Investment thesis
Interactive widgets visualize the load-bearing thesis claims. Drill into each via the linked source pages.
Customer pipeline
Silicon design-win lifecycle — MRVLStage A
Engaged (NDA + spec)
>5
Early conversations / NDA-bounded; spec exchange and architecture review
Stage B
NRE / tape-out
2-3
Engineering work funded; first-silicon tape-outs underway
Stage C
Ramp
4
Production silicon shipping; revenue recognition active. FY26 custom silicon scaled $0 to $1.5B (per Q4 FY26 mgmt) anchored by Trainium 2 + Maia 1 + Arke
Stage D
Steady-state
0
Mature volume; recurring revenue. Mgmt FY28 framework: custom silicon doubles vs FY27 (~$3.6B+) — implies multiple programs in steady-state
Known design wins
| Program | Customer | Stage | Status | Last signal |
|---|---|---|---|---|
| Trainium 2 | AWS / Amazon | C | Fully subscribed; ramping | 2026-03-05 Q4 FY26 earnings call (Murphy): 'Custom business which doubled in fiscal 2026 — scaled from 0 revenue to $1.5 billion in fiscal 2026.' Fully-subscribed framing on Trainium 2 production reaffirmed (companies/mrvl/kb/03_ecosystem/company_profiles/aws_amazon.md). Source: https://www.fool.com/earnings/call-transcripts/2026/03/05/marvell-mrvl-q4-2026-earnings-call-transcript/ |
| Trainium 3 | AWS / Amazon | B | Ramp H1 2026; NRE / early production | 2026-03-05 Q4 FY26 call (Murphy, paraphrased): 'multi-generational program continuity'; defiant 'Do you see me blinking? You do not.' response to Trainium 3/4 share-loss question (companies/mrvl/kb/05_financials/earnings_calls.md L355, L453). Press analyst inference (Trefis, Nextplatform Sep 2025) places Trainium 3 in ramp. |
| Maia 1 (100) | Microsoft Azure | C | Ramping in Azure data centers; OpenAI workload-targeted | 2025-09-03 Microsoft Maia 100 announcement (Nov 2023, Microsoft Ignite); Maia 100 ramp confirmed via Marvell Q1-Q3 FY26 calls under non-naming pattern. Source: https://news.microsoft.com/source/features/ai/in-house-chips-silicon-to-service-to-meet-ai-demand/ + companies/mrvl/kb/03_ecosystem/company_profiles/microsoft_azure.md |
| Maia 2 (200) | Microsoft Azure | B | Slipped from 2025 to 2026 (reportedly H2 2026); design phase | 2026-03-05 Industry reporting confirms Maia 200 production slipped to 2026; Marvell Q4 FY26 call addresses Maia ramp pacing without naming customer (companies/mrvl/kb/03_ecosystem/company_profiles/microsoft_azure.md Recent News). Confidence: secondary (◐). |
| TPU 8i + MPU | Google / Alphabet | A | Design discussions; co-design partner with MediaTek announced at Google Cloud Next 2026 | 2026-04-09 Google Cloud Next 2026 (Apr 9-11): Sundar Pichai unveils 8th-gen TPU (TPU 8t training, TPU 8i inference), MPU, and Marvell + MediaTek partnership for inference-side silicon. Source: https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/ + https://thenextweb.com/news/google-marvell-ai-chips-inference-tpu-broadcom (companies/mrvl/kb/03_ecosystem/company_profiles/google_alphabet.md). Confidence: paraphrased/secondary (◐) — primary verification flagged in KB. |
| MTIA Arke (inference) | Meta Platforms | C | First inference-class Arke chips at pilot deployment 2025; ramping | 2026-04-14 Broadcom press release (Apr 14, 2026) extends Meta MTIA partnership for multi-gigawatts of custom silicon — reinforces Broadcom on Iris (training) while Marvell remains on Arke (inference). Source: https://www.globenewswire.com/news-release/2026/04/14/3273998/19933/en/Broadcom-Announces-Extended-Partnership-with-Meta-to-Deploy-Technology-to-Support-Multi-Gigawatts-of-Meta-s-Custom-Silicon-MTIA.html (companies/mrvl/kb/03_ecosystem/company_profiles/meta_platforms.md). Confidence: paraphrased/secondary (◐). |
| Custom XPU pipeline (unnamed) | 10+ hyperscalers / cloud customers | A | 50+ active design opportunities / 18 multi-generational sockets disclosed | 2026-03-05 Q4 FY26 call (Murphy): 'breadth of our pipeline — 50+ active design opportunities with 10+ customers' and Q2 FY26 disclosure of '18 multi-generational sockets, several represent multibillion-dollar lifetime revenue.' Source: https://www.fool.com/earnings/call-transcripts/2026/03/05/marvell-mrvl-q4-2026-earnings-call-transcript/ + companies/mrvl/kb/05_financials/earnings_calls.md L268, L448. |
