AI Custom Chip (ASIC) Market Explodes, Broadcom and Marvell Core Beneficiaries
AI summary card
AI Custom Chip (ASIC) Market Explodes, Broadcom and Marvell Core Beneficiaries
AI ASIC shipments expected to surpass GPUs by 2027, market size reaching $60-70 billion in 2026; Broadcom and Marvell dominate the high-end market leveraging IP advantages.
- 2026 Digital Custom AI ASIC market opportunity approx. $60-70 billion, future CAGR exceeding 40-50%
- Expected 2027 AI ASIC/XPU unit shipment share to reach 53%, surpassing GPU for the first time
- Broadcom holds 80-85% share of high-end ASIC market, expected FY26 AI revenue exceeds $60 billion
- Marvell holds 10-12% share of high-end ASIC market, expected CY26 data center revenue approx. $9.3 billion
- Cloud giants' self-developed chips rely on semiconductor companies' IP (e.g., SERDES, HBM interfaces) and design expertise
Report interpretation
Overview
This report provides an in-depth analysis of the resurgence trend in the custom chip (ASIC) market driven by artificial intelligence. The report indicates that as large OEMs, cloud service providers (CSPs), and hyperscale data centers seek differentiation at the silicon level, enhance performance, and reduce power consumption, demand for custom ASICs is surging. The market size is expected to reach $60-70 billion in 2026 and maintain a compound annual growth rate of over 40-50% in the coming years. Broadcom and Marvell, as core vendors with rich IP portfolios and design expertise, will dominate this high-growth market.
Core views
The core value of custom ASICs lies in hardware-software co-optimization. When customers control the software stack, they can tailor silicon to specific software, gaining advantages in system-level performance (compute performance/power). While NVIDIA's general-purpose GPUs currently dominate mainstream markets, self-developed chips like Google TPU, Amazon Trainium, Meta MTIA, and Microsoft Maia are driving structural changes due to their extremely high average selling prices (ASP) and strong ROI from AI monetization. In terms of market landscape, Broadcom and Marvell form a duopoly. Broadcom holds an 80-85% share of the high-end ASIC market, possessing the most extensive IP portfolio and advanced packaging technologies (e.g., 2.5D/3D SOIC, CPO). Its clients include Google (TPU v7/v8), Meta (MTIA series), OpenAI, SoftBank/ARM, etc. It is estimated that Broadcom's total FY26 AI revenue will exceed $60 billion (~$20 billion in FY25), and FY27 could track above $150 billion. Marvell holds a 10-12% share of the high-end market, primarily benefiting from projects with Amazon (Trainium 3/4), Microsoft (Maia), and Google (ARM CPU). Marvell has unique advantages in optical DSP (800G/1.6T) and custom HBM architectures; its custom HBM solutions can achieve 33% more HBM stacks per package and reduce interface power consumption by 70%. Estimated Marvell CY26 data center revenue is around $9.3 billion, increasing to $14.6 billion in CY27. Shipment forecasts show total AI accelerator unit shipments will grow from 10.1 million in 2025 to 23.3 million in 2027. Among them, the proportion of ASIC/XPU will rise from 32% in 2025 to 53% in 2027, surpassing GPUs (47%) in quantity for the first time.
Analysis framework
The research report adopts an analytical approach combining 'Supply-Demand Framework' and 'Industry Chain Transmission'. First, it analyzes the economic drivers for cloud giants developing their own chips from the demand side (ROI, differentiation, power consumption), distinguishing the value propositions between ASIC and general-purpose GPUs. Second, it evaluates technical barriers from the supply side, emphasizing that high-speed SERDES, HBM interface IP, and advanced packaging capabilities are key factors preventing cloud giants from fully self-developing and necessitating reliance on Broadcom/Marvell. Finally, by breaking down the roadmap (Pipeline) and process nodes (3nm/2nm) of major clients (Google, Meta, Amazon, Microsoft, etc.), it derives the revenue growth potential and market share of the two core suppliers from top-down.
Methodology notes
Supply and demand balance analysis in semiconductor sub-segments
The report argues the high prosperity of the ASIC market and the stability of the duopoly pattern by analyzing the demand of cloud giants for differentiated, low-power chips (demand side) and the scarcity of suppliers with high-end IP and manufacturing capabilities (supply side).
Value capture of IP licensing and design services in the chip industry chain
The report points out that cloud giants are at the downstream application layer; although they define chip functions, they lack mid-stream core IP (such as SerDes, HBM controllers) and physical design capabilities. Therefore, upstream Broadcom and Marvell can capture significant value. This industry division determines the profit distribution structure.
Revenue forecasting using project pipeline accumulation method
The report compiles the specific mass production schedules, process nodes, and estimated scales of major clients for Broadcom and Marvell (such as Google TPU, Meta MTIA, Amazon Trainium) to accumulate and derive revenue predictions for future years. This is a typical fundamental analysis method based on order visibility.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Broadcom (AVGO.US)Core beneficiary, holding 80-85% share of the high-end ASIC market
- Strengths
- Most extensive IP portfolio (SerDes, HBM interfaces), advanced 2nm/3nm reference platforms, 3D SOIC packaging technology, tied to top-tier clients like Google/Meta/OpenAI
- Comparison
- Compared to Marvell, Broadcom holds an absolute leading position in the core logic chips (XPU) of AI accelerators, with a more premium client roster
- Marvell (MRVL.US)Core beneficiary, holding 10-12% share of the high-end ASIC market
- Strengths
- Unique advantages in optical interconnects (DSP/CPO) and custom HBM architectures, deeply tied to Amazon and Microsoft, SmartNIC/DPU business strong
- Comparison
- Compared to Broadcom, Marvell is more distinctive in interconnect and storage interface areas, being the main partner for Amazon Trainium and Microsoft Maia
Key data
- 2026 Custom AI ASIC Market Size$60-70 billionFuture CAGR expected to exceed 40-50%
- 2027 ASIC/XPU Shipment Share53%Surpasses GPU's 47% for the first time, establishing quantity advantage
- Broadcom High-End ASIC Market Share80-85%Market position remains stable
- Broadcom FY26 AI Revenue Forecast>$60 billionSignificant growth compared to ~$20 billion in FY25
- Marvell CY26 Data Center Revenue Forecast~$9.3 billionGrowth compared to ~$6.1 billion in CY25
- Google TPU v7/v8 Backlog6.5 million+ shipmentsCorresponding to approx. $100 billion revenue potential
Impact & implications
The report believes that AI computing infrastructure is evolving from general-purpose GPUs to a 'GPU + Custom ASIC' hybrid architecture. For investors, this means the focus of the semiconductor value chain is shifting towards custom chip design service providers with core IP and advanced packaging capabilities. Broadcom and Marvell are not only component suppliers but also key partners in the AI strategy implementation of cloud giants; their earnings visibility increases significantly as large customer projects go into mass production. Additionally, the introduction of custom HBM and CPO (Co-Packaged Optics) technologies will further consolidate the technological moats and bargaining power of these two companies.
What to watch
- Progress in design wins for Broadcom and Marvell at the 2nm process node
- Mass production ramp-up status of Google TPU v8/v9 and Anthropic projects
- Early volume performance of Microsoft Maia 3nm ASIC in CY27
- Penetration rate of custom HBM and CPO technologies in next-generation AI clusters