Broadcom AI Networking and Custom Chips: Deep Investment Analysis

The Quiet AI Giant: How Broadcom Became Essential

Broadcom does not have an AI narrative problem. It has a marketing problem. While NVIDIA's Jensen Huang commands keynote stages in a leather jacket and Palantir's Alex Karp philosophizes about Western civilization, Broadcom's Hock Tan delivers deliberately understated earnings calls, provides conservative guidance, and lets the numbers speak. This communication style has contributed to a persistent valuation discount relative to Broadcom's actual AI exposure — a discount we believe represents an opportunity.

Consider the facts. Broadcom designs the custom AI accelerator chips (XPUs) that power Google's TPU infrastructure. It designs custom chips for Meta's AI training and inference workloads. It designs custom chips for ByteDance's recommendation and video processing systems. These are three of the five largest AI compute buyers on Earth. In parallel, Broadcom supplies the networking silicon — switches, routers, and network interface cards — that connects GPU and TPU clusters in every major AI data center globally. If NVIDIA provides the compute engine for the AI revolution, Broadcom provides both alternative engines and the nervous system that ties them together.

Total AI-related semiconductor revenue was approximately $12.2 billion in fiscal 2024 (ending October) and is tracking toward $16–20 billion in fiscal 2025. This represents roughly 45% of Broadcom's total semiconductor revenue, up from 15% just two years ago. The pace of this transformation is extraordinary and not yet fully reflected in the stock's valuation.

XPU Custom Chips: The Anti-NVIDIA Play

Why Hyperscalers Want Custom Silicon

The motivation for custom AI chips is straightforward: cost and efficiency. NVIDIA's H100 and B200 GPUs are general-purpose accelerators designed to run any AI workload. This flexibility comes at a price — literally (H100s cost $25,000–40,000 each) and in terms of computational efficiency. A chip designed for one specific task will always outperform a general-purpose chip on that task, just as a Formula 1 car outperforms a Toyota Camry on a racetrack.

Google understood this earliest. Its first TPU (Tensor Processing Unit), designed in collaboration with Broadcom, was deployed in 2015 — years before the current AI boom. TPUs are optimized for the matrix multiplications that dominate neural network computation, and they power the majority of Google's internal AI workloads: Search ranking, YouTube recommendations, Gmail spam filtering, Translate, and Gemini inference. Google has never disclosed the exact number of TPUs deployed, but estimates range from 1–2 million chips across its global data center fleet.

Meta followed a similar logic. Its MTIA (Meta Training and Inference Accelerator) chips, designed with Broadcom, are optimized for the recommendation models that drive Facebook and Instagram ad targeting. These models are the economic engine of Meta's $160 billion annual revenue — even a 5% improvement in recommendation quality translates to billions in incremental ad revenue. ByteDance, operator of TikTok, has similar motivations for its custom AI chips, which optimize for the video recommendation algorithms that power the world's fastest-growing social media platform.

Broadcom's Role: Design IP, Not Manufacturing

A critical nuance: Broadcom does not manufacture custom chips. It provides the design IP, chip architecture, SerDes (high-speed serializer/deserializer) technology, packaging design, and TSMC manufacturing relationship that enables hyperscalers to bring custom silicon to market. Think of Broadcom as the architect and general contractor, while TSMC is the construction company and the hyperscaler is the building owner.

This model carries significant advantages. Broadcom does not bear the capital expenditure of building fabrication facilities (which cost $20–40 billion). It does not take inventory risk on finished chips. Its revenue is driven by design fees, licensing royalties on IP blocks, and per-chip royalties on manufactured units. Gross margins on XPU revenue are estimated at 70–75%, comparable to NVIDIA's data center margins, but with lower capital intensity and more predictable revenue streams from multi-year design engagements.

AI Networking: The Overlooked $15 Billion Opportunity

Investors fixated on the GPU and custom chip narrative are missing the networking story. Training a large language model on 10,000+ GPUs requires each GPU to communicate with thousands of others, exchanging gradient updates, activation values, and model parameters at terabit-per-second speeds. The networking fabric that connects these GPUs is as critical to training performance as the GPUs themselves. A cluster with inferior networking is like a supercar on a dirt road — the engine's potential is wasted.

NVIDIA has historically dominated this market through InfiniBand, a high-performance networking standard it acquired with its 2020 purchase of Mellanox Technologies. InfiniBand delivers superior latency and bandwidth for tightly coupled workloads, which is why it powers the majority of existing large-scale training clusters.

But the market is shifting. Ethernet — the open networking standard that Broadcom dominates — is closing the performance gap while offering significant advantages in cost, vendor flexibility, and operational simplicity. Broadcom's Tomahawk 5 switch ASIC delivers 51.2 terabits per second of switching capacity, sufficient for the largest AI clusters. Its Jericho3-AI chip enables ultra-low-latency fabric architectures optimized for AI training. And because Ethernet is an open standard, hyperscalers can avoid vendor lock-in with NVIDIA's proprietary InfiniBand ecosystem.

The Ultra Ethernet Consortium — backed by Microsoft, Meta, AMD, Broadcom, and Intel — is explicitly building Ethernet specifications for AI workloads, aiming to match InfiniBand's performance while maintaining Ethernet's cost and flexibility advantages. Meta has publicly committed to Ethernet-based AI networking for its next-generation data centers. Microsoft is deploying Ethernet alongside InfiniBand. Google has always used Ethernet internally.

