP allows:

• Higher bandwidth per lane
• Better signal integrity at long distances
• Lower error rates

This is why 800G / 1.6T optics still heavily rely on InP-based devices

It enables integration (PICs – Photonic Integrated Circuits)

Instead of:

• Discrete laser + modulator + detector

You can build:

• Integrated photonic chips (PICs) on InP

That means:

• Smaller size
• Lower power
• Better thermal performance

This is critical for:

• Dense data center optics
• Co-packaged optics (future)

It is a key vertical integration lever for AAOI

For Applied Optoelectronics:

• AAOI grows InP wafers (epitaxy) internally (MBE/MOCVD)
• Fabricates:
  • Laser diodes
  • Optical components

This gives:

• Cost control (vs buying lasers from Coherent / Lumentum)
• Supply security
• Custom device design

This is one of AAOI’s key differentiators vs “assembly-only” players.

InP vs Silicon Photonics

Industry trend:

• InP still dominates lasers
• Silicon photonics is growing for integration
• Many systems = hybrid (InP laser + SiPh modulator)

InP substrate suppliers

• AXT Inc
• Sumitomo Electric
• JX Advanced Metals