mstack

CRDO

5 min read

CRDO — Credo Technology

TL;DR

Credo dominates the Active Electrical Cable (AEC) market for short-reach (3–7 meter) datacenter interconnects and is embedded deep in AWS Trainium’s scale-up fabric — with a stock-warrant deal structure so favorable to AWS that Credo effectively paid Amazon to adopt AECs. The near-term catalyst is 1.6T AEC ramp in 2H 2026: Meta VR200 uses 1.6T AECs for NIC-to-TOR; xAI is expected to use 1.6T AECs across the full fabric (NIC-to-TOR and switch-to-switch), giving Credo pricing power. The structural risk: Irrational Analysis OFC 2026 — “the 3-7m niche dies as LPO, XPO, NPO, and CPO proliferate.”123

Business

Designs AECs (Active Electrical Cables) — copper-based point-to-point interconnects with active signal conditioning for 3–7 meter reach within datacenters. Also makes retimers, PCIe switches, and optical DSPs. AECs serve scale-up (chip-to-chip across racks) and NIC-to-TOR connections in hyperscaler rack deployments. AWS is the largest known customer, with a multi-year strategic partnership anchored by stock warrants. Citrini Research covered CRDO as a connectivity long thesis.24

Thesis

  • AEC dominance in AWS Trainium ecosystem. Trainium2 NL32x2 3D Torus uses ~6,100 cross-rack AECs for its scale-up fabric. Trainium3 NL32x2 Switched uses ~5,100 AECs; Trainium3 NL72x2 Switched (the larger liquid-cooled SKU, 144-chip domain) uses 11,520 AECs for cross-rack NeuronLink connectivity. Credo is the embedded supplier for these links.2
  • AWS warrant deal = structural lock-in. AWS receives CRDO stock warrants that vest as AWS hits AEC purchase milestones. The strike price was low enough that the warrant value at market prices exceeded the total AEC spend — Credo “effectively paid Amazon to take AECs.” This structure (similar to ALAB-AWS but with an even larger effective rebate) creates deep supply-chain stickiness in exchange for equity dilution.2
  • 1.6T AEC ramp in 2H 2026. SemiAnalysis Vera Rubin: Meta VR200 will use 1.6T AECs for NIC-to-TOR connections once 1.6T AECs are broadly available (expected 2H 2026). xAI is expected to adopt 1.6T AECs for both NIC-to-TOR and switch-to-switch connectivity — replacing most transceivers at leaf, spine, and core layers. “This can give Credo plenty of pricing power” on the xAI deployment.1
  • Optical DSP diversification. Credo demonstrated an impressive optical DSP interoperability demo at OFC 2026 — useful for the 800G/1.6T transceiver market and as a DSP alternative to Marvell. Irrational Analysis: “Credo optical DSP being really good is hyper-bearish Marvell, neutral Credo” — it captures DSP share but enters a market the author views as structurally shrinking.3

Risks

  • Structural niche erosion. Irrational Analysis OFC 2026: “Credo operates in a niche of 3–7 meters. This niche dies as advanced optics such as LPO, XPO, NPO, and CPO proliferate.” As LPO (linear pluggable optics) and CPO push optical connectivity closer to the chip, the distance range where AECs are the best solution compresses. Terminal value of AEC is structurally uncertain.3
  • Hyperscaler warrant dilution. The AWS deal structure that creates stickiness also creates equity dilution: Credo issued warrants worth more than the AEC revenue. If replicated with additional customers, the per-share economics deteriorate even as top-line revenue grows.
  • Optical DSP entering a declining market. Even a strong optical DSP product may not generate durable earnings if the DSP-retimed transceiver segment is replaced by LPO, NPO, and CPO — Irrational’s read is “questionable terminal value” for retimed optics.3
  • Customer concentration. AWS/Trainium is the anchor customer. Any design change away from AEC scale-up topology (e.g. optical scale-up) in a future Trainium generation removes the primary demand driver.

Recent catalysts

  • 2026-03-27 — OFC 2026: Credo optical DSP interoperability demo impressive; Irrational Analysis flags the “Optics Reach and Credo Dilemma” — niche viability question raised publicly.3
  • 2026-02-25 — SemiAnalysis Vera Rubin: Meta VR200 + xAI deployments identified as 1.6T AEC customers; ramp expected 2H 2026; xAI single-plane full-fabric AEC design gives Credo pricing power.1
  • 2025-12-04 — SemiAnalysis Trainium3: AEC content model revealed — 5,100–11,520 AECs per rack SKU; AWS warrant structure described; Trainium3 cross-rack scale-up relies on OSFP-XD AECs.2

Second-order reads

  • 2026-03-27 — Irrational Analysis, OFC 2026 Recap — Credo DSP capturing share from Marvell is bad for MRVL DSP franchise; AEC niche dying is a long-run structural negative for CRDO itself.
  • 2025-12-04 — SemiAnalysis, Trainium3 — AWS PCIe retimer deal (ALAB) and AEC deal (CRDO) both embed AWS warrant structures → positive for ALAB on the retimer/switch side; copper scale-up AECs in Trainium4 trajectory matters for both CRDO and ALAB volume outlook.
  • 2026-02-25 — SemiAnalysis, Vera Rubin — xAI single-plane 1.6T AEC fabric means xAI is NOT buying 1.6T pluggable transceivers from LITE COHR AAOI for those NIC-to-TOR links — zero-sum at the port level.

Valuation & positioning

Pending. Citrini Research published a long thesis on CRDO (September 2025) but it was paywalled. Chipstrat Jan 2026 review notes connectivity basket (including CRDO) more than tripled from inception but CRDO underperformed optical names YTD 2026 as market rotated toward laser/photonics.4

Sources

ALAB — PCIe retimer/switch at AWS; same warrant deal structure; complementary to CRDO AECs in Trainium BoM MRVL — optical DSP competitor; CRDO DSP gaining share at MRVL’s expense LITE COHR AAOI — optics competitors at reach ranges where AEC is substituted by LPO/CPO FN — transceiver manufacturer; AEC vs. optics substitution is zero-sum at the port SMTC — AEC/CDR component supplier

Footnotes

  1. SemiAnalysis — Vera Rubin: Extreme Co-Design — 2026-02-25 ↩ ↩2 ↩3

  2. SemiAnalysis — AWS Trainium3 Deep Dive — 2025-12-04 ↩ ↩2 ↩3 ↩4 ↩5

  3. Irrational Analysis — OFC 2026 Irrational Recap — 2026-03-27 ↩ ↩2 ↩3 ↩4 ↩5

  4. Chipstrat — This Month in Review: Jan 26 — 2026-01-30 ↩ ↩2

Graph View

Backlinks