Marvell Technology (MRVL)
Custom AI ASIC design services + high-speed data center interconnect optics — Broadcom's closest competitor in the custom-silicon duopoly.
1. Core Product / Service
Marvell (NASDAQ: MRVL) is a semiconductor company that competes directly with broadcom in two AI-critical areas:
Custom AI accelerators (XPUs) — Marvell designs custom AI ASICs for hyperscalers. Confirmed programs include Amazon's Trainium (both generations), Microsoft Maia inference chip, and additional unnamed hyperscaler programs. Marvell closed the Celestial AI acquisition (Q1 calendar 2026) to add photonic fabric technology for intra-rack/intra-system connectivity — a key enabler for 1.6T and 3.2T optical connections inside AI clusters.
Data center interconnect (optics + networking) — Marvell's Brightlane PHY and custom SerDes power 800G/1.6T optical modules; the Celestial AI deal adds co-packaged optics (CPO) roadmap toward 200T switch systems. Optical interconnect business growing >50% YoY.
Storage controllers — PCIe/storage controllers for AI servers.
Marvell is the second-largest custom ASIC design service provider after broadcom, with estimated ~25% share of the custom AI accelerator market vs Broadcom's ~60%+ (Counterpoint, 2026 estimates).
2. Target Users & Pain Points
- Hyperscalers building custom silicon — Amazon (Trainium + Inferentia), Microsoft (Maia), plus unnamed additional customers
- Optical module makers — Marvell's PHY/SerDes inside 800G/1.6T modules from coherent, lumentum, innolight and others
- AI cluster operators — need lower-latency, lower-power optical interconnects as GPU counts scale
Pain solved: same as broadcom — two paths out of NVIDIA dependency (custom ASIC + open Ethernet/optical fabric). Marvell differentiates with optical integration (Celestial AI photonic fabric) and a more modular engagement model vs Broadcom's full-stack design-service approach.
3. Competitive Landscape
| Company | Approach | Positioning vs Marvell |
|---|---|---|
| broadcom | Full custom ASIC + Tomahawk switch silicon | ~60% custom ASIC share; Marvell at ~25%; Broadcom is the scale leader |
| nvidia | Merchant GPU + InfiniBand + NIM | Direct competitor for AI accelerator dollars; custom ASIC is the alternative |
| Intel Gaudi | Merchant AI accelerator | Limited AI datacenter traction |
| Astera Labs / Credo | Connectivity retimers | Adjacent; no custom ASIC capability |
Broadcom's moat: deeper hyperscaler integration (Google, Meta, OpenAI, Anthropic are all Broadcom programs). Marvell's moat: Amazon as anchor customer (Trainium/Inferentia) + Celestial AI photonics + faster growth rate (~42% revenue growth in FY2026 vs Broadcom's ~24%).
4. Unique Observations
- FY2026 full-year: Revenue $8.2B (+42% YoY) — record, driven by AI datacenter demand [1]. Data center revenue surpassed $6B for FY2026 [1]. Custom business doubled to $1.5B in FY2026 [3].
- FY2027 guidance raised sharply: approaching $11B (from prior $10B guidance in Dec 2025 and $9.5B in Sep 2025) — >30% YoY growth [1][3]. Q4 FY2026 was record: $2.219B revenue (+22% YoY) [3].
- Data center revenue growth: FY2027 expected +40% YoY; interconnect business growing >50% YoY [1][3].
- Custom ASIC >20% growth in FY2027, doubling in FY2028 — the custom ramp is just beginning [3].
- Celestial AI acquisition: adds photonic fabric for intra-rack connectivity; target $500M annualized revenue by Q4 FY2028, $1B by Q4 FY2029 [1][3]. This is the most strategically significant deal — CPO/co-packaged optics is the next bottleneck after 1.6T optical modules.
- Customer concentration risk is real: MRVL's top 2-3 customers likely represent >50% of custom revenue. Amazon Trainium/Inferentia is the anchor; any program loss would be material.
- Key risk — vertical integration: If Amazon (MRVL's largest custom customer) chooses to bring more chip design in-house (as they've done with Nitro, Graviton, Inferentia), MRVL's Amazon revenue is at risk. OpenAI's chip team could similarly displace future MRVL programs.
- vs Broadcom on token cost chain: MRVL's AI story is arguably faster-growing but narrower — less switching silicon exposure and less hyperscaler diversity than AVGO. Where Broadcom captures 30-50% of silicon BOM on custom silicon clusters, Marvell captures less but has more optics upside via Celestial AI CPO roadmap.
5. Financials / Funding
- Listed: NASDAQ: MRVL; market cap ~$90B+ [2026-05-11]
- FY2026 full-year revenue: $8.2B (+42% YoY) [1]
- FY2026 data center revenue: >$6B [1]
- FY2026 custom business: $1.5B (doubled from FY2025) [3]
- FY2027 revenue guidance: approaching $11B [1]
- FY2027 data center revenue growth: ~+40% YoY [1]
- FY2027 interconnect growth: >+50% YoY [1]
- FY2026 Q4 revenue: $2.219B (+22% YoY) [3]
- Q1 FY2026 data center: $1.44B (+76.5% YoY) [1]
- Celestial AI target: $500M annualized revenue by Q4 FY2028; $1B by Q4 FY2029 [3]
6. People & Relationships
- CEO: Murphy (since 2016)
- Strategic acquisitions: Celestial AI (photonics, 2026), XConn (connectivity ICs, 2026)
- Anchor customer: Amazon (Trainium + Inferentia custom ASIC programs)
- Additional hyperscaler programs: Microsoft Maia, unnamed others
- Optical DSP customers: same optical module ecosystem as broadcom
- Strategic supplier: tsmc (advanced node + CoWoS)