Invest Alpha

Connectivity

Connectivity Architecture

Inside a server / GPU box

  • NVLink / NVSwitch (fast GPU fabric)
  • PCIe / CXL (CPU ↔ GPU, GPU ↔ NIC, GPU ↔ SSD, memory pooling)
  • Astera/CRDO retimers help extend these links.

Within a rack

Two paths exist:

a) Server-to-switch links (TOR switch)

Usually via:

  • Short DAC (<3m)
  • AECs (Credential DSP-based copper)
  • Optics (AAOI)

These are still Ethernet, but only at the NIC edge — not for internal device communication.

b) GPU-to-GPU within the same tray/backplane

Via NVLink or PCIe.

Between racks

  • Long-reach optics (AAOI) using 400G/800G QSFP-DD/OSFP
  • Network fabric is Ethernet or InfiniBand

Retimers

Retimers are digital signal processing (DSP) chips used to:

  • Clean up high-speed electrical signals
  • Extend reach over copper inside servers (PCIe/CXL)
  • Extend reach within the rack (as part of an AEC cable)
  • Enable reliable connectivity at 100G / 200G per lane SerDes speeds

Retimers DO NOT replace optics and DO NOT replace AECs — they complement them.

Retimers appear in two places:

1️⃣ Inside an AEC Cable (Cable Retimers)

AEC = Active Electrical Cable
Retimers are embedded inside the cable ends.

They:

  • Recover the signal (equalization + CTLE + DFE)
  • Reclock the data
  • Reduce bit errors
  • Make copper work reliably at 400G / 800G / 1.6T

Without retimers, copper fails beyond ~1 meter at 100G/lane speeds.

This is why Credo is dominant:

  • Credo makes the retimer/DSP silicon that OEMs insert inside AECs.
  • Broadcom/Marvell also make DSPs, but Credo has merchant leadership.
  • Cable vendors (Luxshare, Amphenol) use Credo chips inside their AECs.

If you buy a high-speed AEC, it almost certainly contains:

➡ Credo retimers or
➡ Marvell retimers or
➡ Broadcom retimers (captive)

So:
Retimers are the heart of AECs.

2️⃣ Inside Servers for PCIe / CXL Connectivity (System Retimers)

Used inside:

  • GPU servers
  • CPU motherboards
  • NICs
  • AI accelerator cards
  • PCIe/CXL expansion boxes

These retimers (also called “redrivers + clock data recovery”) extend reach for:

PCIe 5.0 (32GT/s)

PCIe 6.0 (64GT/s)

CXL 2.0 / 3.0 memory expansion

This is the segment where Astera Labs (ALAB) is the market leader.

ALAB products:

  • Aries PCIe retimers
  • Leo CXL memory connectivity chips

Astera is not an AEC supplier — they do system-level retimers, not cable-level retimers.

The Three Types of Interconnect Silicon

LayerTechnologyKey VendorsDistanceUsed For
1. Copper Inside RackAEC DSP RetimersCredo, Marvell1–5mAEC cables inside a rack
2. Copper Inside ServerPCIe/CXL RetimersAstera, Broadcomcm–1mCPU–GPU–NIC signaling
3. Optical Between RacksOptical TransceiversAAOI, Innolight, Coherent5–100m+Inter-rack AI networking

Notice how each vendor sits in a different layer:

  • Credo → AEC copper (retimers inside cables)
  • Astera → PCIe/CXL inside server (system retimers)
  • AAOI → Optical between racks
  • Broadcom/Marvell → Both system PHYs + cable DSPs

They do not directly compete, except for Broadcom/Marvell who overlap everyone.

Putting It All Together in an AI Rack

Inside the GPU server:

Retimers (Astera, Broadcom)
→ Extend PCIe/CXL paths CPU ↔ GPU ↔ NIC

Inside the rack:

AECs (Credo DSP retimers inside cables)
→ Connect server NICs to TOR switches

Between racks:

Optical transceivers (AAOI)
→ Connect TOR to spine/superspine switches
→ Scale GPU clusters across racks