We estimate Broadcom's AI networking revenue at $4–5 billion in fiscal 2025, growing at 40%+ annually. By fiscal 2027, this could reach $8–10 billion as Ethernet captures a majority share of new AI data center builds. Combined with XPU revenue, Broadcom's total AI-related semiconductor revenue is on a path to $25–30 billion by fiscal 2027 — a figure that would represent 55–60% of total company revenue. For a broader perspective on how the custom chip versus GPU dynamic is evolving, see our deep dive on custom AI chips versus NVIDIA GPUs.

VMware: The Cash Cow the Market Undervalues

Broadcom closed its $69 billion acquisition of VMware in November 2023, making it one of the largest technology acquisitions in history. The market's initial reaction was skepticism — VMware was a mature infrastructure software company with decelerating growth, and the acquisition added significant debt to Broadcom's balance sheet. Two years later, the integration has exceeded expectations.

Hock Tan's playbook is by now well understood: acquire large software companies, eliminate cost redundancies, convert perpetual licenses to subscriptions, and focus R&D on the highest-margin product lines. With VMware, the subscription conversion is tracking ahead of plan. VMware's annualized booking value (ABV) for subscription licenses exceeded $5 billion in the most recent quarter, up from near zero pre-acquisition. Operating margins for the VMware segment have expanded from the mid-20s to mid-30s percent, with a path to 65%+ as the subscription transition completes — consistent with Broadcom's other software businesses (CA Technologies, Symantec).

The strategic relevance for the AI thesis: VMware's VCF (VMware Cloud Foundation) is the virtualization layer that many enterprises use to manage their on-premise and hybrid cloud infrastructure. As enterprises deploy AI workloads on-premise (for data security and latency reasons), VCF becomes the management plane for those deployments. This creates a synergy between Broadcom's AI semiconductor business and its software business that no competitor can match — the chips and the software that manages them come from the same company.

Broadcom vs. AI Semiconductor Peers

Risk Factors: The Fourth Customer and Concentration Concerns

Customer Concentration in XPU

Three customers represent the entirety of Broadcom's XPU revenue. Google alone may account for 50%+ of XPU revenue given the scale and maturity of its TPU program. This concentration creates outsized risk from any single customer's decision to insource chip design, shift to a competitor (Marvell is the primary alternative), or reduce AI capex spending. The loss of even one XPU customer would remove $4–6 billion in revenue and trigger a significant re-rating of the stock.

The much-anticipated “fourth customer” — rumored to be Apple or Microsoft — would meaningfully diversify this risk and could add $5–8 billion in incremental revenue by fiscal 2027–2028. Hock Tan has been cagey about confirming additional customers, stating only that “we are in active engagement with several potential XPU partners.” A formal announcement of a fourth customer would be a significant positive catalyst.

TSMC Dependency and Geopolitical Risk

All of Broadcom's advanced chips are manufactured at TSMC, primarily at the 5nm and 3nm nodes. Any disruption to TSMC — whether from a Taiwan Strait conflict, earthquake, or capacity allocation dispute — would halt Broadcom's AI chip production entirely. This is not a Broadcom-specific risk (NVIDIA and AMD face the same exposure), but it is worth noting that Broadcom has no viable manufacturing alternative for its most advanced designs. Intel Foundry Services and Samsung are years behind TSMC on leading-edge process technology.

VMware Integration Execution

While the VMware integration is tracking well by most metrics, Broadcom's aggressive license conversion strategy has alienated some customers. Reports of 2–5x price increases on VMware renewals have pushed some enterprises to evaluate alternatives like Nutanix, Red Hat OpenShift, and Proxmox. If customer attrition exceeds expectations, it could offset the margin expansion that the subscription conversion is designed to deliver. We estimate VMware customer churn at 10–15% by count but less than 5% by revenue, as larger customers with more complex environments are the least likely to migrate.

Investment Thesis: Why We Prefer Broadcom to NVIDIA at These Levels

This is our contrarian take, and we recognize it will be unpopular. We believe Broadcom offers better risk-adjusted returns than NVIDIA for new money entering the AI semiconductor trade today. Here is the reasoning.

First, diversification. Broadcom generates roughly 55% of revenue from non-AI businesses (legacy semiconductor + VMware software), which provides downside protection if AI spending decelerates. NVIDIA generates 80%+ from data center, making it more exposed to any AI capex slowdown. Second, the VMware software business produces highly predictable recurring revenue at 65%+ operating margins, creating a valuation floor that NVIDIA lacks. Third, Broadcom's XPU customers (Google, Meta, ByteDance) are locked into multi-year design partnerships that provide forward revenue visibility NVIDIA does not have — NVIDIA sells GPUs on shorter-cycle procurement.

Fourth, and most importantly, Broadcom trades at a similar forward P/E to NVIDIA (~30x) but with a fundamentally different risk profile. NVIDIA's 30x requires the AI capex boom to continue unabated and for custom silicon not to erode its market share. Broadcom's 30x requires more modest AI growth continuation plus the VMware margin expansion that is already in motion. The downside scenario for Broadcom (AI spending slows but VMware grows steadily) is far less severe than the downside for NVIDIA (AI spending slows, period).

We would add to Broadcom positions on any pullback to the $200–220 range (25–27x forward earnings), which would likely coincide with a broader semiconductor sector correction. For additional context on how the AI semiconductor value chain is evolving, see our analysis of where smart money is deploying in the AI capex boom.

Frequently Asked Questions