Broadcom Deep Dive: How Custom Silicon, Networking & VMware Are Powering the AI Boom
Executive Summary
Broadcom has undergone one of the most consequential transformations in the semiconductor industry. Once known as a diversified chipmaker with a sprawling portfolio, the company has pivoted decisively toward AI infrastructure—and the results are staggering. In Q2 FY2026, Broadcom's AI semiconductor revenue reached $10.8 billion, up 143% year-over-year, representing 49% of total revenue. For the full fiscal year 2026, management projects AI semiconductor revenue of $56 billion, up approximately 180% from fiscal 2025. By fiscal 2027, that figure is expected to exceed $100 billion.
This transformation rests on three pillars. First, custom AI accelerators (XPUs): Broadcom co-designs specialized silicon for hyperscalers including Google, Meta, Anthropic, OpenAI, and two additional undisclosed customers. Second, AI networking: Broadcom's Ethernet switches—led by the Tomahawk 6, the world's first 102.4 Tbps switch—are the connective tissue binding massive AI clusters together. Third, infrastructure software: The $69 billion acquisition of VMware has created a high-margin, recurring-revenue anchor that balances semiconductor cyclicality while positioning Broadcom at the center of enterprise private cloud and hybrid infrastructure.
The core editorial question asks whether Broadcom is becoming the infrastructure company behind the AI boom. The evidence suggests the answer is yes—but with important caveats. Broadcom's position across custom silicon, networking, and enterprise software is uniquely defensible, yet it faces intensifying competition from NVIDIA, Marvell, and the ever‑present risk of hyperscaler insourcing. What makes Broadcom distinctive is not dominance in any single layer, but control across multiple layers of the AI infrastructure stack.
Company Overview
Broadcom Inc. (NASDAQ: AVGO) is a global technology company headquartered in Palo Alto, California, with a dual business model spanning semiconductor solutions and infrastructure software.
The company's semiconductor portfolio includes custom AI accelerators (XPUs), networking chips (Ethernet switches, DSPs, PHYs), storage controllers, broadband components, and wireless connectivity solutions. Its infrastructure software segment, anchored by VMware, provides virtualization, cloud management, and private cloud platforms.
Key Financial Metrics (FY2025)
| Metric | FY2025 | YoY Change |
|---|---|---|
| Total Revenue | $64 billion | +24% |
| Semiconductor Revenue | $37 billion | Record |
| AI Semiconductor Revenue | $20 billion | +65% |
| Infrastructure Software Revenue | $27 billion | +26% |
| AI Backlog | $73 billion | — |
Source: Broadcom Q4 FY2025 earnings
Key Financial Metrics (Q2 FY2026)
| Metric | Q2 FY2026 | YoY Change |
|---|---|---|
| Total Revenue | $22.2 billion | +48% |
| AI Semiconductor Revenue | $10.8 billion | +143% |
| AI as % of Total Revenue | 49% | — |
| Operating Margin | 67% | Record |
| Bookings | $30B+ AI semiconductors | — |
Source: Broadcom Q2 FY2026 earnings
Broadcom's Transformation
From Diversified Semiconductor to AI Infrastructure Platform
Broadcom's evolution is a story of strategic consolidation and deliberate pivot. Under CEO Hock Tan, the company built its semiconductor business through a series of acquisitions—including LSI, Avago, Brocade, and CA Technologies—creating a diversified portfolio spanning storage, networking, broadband, and enterprise software. But the real transformation began with the recognition that AI would reshape the entire computing landscape.
The pivot has been decisive. In fiscal 2025, AI semiconductor revenue reached $20 billion, accounting for 31% of total revenue. By Q2 FY2026, that share had jumped to 49%. The company's AI revenue has gone from representing just 15% of total revenue two years ago to 44% in early 2026. This is not incremental growth—it is a fundamental repositioning of the company.
Strategic Acquisitions
Two acquisitions stand out as foundational to Broadcom's current position:
- The VMware Acquisition ($69 billion, closed November 2023): The VMware acquisition is the single most transformative deal in Broadcom's history. It added approximately $12 billion in annual recurring revenue and gave Broadcom a commanding position in enterprise virtualization and private cloud. The integration is expected to cost approximately $1 billion in transition spending, but the combined entity is projected to generate $50 billion in revenue.
- Networking and Connectivity Acquisitions: Broadcom's networking dominance was built through organic innovation (Tomahawk, Jericho, Trident switch families) supplemented by strategic acquisitions in optical and connectivity technologies.
