Skip to content
MRVL
~3 min read · 743 words ·updated 2026-04-29 · confidence 72%

Executive Description

Alphabet Inc. (NASDAQ: GOOG / GOOGL) — through Google Cloud and the TPU program — has been historically a Broadcom-only custom-silicon customer (TPU v1 through v7 designed in partnership with Broadcom over a 7+ year relationship). At Google Cloud Next, April 2026, Alphabet disclosed a strategic diversification: Marvell joins Google’s TPU ecosystem on the inference side, partnered with MediaTek, while Broadcom retains the training-TPU position. The April 2026 announcement also reveals Google’s Memory Processing Unit (MPU) — a memory-centric companion ASIC to the TPU stack on which Marvell has design-win positioning. Marvell’s FY2026 Google revenue is estimated at ~$350-550M (mostly existing optical/DSP supply), but the FY2027-2028 ramp from the MPU and inference-TPU programs is the key upside lever. Cross-link customer wallet share Part 2.3 for the full wallet-share model.

Marvell Relationship

  • Inference TPU partnership (NEW, April 2026): Marvell, with MediaTek, is named as a co-design partner on Google’s TPU 8i (inference-optimized variant). Specific BOM split between MediaTek and Marvell not publicly disclosed. ◐
  • Memory Processing Unit (MPU): Announced at Google Cloud Next 2026 as a new ASIC class designed to relieve memory-bandwidth bottlenecks adjacent to TPU compute. Marvell’s role is on architecture/physical-implementation; specific scope undisclosed. ◐
  • Training TPU (TPU 8t): Remains a Broadcom design — Google has not displaced its long-standing partner on training silicon. Cross-link broadcom. ✓
  • Optical DSP & interconnect: Marvell already supplies optical DSPs and pluggable modules to Google data centers (existing business, $300-500M FY2026 estimate). ◐
  • FY2026 revenue estimate: $350-550M total (optical dominant; TPU work nascent). Cross-link customer wallet share Part 2.3. ◐
  • Forward upside: If MPU + inference-TPU revenue ramps to $2-3B by FY2028, Google wallet share grows from ~0.75% (FY2026) toward 4-6% — meaningful contributor to Marvell’s $15B FY2028 framing.
  • Strategic logic on Google’s side: Diversifying away from a single-vendor Broadcom dependency on TPUs reduces concentration risk and creates competitive pricing dynamics. Google’s posture mirrors AWS’s multi-vendor logic (Trainium with Marvell + Annapurna in-house).

Recent News

  • 2026-04-09 to 2026-04-11 — Google Cloud Next 2026: Sundar Pichai unveils 8th-gen TPU (TPU 8t for training, TPU 8i for inference), MPU, and Marvell + MediaTek partnership for inference-side silicon. Marvell stock rallies on disclosure. ✓
  • 2026-04 — Marvell stock price surge on Google AI chip partnership coverage (multiple analyst notes upgrade Marvell’s Google wallet trajectory). ✓
  • 2026-Q1 (March 2026) — Marvell Q4 FY2026 earnings: management hints at “significant new design-win activity” in custom XPU pipeline; April Google announcement provides the proof point. ✓
  • 2025-Q4 — Industry reporting that Google was actively shopping a second TPU vendor; Marvell and MediaTek surface as front-runners. ◐
  • 2024-2025 — Broadcom maintains exclusive TPU partnership for TPU 6/7 generation; Google CFO commentary confirms multi-year Broadcom commitment.

Risks & Catalysts

Risks

  • Nascent revenue, execution-dependent: TPU 8i and MPU programs are early; revenue contribution in FY2026 is minimal. Slips on either program directly delay Marvell’s Google upside.
  • MediaTek BOM-share split: The Marvell/MediaTek partnership terms are not disclosed. If MediaTek captures the lion’s share of design-win value, Marvell’s Google revenue at maturity could be lower than wallet-share modeling suggests.
  • Broadcom defense: Broadcom retains training-TPU exclusivity and may aggressively bid to expand into inference; cross-link broadcom.
  • TPU volume sensitivity to Google AI infrastructure capex: Google has guided $175-185B FY2026 capex; any sharp moderation reduces TPU build rates and Marvell-attributable revenue.
  • Inference vs. training mix shift: If Google’s training-side compute grows faster than inference-side (less likely given agentic-AI tailwinds), Marvell’s inference-only positioning underperforms.

Catalysts

  • TPU 8i first-volume deployment confirmation (likely H2 2026 or 2027) — would validate Marvell’s inference role.
  • Marvell explicit Google revenue disclosure — first quarterly call where Google is implicitly identified as a top-3 customer.
  • MPU design-win details — any technical disclosure on MPU would tighten the Marvell BOM-share modeling.
  • Multi-gigawatt TPU pod build commitments — Google has indicated multi-GW deployments by FY2028; each gigawatt is roughly proportional to TPU 8i volume.
  • Google Cloud Next 2027 follow-on disclosures — next pinch point for visibility.

Sources

KB cross-references