护城河对比
最多三家已覆盖的公司，逐项对照。每一项判断都是人工编辑的判断，而非公司披露的数字——并且每一项都附有其引用依据。
| Marvell Technology | Cerebras Systems | Palantir Technologies | |
|---|---|---|---|
| 护城河评级 | 窄 FY2026 10-K (filed 2026-03-11): differentiated platform IP — over 10,000 issued patents and pending applications as of 2026-01-31, plus a proven custom ASIC platform leveraging ultra-high-speed SerDes, silicon photonics, co-packaged optics and custom HBM — but Marvell itself calls its markets 'intensely competitive' with 'pricing pressures', notes customers 'have chosen to develop certain semiconductor products internally', and discloses two >=10% customers with the ten largest at 82% of FY2026 net revenue. Real, defensible IP in a concentrated, contestable customer base = narrow, not wide. https://www.sec.gov/Archives/edgar/data/1835632/000183563226000011/mrvl-20260131.htm | 窄 One-of-a-kind silicon sold to very few buyers: the IPO prospectus discloses G42 at 24.0% of 2025 revenue (85.0% in 2024), MBZUAI at 62.0% of 2025 revenue, and a December 2025 MRA with OpenAI representing 'a substantial portion of our projected revenues over the next several years' — against a competitive field the same document names as NVIDIA, AMD, Intel, hyperscaler accelerators and the major clouds. | 窄 The FY2025 10-K competition section concedes that some competitors have 'greater name recognition, longer operating histories, and larger customer bases', 'larger and more mature intellectual property portfolios', and 'substantially greater financial, technical, and other resources' — and that Palantir is 'fundamentally competing with the internal software development efforts of our potential customers'. A real moat exists but the filing's own disclosure caps it below wide. |
| 护城河类型 | 无形资产与知识产权 The moat rests on hard-to-replicate mixed-signal IP: the 10-K describes the custom ASIC platform built on ultra-high-speed SerDes, ARM compute, security, storage, silicon photonics and advanced packaging (die-to-die interconnects, chiplets, CPO, custom HBM), with multiple 5nm designs executed, 3nm in progress and a 2nm platform in development; a secondary switching-cost element comes from multi-year custom design wins co-developed to individual customer specifications. https://www.sec.gov/Archives/edgar/data/1835632/000183563226000011/mrvl-20260131.htm | 无形资产与知识产权 Wafer-scale integration is the moat's substance: the prospectus contrasts the WSE-3's 4 trillion transistors on 46,225 mm² of silicon with NVIDIA's B200 at 208 billion on roughly 1,600 mm² — a commercial chip category with one occupant. | 转换成本 The 10-K reports that 'our top three customers by revenue for the year ended December 31, 2025, have been with us for an average of ten years' and that a significant portion of revenue comes from existing customers expanding their relationships. The filing explicitly disclaims patent dependence — 'we are not dependent on any particular patent or application' — so the durable source is embedded deployment, not IP. |
| 领先地位 | 并列领先者 Leader in its optics niche, #2 in custom silicon: ~60% of high-end PAM4 DSP share (36kr, 2026-06-27, https://eu.36kr.com/en/p/3870758441178373) but an estimated 20-25% of custom AI ASIC design services versus Broadcom's ~70% (hashrateindex, 2026-05-13, https://hashrateindex.com/blog/design-partners-ai-asic-market-part-2/) — net, a co-leader in AI data-center connectivity/custom silicon behind Broadcom overall. | 明确领先者 Scoped to the category the record supports: wafer-scale integration has no other commercial occupant per the prospectus's own comparison, and the Q2 release claims — attributed, via AMD and AWS partnerships — leadership in disaggregated inference, a technique the prospectus defines. | 势均力敌 The 10-K claims only that 'we believe we generally compete favorably with our competitors', while conceding rivals' greater scale, recognition and resources. The filing does not assert market leadership, so a leader band is not groundable from it. |