Semiconductor + Software: A Complementary Model
The genius of Broadcom's strategy lies in how its two businesses complement each other:
- Semiconductor provides high-growth, high-volatility revenue tied to AI infrastructure buildouts.
- Software provides stable, high-margin recurring revenue that smooths cyclicality.
Infrastructure software gross margins reach approximately 93%, with operating margins above 78%. This software annuity funds AI semiconductor expansion while insulating the company from the worst of semiconductor cyclicality.
Custom AI Accelerators
The XPU Opportunity
Broadcom's custom AI accelerator business is the primary engine of its AI growth. Unlike NVIDIA, which sells merchant GPUs to any customer, Broadcom co-designs application-specific integrated circuits (ASICs)—custom chips optimized for a specific hyperscaler's workloads.
These chips are known as XPUs (a broad term encompassing TPUs, custom accelerators, and other specialized AI silicon). Broadcom does not brand or sell these chips independently; instead, it acts as a design and manufacturing partner, coordinating fabrication through TSMC and delivering finished silicon to its customers.
Customer Relationships
Broadcom's confirmed custom silicon customers include:
| Customer | Program | Key Details |
|---|---|---|
| TPU | Long-term agreement signed April 2026 for multiple generations; projected to exceed $10B from TPU alone in 2025 | |
| Meta | MTIA XPU | 3 GW deployment through end of 2028; initial 1 GW order begins H2 2027 |
| Anthropic | TPU-based compute | 1+ GW in 2026; additional 5 GW beginning 2027 |
| OpenAI | Custom XPU | First-gen ships in 2026; 1.3 GW commitment in 2027 as part of 10 GW agreement by 2029 |
| Two additional | Undisclosed | $6 billion in purchase orders secured |
Source: Broadcom Q2 FY2026 earnings call
Broadcom is estimated to hold approximately 75% market share in custom ASIC AI accelerators.
Why Hyperscalers Are Developing Custom Silicon
The shift toward custom silicon is not a niche trend—it is a structural transformation in how hyperscalers approach AI infrastructure. The logic is compelling:
- Cost Efficiency: Custom AI chips cost between $6 billion and $11 billion per gigawatt, compared to approximately $19 billion for NVIDIA GPU racks. At hyperscale, every percentage point of efficiency matters when spending tens of billions annually on AI infrastructure.
- Workload Optimization: GPUs are general-purpose processors designed for a wide range of workloads. XPUs can be optimized for specific models, inference patterns, and software stacks. Once models are trained (which still favors NVIDIA GPUs), they spend years handling inference—workloads that often reward lower costs and higher efficiency rather than raw compute power.
- Supply Chain Diversification: With combined AI infrastructure CapEx exceeding $630 billion planned for 2025, hyperscalers are diversifying away from single-vendor GPU dependency—a lesson the server CPU market taught a decade ago.
- Strategic Control: Owning the silicon roadmap gives hyperscalers control over their own destinies.
Competitive Position Against NVIDIA
Broadcom and NVIDIA are not direct competitors in the traditional sense—they operate in different segments of the AI chip market.
- NVIDIA's Advantage: Dominant position in AI training with GPUs; CUDA software ecosystem creates powerful lock‑in; data center revenue of $75.2 billion in Q1 FY2027, up 92% YoY.
- Broadcom's Advantage: Custom silicon is more cost‑effective at scale for inference workloads; ~75% market share in custom ASIC AI accelerators; relationships with six of the world's largest AI infrastructure builders; networking portfolio creates additional value beyond the accelerator itself.
However, NVIDIA is not standing still. The company is reportedly supporting Marvell to suppress Broadcom's market share monopoly, leveraging Marvell as a strategic counterweight in custom silicon.
AI Networking
The Networking Imperative
As AI clusters scale from thousands to millions of XPUs, the network has become a critical bottleneck. Training a large language model requires moving massive amounts of data between accelerators with minimal latency. High-speed networking is no longer a supporting function—it is a core determinant of cluster performance.
Broadcom's networking portfolio addresses this challenge across multiple dimensions:
- Ethernet Switches: Tomahawk and Jericho families provide switching capacity that binds AI clusters together.
- Optics and Interconnects: DSPs, lasers, co‑packaged optics (CPO), and fabric solutions enable high‑speed data movement.
- Scale‑Up and Scale‑Out: Technology supports both connecting accelerators within a rack and connecting racks across a data center.
Tomahawk 6: A Breakthrough Product
In June 2025, Broadcom began shipping the Tomahawk 6, the world's first 102.4 Terabits per second Ethernet switch—double the bandwidth of any Ethernet switch on the market.
