MRVL

MRVL — Marvell Technology

TL;DR

Marvell’s two-act structure: a DSP business under structural threat from LPO/CPO/NPO adoption, and a custom ASIC + photonic interposer business that could more than replace it. The bear case — Irrational Analysis: “everyone hates optical DSP… going from Marvell 80%, Broadcom 20% to…” — is visible in the OFC 2026 trend. The bull case: Celestial AI acquisition (Dec 2025) targets Amazon Trainium4 scale-up CPO with a $1B revenue run rate by end of CY2028, and Google is in early-stage discussions to use Marvell as a TPU design partner. Stock +20% on March 2026 earnings (short squeeze vs. broad semis -10%).¹²³

Business

Designs networking silicon (DSPs for optical transceivers, PCIe switches, DPUs) and custom ASICs for hyperscalers. Three primary revenue streams: (1) Optical DSPs — historically ~80% market share in datacenter transceiver DSPs; under pressure from LPO/CPO/NPO alternatives that eliminate or integrate DSP; (2) Custom ASIC — designed AWS Trainium2; lost Trainium3 socket to Alchip (execution failure); in early discussions with Google as a TPU design partner for high-speed interconnect work; (3) Celestial AI — acquired Dec 2025; EAM-based photonic interposer CPO targeting scale-up links; first commercial milestone is $500M revenue run rate by Jan 2028.⁴¹²

Thesis

• Celestial AI — photonic interposer for scale-up CPO. Celestial AI’s Photonic Fabric (PF) Chiplet is a TSMC 5nm chip incorporating die-to-die interfaces (UCIe, MAX PHY) that delivers 16 Tbit/s scale-up bandwidth per first-gen device (64 Tbit/s in gen-2). Uses EAM modulators (not MRMs), enabling better thermal stability above 50°C — key for placement beneath high-power XPUs. Primary target is Amazon Trainium4 scale-up CPO. Marvell projects $500MrevenuerunratebyF1/28(Jan2028)and$1B run rate by end of CY2028; $2.25Bearn-outcontingenton$2B cumulative revenue by Jan 2029.⁴
• Google TPU design partnership emerging. FundaAI (Apr 2026): Google is actively negotiating TPU development projects with Marvell, who would play a design-service role (similar to MediaTek on other hyperscaler programs), leveraging Marvell’s strength in high-speed interconnects. Google is also exploring a dedicated inference accelerator (LPU-analog) with Marvell involvement. Projects are in early-stage specification definition — not yet contracted.²
• Coherent-Lite as a DSP bridge. At ISSCC 2026, Marvell presented an 800G Coherent-Lite transceiver targeting datacenter campus links (up to tens of km). O-band, dual polarization QAM, 3.72 pJ/bit (excluding SiPho) — half the power of full coherent. Latency <300ns over 40km. Positioned as a middle ground between direct detect and full coherent — relevant for hyperscale campus topologies that don’t need 80km reach.⁵

Risks

• DSP market share collapse. Irrational Analysis OFC 2026: “everyone fucking hates optical DSP.” LPO, CPO, NPO, and XPO (Arista’s new variant) all eliminate or marginalize external DSP. Market share that was ~80% Marvell / ~20% Broadcom is structurally degrading. Credo’s DSP performing well makes it “hyper-bearish Marvell” on the DSP side — the only bear-mitigation argument is Celestial.¹
• Custom ASIC execution credibility gap. Lost Trainium3 to Alchip after failing on Trainium2 timelines and RDL interposer design. SemiAnalysis: “Marvell ends up being the big loser from this.” Alchip had to step in to rescue the Trainium2 package design. Google discussions are early-stage — not contracted, and Marvell’s execution track record creates a credibility overhang.³
• Celestial earn-out at risk. The $2.25Bcontingentpayoutrequires$2B in cumulative revenue by Jan 2029. A single customer concentration risk (Trainium4) and an aggressive CPO commercialization timeline (first production CY2027) make the earn-out challenging. Any slip in Trainium4 schedule or scope flows directly to Celestial revenue.⁴
• Coherent-Lite market size unclear. Campus coherent is a niche between direct detect and full DCI coherent. Hyperscaler adoption of this tier over LPO (which also eliminates DSP complexity) is uncertain.

Recent catalysts

• 2026-04-15 — ISSCC 2026: Marvell presented 800G Coherent-Lite transceiver (O-band, dual pol QAM, 3.72 pJ/b, 40km, <300ns) targeting datacenter campus applications.⁵
• 2026-04-14 — FundaAI: Google in early-stage negotiations with Marvell for TPU design services + potential inference accelerator design role.²
• 2026-03-06 — MRVL earnings (March 5, 2026): stock +20% while broad semis/optics -10%. Characterized as a short squeeze vs. significant short interest. Paywalled earnings details.⁶
• 2026-03-27 — OFC 2026: Irrational Analysis flags MRVL DSP business “panicking” — anti-DSP jihad across the industry; Marvell throwing ideas at the wall. Only bull case is Celestial.¹
• 2025-12-04 — SemiAnalysis: Marvell confirmed loser on Trainium3; Alchip won the socket after Marvell’s RDL interposer execution failure on Trainium2.³
• 2025-12-02 — Marvell announced acquisition of Celestial AI (EAM-based photonic interposer, TSMC 5nm).⁴

Second-order reads

• 2026-03-27 — Irrational Analysis, OFC 2026 Recap — Credo optical DSP performing well is “hyper-bearish Marvell” → positive for CRDO at MRVL’s expense; Marvell DSP erosion is negative second-order for FN if their transceiver volumes use Marvell DSP.
• 2025-12-04 — SemiAnalysis, Trainium3 — Alchip wins Trainium3 design → positive for Alchip (private); Marvell losing socket is negative; Trainium3 uses TSMC N3P → positive for TSM advanced packaging.
• 2026-04-14 — FundaAI, AI Infra 2026 — Google may use Marvell for TPU design services → if contracted, negative second-order for AVGO (AVGO currently designs Google TPU; Marvell as alternative vendor reduces AVGO’s sole-source advantage).

Valuation & positioning

Pending. Earnings post (March 5, 2026) is paywalled; no specific revenue/EPS figures available from corpus. The +20% earnings move on March 5 vs. -10% in broad semis/optics is the key market signal — squeeze dynamics rather than fundamental re-rating by Irrational Analysis’s reading.⁶

Sources

Related

AVGO — AVGO designs Google TPU; MRVL in early-stage discussions as alternative design partner CRDO — Credo DSP performing well = bearish MRVL DSP franchise NVDA — Nvidia CPO uses its own internal DSP, not MRVL AMD — Marvell reportedly involved in MI450 interconnect ecosystem COHR LITE — optical component supply chain FN — transceiver contract manufacturer; downstream of MRVL DSP TSEM TSM — foundry; Celestial AI on TSMC N5

Footnotes

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

2. FundaAI — Deep: AI Infra 2026, Shifting from Brain to Body — 2026-04-14 ↩ ↩² ↩³ ↩⁴

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

4. SemiAnalysis — Co-Packaged Optics (CPO) Book Scaling — 2026-01-01 ↩ ↩² ↩³ ↩⁴

5. SemiAnalysis — ISSCC 2026: Nvidia and Broadcom CPO — 2026-04-15 ↩ ↩²

6. 2026 Earnings: Hilarious Short Squeeze — 2026-03-06 ↩ ↩²