| 定价权 | 中等 The 10-K characterizes Marvell's markets as having 'pricing pressures' and intensifying competition (https://www.sec.gov/Archives/edgar/data/1835632/000183563226000011/mrvl-20260131.htm), yet the Q1 FY2027 release reports 52.1% GAAP / 58.9% non-GAAP gross margin on record revenue (https://www.sec.gov/Archives/edgar/data/1835632/000183563226000014/q127_8kx522026ex-991.htm) — differentiated-IP margins, tempered by hyperscaler buyer power. | 中等 The CFO states Q2 'significantly improved core gross and operating margins compared to a year ago' — improving, from a base where three named customers hold most of the revenue and the leverage that comes with it. | 强 FY2025 gross profit of $3.7 billion on a gross margin of 82% (84% excluding stock-based compensation), up from 80% (83% ex-SBC) in FY2024 — a sustained and rising 80%+ gross margin. |
| 综述 | Marvell is a fabless data-infrastructure silicon supplier whose center of gravity has shifted decisively to the AI data center: the data center end market was $6,100.3M, 74% of FY2026 revenue, up from 40% two fiscal years earlier (FY2026 10-K). Its strongest position is electro-optics — in high-end PAM4 optical DSPs for 400G+ transceivers it holds roughly 60% share on Inphi-inherited SerDes/FEC IP, with Broadcom above 30%, the two together over 90% (36kr, 2026-06-27). In custom AI silicon it is the structural #2 design partner at an estimated 20-25% of the custom AI ASIC design-services market versus Broadcom's ~70%, anchored by AWS Trainium and Microsoft Maia wins (hashrateindex, 2026-05-13). The Q1 FY2027 release (2026-05-27) shows the flywheel turning — record $2.418B revenue (+28% YoY), Q2 guided to $2.7B mid-point (+35% YoY), management citing 'exceptional AI-related bookings' across 800G/1.6T optics, 51.2T switches, CPO/NPO and custom XPU — and the Celestial AI (Photonic Fabric) and XConn (PCIe/CXL switching) acquisitions closed in February 2026 extend the interconnect moat toward scale-up fabrics. The offsets that keep the moat narrow are in Marvell's own filing: intense competition (AMD, Alchip, Astera, Ayar, Broadcom, Credo, GUC, Lightmatter and others), hyperscaler in-housing risk, and heavy customer concentration. | Cerebras bet that AI compute should be one enormous chip rather than many small ones — 'quantities of compute and memory never before assembled on a single commercial chip,' avoiding 'the latency and the power-draw induced by the traditional approach,' in the prospectus's words — and is converting that bet into an inference cloud: Q2 2026 cloud revenue grew 281% with 600 MW of data center capacity under contract and $25.4 billion of remaining performance obligations. The concentration that funded the bet is the risk that remains: two related-party Gulf customers were 86% of 2025 revenue between them, with the OpenAI MRA now layered on top. | Customers who deploy Foundry model their operations in the Ontology and build workflows on top of it, so displacement means rebuilding that layer rather than swapping a vendor — which is why the 10-K frames its main competitor as the customer's own internal build effort. Ten-year average tenure among the top three customers is the filing's own evidence that those deployments persist. |
| 产业链位置 | Fabless supplier spanning 'data center core to network edge': it sits between hyperscaler AI compute (custom XPU/XPU-attach ASICs) and the optical layer (PAM4/coherent DSPs, CPO/LPO, DCI, AEC, PCIe retimers), outsourcing fabrication to independent CMOS foundries; the Feb-2026 Celestial AI and XConn acquisitions push it further into scale-up photonic fabric and PCIe/CXL/UALink switching (FY2026 10-K, https://www.sec.gov/Archives/edgar/data/1835632/000183563226000011/mrvl-20260131.htm). | Layer-4 wafer-scale accelerators plus an inference cloud — hardware vendor and specialized cloud in one node. | Application-layer software sitting above the cloud and model providers; AIP connects to third-party LLMs rather than supplying them. |
| 产品（份额 / 壁垒） |
|
|
|
| 长周期投票 | +0.13 权重 0.20 · 群体判断 bullish 编辑先验，未经回测。 | +0.20 权重 0.20 · 群体判断 bearish 编辑先验，未经回测。 | +0.11 权重 0.20 · 群体判断 bullish 编辑先验，未经回测。 |