Competitor matrix
7 vendors · 6 axes · MRVL focus highlighted| Vendor | Custom AI silicon | Optical DSP / interconnect | DPU / SmartNIC | Switching silicon | PCIe / CXL fabrics | Market cap / scale |
|---|---|---|---|---|---|---|
MarvellMRVLfocus Focus company — anchored by data-center segment ($6.1B FY26, +46% YoY) | leader AWS Trainium 2/3 anchor; Maia, Arke, TPU 8i/MPU expansion | leader Ara 1.6T, Electra DSP, COLORZ pluggables; Polariton 3.2T+ roadmap | fast-follower OCTEON DPU; legacy strength | fast-follower Teralynx 10 at 102.4 Tbps; smaller share than AVGO | present XConn close 2026-02-10; Apollo CXL+PCIe switch IP | $133B Apr 28 2026 close $154.83 (~$133.1B mkt cap) |
BroadcomAVGO MRVL's primary direct competitor; ~7x larger overall; AI-semi gap ~1.4x | leader Google TPU 1-8t (training, 7+ year exclusive); Meta MTIA Iris (Apr 2026 multi-GW extension); reported OpenAI ASIC | fast-follower Trident-pair DSP, Bailly CPO; CPO momentum | limited Not a primary DPU vendor | leader Tomahawk 6 at 102.4 Tbps; ~75%+ Ethernet switching share | limited Selective IP; not a CXL switch leader | $1.5T+ Multi-trillion EV; FY25 AI-semi $8.4B (+106% YoY) |
NVIDIANVDA Coopetition: customer + competitor + investor ($2B reported equity, 2026) | absent Merchant GPU vendor — not custom silicon for hyperscalers (and Marvell is a target of $2B NVDA investment) | present In-house silicon photonics + CPO investments; uses merchant DSPs (incl. MRVL) in Spectrum-X / ConnectX | leader BlueField-3/4 SmartNIC / DPU dominant | leader Spectrum-X Ethernet + Quantum InfiniBand; NVLink Switch fabric | limited NVLink ecosystem (proprietary); NVLink Fusion opens to 3P silicon | $3T+ Dominant AI accelerator share (~70-85% by value) |
AMDAMD AI GPU fast-follower; competes with NVDA more than MRVL directly | limited Merchant accelerator (MI300/325/355); not hyperscaler custom silicon | absent Not a meaningful optical DSP supplier | fast-follower Pensando DPU acquisition (2022) | limited Pensando-derived switching; not a primary switching-silicon vendor | fast-follower CXL early-mover via EPYC + Pensando | $200B+ Merchant AI GPU fast-follower to NVDA |
IntelINTC Limited overlap with MRVL in AI infrastructure | limited Habana Gaudi 3 niche; not winning hyperscaler custom-silicon book | limited Silicon photonics IP (PSM4/CWDM4); commercial impact muted | limited Mount Evans IPU narrow uptake | limited Tofino switch silicon (Barefoot acq.) discontinued | fast-follower CXL standardization leader; Sapphire/Granite Rapids host | $100-150B Foundry pivot; AI silicon footprint limited |
Astera LabsALAB Direct competitor on PCIe/CXL retimer/expander; XConn acquisition partly closes this gap for MRVL | absent Not a custom-AI-silicon vendor | absent Not an optical DSP vendor | absent Not a DPU vendor | limited Scorpio fabric switch (PCIe/CXL focus) | leader Aries / Taurus / Leo retimer + memory expansion; PCIe/CXL pure-play leader | $15-25B Connectivity-IP pure-play |
CredoCRDO Leader in AEC (Active Electrical Cables) + signal-conditioning IP; flanks MRVL on the copper-to-optical edge | absent Not a custom-AI-silicon vendor | fast-follower Optical DSP (Dove); secondary vs MRVL/AVGO | absent Not a DPU vendor | absent Not a switching-silicon vendor | limited Selective signal-conditioning IP | $10-15B AEC + signal-conditioning IP pure-play |
Legend:LeaderFast followerPresentLimitedAbsent
Categorical positioning across six AI-infrastructure segments. Marvell holds leader status on custom AI silicon and optical DSP; competes most directly with Broadcom on custom AI + switching, with Astera on PCIe/CXL, and with Credo on optical DSP / AEC.
Drill-down pages
Bull case →
Five-pillar thesis: foundry-PDK lock-in · reliability gate cleared · pipeline filling · Marvell-Polariton validation · cap-table can wait
Bear case →
Five-pillar thesis: TFLN winning the volume window · "nearly there" since 2014 · materials-licensing unproven at scale · persistent dilution · sentiment priced in
Risk register →
Technical · customer · foundry · capital · IP · competitive · market
Catalyst calendar →
Q2 2026 → 2029 · earnings, foundry tape-outs, customer reveals, IP cliffs
Open questions →
What this KB does not yet resolve · research paths for each