- 102.4 Tbps switching capacity in a single chip
- Support for 1 million+ XPU clusters
- Cognitive Routing 2.0 with advanced telemetry, dynamic congestion control, and adaptive flow control
- Co‑packaged optics (CPO) options for lowest power and latency
- Unified scale‑up and scale‑out Ethernet fabric, eliminating proprietary lock‑in
Tomahawk 6 enables a 128,000‑XPU network using only two switch tiers. This represents a step‑function improvement in AI cluster economics.
Ethernet vs. InfiniBand
NVIDIA's networking business is built on InfiniBand, a proprietary interconnect technology. Broadcom champions Ethernet—an open, standards‑based fabric. Tomahawk 6 gives hyperscalers an open fabric free of proprietary lock‑in.
Networking Revenue
Networking represented nearly 40% of Broadcom's Q2 FY2026 AI revenue. Management expects this to normalize closer to 30% of AI revenue as custom XPU shipments accelerate.
VMware Strategy
The Acquisition
Broadcom's $69 billion acquisition of VMware, completed in November 2023, was the largest technology acquisition of its era. VMware generated approximately $12 billion in annual revenue at the time of acquisition.
Integration and Restructuring
- Licensing Shift: Eliminated perpetual licensing in favor of subscription‑based licensing.
- Portfolio Refocus: Refocusing VMware on core private and hybrid cloud environments; divesting non‑core assets including Carbon Black and End‑User Computing.
- Pricing Restructuring: VCF subscription list prices slashed in half while adding higher support service levels.
- Customer Disruption: Significant customer pushback over pricing increases and forced migrations.
VMware Cloud Foundation (VCF)
The strategic centerpiece is VMware Cloud Foundation 9.0, positioned as an "AI Native Platform". VCF 9.1 (May 2026) delivered AI optimizations including up to 40% reduction in server costs, up to 39% lower storage TCO, and 4x faster cluster upgrades.
Strategic Importance
- Financial Stability: Infrastructure software generated ~$27B in FY2025, with operating margins above 78%.
- Enterprise AI Exposure: Positions Broadcom for enterprise private AI infrastructure.
- Customer Relationships: Enterprise customers provide cross‑sell opportunities across the portfolio.
Hyperscaler Relationships
Confirmed customers: Google, Meta, Anthropic, OpenAI, plus two additional undisclosed customers, and Apple reported as a new customer in 2026. Each relationship involves multi‑year, multi‑generation agreements, creating significant switching costs.
AI Infrastructure Revenue
Revenue Breakdown
- Custom AI Accelerators (XPUs): Largest component; Q2 FY2026 revenue $10.8B, up 143% YoY.
- AI Networking: ~40% of Q2 AI revenue; expected to normalize at ~30%.
- Infrastructure Software: VMware VCF increasingly positioned as AI infrastructure; Q1 FY2026 software revenue $6.8B.
Growth Trajectory
| Period | AI Semiconductor Revenue | YoY Growth |
|---|---|---|
| FY2025 | $20B | +65% |
| Q4 FY2025 | $6.2B | +74% |
| Q1 FY2026 | $8.4B | +106% |
| Q2 FY2026 | $10.8B | +143% |
| Q3 FY2026 (guide) | $16.0B | +200%+ |
| FY2026 (projected) | $56B | +180% |
| FY2027 (projected) | $100B+ | — |
Source: Broadcom earnings; Goldman Sachs estimates
Competitive Moat
- Semiconductor IP and Engineering: Decades of expertise in SERDES, Ethernet switching, co‑packaged optics, advanced packaging, and ASIC design.
- Customer Integration and Switching Costs: Co‑design investment, software optimization, multi‑generation agreements, and supply chain integration create high barriers to switching.
- The Custom Silicon Moat: Scale, foundry relationships, customer trust, and track record across multiple generations.
Competitive Landscape
Comparison with key competitors:
| Factor | Broadcom | NVIDIA | Marvell | AMD |
|---|---|---|---|---|
| AI Revenue (latest Q) | $10.8B | $75.2B | ~$2.4B | $5.8B |
| Business Model | Custom ASIC | Merchant GPU | Custom ASIC | Merchant GPU/CPU |
| Gross Margin | ~67% | ~75% | ~52% | ~50% |
| Networking | Ethernet leader | InfiniBand | Ethernet | No |
| Software | VMware (enterprise) | CUDA (developer) | Limited | ROCm |
| Moat | Integration + switching costs | CUDA ecosystem | Niche | CPU/GPU portfolio |
Sources: Company earnings
Growth Drivers
- Hyperscaler AI Infrastructure Buildout: $700B+ planned CapEx in 2025.
- Enterprise AI Adoption: VMware Cloud Foundation as "AI Native Platform".
- Inference at Scale: Custom silicon more cost‑effective than GPUs for inference.
- Networking Upgrades: Tomahawk 6 drives replacement cycles.
- Customer Expansion: From three to six hyperscaler customers.
Risks & Challenges
- Customer Concentration – heavy reliance on a few hyperscalers.
- Hyperscaler Insourcing – risk that customers bring chip design in‑house.
- NVIDIA Competitive Pressure – NVIDIA supporting Marvell as a counterweight.
- Semiconductor Cyclicality – non‑AI segments face downturns.
- VMware Customer Disruption – pricing and licensing changes may cause churn.
- Regulatory Considerations – antitrust scrutiny.
- Integration and Execution Risks – scaling AI while integrating VMware.
- AI Infrastructure Spending Normalization – potential slowdown in CapEx growth.
Strategic Positioning
Broadcom is positioning itself as an integrated AI infrastructure provider spanning custom silicon, networking, and enterprise software. The VMware acquisition provides a stable software annuity and cross‑sell opportunities. The company's "gatekeeper" position is defensible due to deep customer relationships, switching costs, and technology leadership across multiple layers.
Company Deep Dive Conclusion
Is Broadcom becoming the infrastructure company behind the AI boom? The evidence strongly supports an affirmative answer—with important qualifications. Broadcom has established itself as the leading supplier of custom AI accelerators to hyperscalers, with ~75% market share and relationships with six of the world's largest AI infrastructure builders. Its networking portfolio—led by Tomahawk 6—provides the connective tissue for massive AI clusters. And its VMware acquisition has created a high‑margin, recurring‑revenue software business that balances semiconductor cyclicality while positioning Broadcom for enterprise AI adoption.
The financial trajectory is extraordinary: AI semiconductor revenue projected from $20B in FY2025 to $56B in FY2026 to over $100B in FY2027. However, risks include customer concentration, hyperscaler insourcing, NVIDIA competitive pressure, and VMware integration challenges. What makes Broadcom distinctive—and potentially defensible—is its position across multiple layers of the AI infrastructure stack. No other company combines custom AI silicon, high‑speed Ethernet networking, and enterprise infrastructure software at Broadcom's scale.
Key Metrics and Financial Data
| Metric | Value | Period |
|---|---|---|
| Total Revenue | $22.2B | Q2 FY2026 |
| AI Semiconductor Revenue | $10.8B | Q2 FY2026 |
| AI as % of Total Revenue | 49% | Q2 FY2026 |
| Networking as % of AI Revenue | ~40% | Q2 FY2026 |
| Operating Margin | 67% | Q2 FY2026 |
| AI Bookings | $30B+ | Q2 FY2026 |
| FY2025 Total Revenue | $64B | FY2025 |
| FY2025 AI Revenue | $20B | FY2025 |
| FY2025 Infrastructure Software | $27B | FY2025 |
| AI Backlog | $73B | FY2025 |
| FY2026 AI Revenue Guidance | $56B | FY2026 |
| FY2027 AI Revenue Guidance | $100B+ | FY2027 |
Sources: Broadcom earnings, Goldman Sachs
Sources
- Broadcom Q2 FY2026 Earnings Call Highlights — Nasdaq
- Broadcom Q4 FY2025 Earnings Call Transcript — StockAnalysis
- Broadcom Q4 FY2025 Earnings Call Transcript — The Motley Fool
- "Broadcom's $100B custom silicon bet" — Jon Peddie Research
- "Broadcom Builds Custom Chips for Google, Meta, Anthropic, and OpenAI" — Nasdaq
- "Broadcom and Marvell Are Bigger Threats to Nvidia Than You Know" — Yahoo Finance
- "VMware integration to cost Broadcom $1B" — SDxCentral
- Broadcom Ships Tomahawk 6 — Broadcom Investor Relations
- "Broadcom Is Flexing Its AI Infrastructure Software Muscle" — Seeking Alpha
- "Broadcom's AI Story Is Still Early" — Seeking Alpha
- Goldman Sachs urges investors to buy Broadcom dip — Proactive Investors
- "How Much Data Center Revenue Do AI Companies Bring In?" — The Motley Fool
- "Broadcom Statistics 2026" — Axis Intelligence
- "Broadcom Q2 AI Surge Tests Valuation" — Yahoo Finance
- "Broadcom Makes VMware Cloud Foundation an AI Native Platform" — Broadcom News
Reviewed by Erwin Castro
on
Sunday, August 09, 2026
Rating: