Data Center AI Chips Could Change the Winners: Broadcom and Marvell

Published November 20, 2026, 6:25pm ET · 8 min read

This post may contain links from our sponsors and affiliates, and Flywheel Publishing may receive compensation for actions taken through them.

A long, symmetrical shot down a hallway in a dark blue data center. Rows of black server racks with glowing blue and green lights stretch into the distance on both the left and right sides. A large, bright blue graphic of a stylized computer chip with the letters 'AI' in white is projected onto the ceiling, casting an inverted reflection of itself onto the concrete floor. The scene is futuristic and highly technological.
A state-of-the-art data center, illuminated by an AI chip graphic, underscores the immense infrastructure driving the technology sector's leadership in artificial intelligence, as discussed by Franklin Templeton's Jonathan Curtis. © Shutterstock
  • Broadcom and Marvell stand to benefit as hyperscalers expand custom AI processors alongside merchant GPUs.
  • AI training and inference impose different demands on computing, memory bandwidth, power and cost.
  • The Information Network’s HBM model shows how processor architecture and memory content change demand.

Introduction

The selloff in semiconductor stocks has intensified questions about the sustainability of artificial intelligence infrastructure spending. Investors are reassessing the enormous capital commitments being made by hyperscalers, the returns those investments will generate, and whether the companies that dominated the first phase of the AI boom will maintain their competitive advantages.

One important development is receiving less attention than the fluctuations in Nvidia’s share price: hyperscalers are expanding their use of custom-designed AI processors. These chips can improve the economics of training and inference, reduce energy consumption, and give cloud operators greater control over their computing infrastructure. Broadcom (AVGO | AVGO Price Prediction) and Marvell (MRVL) are positioned to participate in this transition through custom silicon, networking and connectivity.

The investment question is not whether custom ASICs eliminate GPUs. It is how an expanding mix of processors changes the distribution of semiconductor revenue and HBM demand, even if total AI computing continues to grow.

Broadcom and Marvell: Two Ways to Invest in Custom Silicon

Broadcom is a leading supplier of custom AI processors and the networking chips that connect them. In April 2026, Broadcom disclosed a long-term agreement to develop and supply future generations of Google TPUs, along with networking and other components for Google’s next-generation AI racks through up to 2031. That agreement provides a tangible commercial basis for its custom silicon opportunity, rather than relying solely on broad market forecasts.

Broadcom’s competitive advantage includes the engineering resources needed to design sophisticated accelerators and the high-speed switching technologies required to connect thousands of devices. These two opportunities reinforce one another: expanding ASIC deployments can increase demand for network bandwidth even when the processor itself is designed for a narrower set of workloads. Its risks include customer concentration, expensive design programs, and pressure on pricing as cloud customers grow more technically sophisticated.

Marvell Technology offers a different mix of exposure. Its August 2026 Google agreement covers custom silicon programs associated with the TPU ecosystem, including inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory computing. Marvell is therefore competing for semiconductor content throughout the AI system, not simply for the central accelerator.

The Google agreement also contains a purchase-linked warrant arrangement: most of the shares vest in tranches tied to cumulative custom-product revenue, with one tranche for each $500 million. This creates a measurable commercial incentive, although the contractual framework does not guarantee that all contemplated revenue will materialize. Marvell’s opportunity is substantial, but its dependence on large customers and program execution warrants particular attention.

The two companies are not identical investment propositions. Broadcom has an established custom accelerator and networking franchise; Marvell offers potential incremental share gains across accelerators, controllers and connectivity. For investors, the relevant measures are design wins, volume ramps, margins and recurring generations of products—not simply the number of chips announced.

Nvidia and AMD Remain Important Competitors

Nvidia (NVDA) continues to supply the flexible GPUs and integrated computing platforms that dominate many large-model deployments. Its software ecosystem and ability to support evolving models are meaningful competitive advantages. A hyperscaler may choose a custom ASIC for a stable, high-volume workload while continuing to use Nvidia GPUs for research, new model development and demanding inference tasks. The two approaches can coexist within the same data center.

Advanced Micro Devices (AMD) also supplies GPUs for AI acceleration and offers FPGA technology for specialized, reconfigurable inference. The company has opportunities across multiple processing architectures, although it must compete against Nvidia’s installed software base and against custom silicon programs designed specifically for hyperscaler needs.

Nvidia also faces a leadership transition following Jensen Huang’s November 12 departure and the appointment of Colette Kress as interim CEO. That development creates management uncertainty, but it would be a mistake to attribute the rise of custom ASICs primarily to the transition. Hyperscalers began developing proprietary processors years earlier because of the economics of computing at enormous scale.

Training and Inference: Why the Distinction Matters

Training develops an AI model by processing large datasets and adjusting model parameters. Inference uses the trained model to generate answers, analyze information or perform tasks. Both activities can require significant computing resources, but their performance requirements differ. Training emphasizes throughput and the coordination of many accelerators; inference often emphasizes cost per token, response latency, energy consumption and the ability to serve many simultaneous users.

Inference itself is not one uniform workload. Processing an incoming prompt, known as prefill, can make effective use of parallel computation. Generating an answer token by token, known as decoding, may be constrained by the speed at which model parameters are read from memory. Long contexts and multi-step agents increase memory and computational requirements, complicating claims that inference always needs less capable processors.

GPUs offer programmability and flexibility across training and inference. Custom ASICs can be tuned to particular workloads and deliver favorable economics when deployed at scale. CPUs handle smaller models and supporting tasks; FPGAs can serve specialized low-latency applications; NPUs support edge and client-device inference. SRAM-oriented architectures provide still another approach to moving data efficiently. These processors do not all use HBM, and their adoption cannot be translated directly into HBM shipments.

What Processor Architecture Means for HBM

High-bandwidth memory consists of stacked DRAM dies that provide high data-transfer rates close to the processor. The number of HBM stacks depends on the chip design, its memory capacity and bandwidth targets, and the generation of memory employed. A GPU can be used for inference, and a custom ASIC can be used for training; architecture should not be confused with workload.

According to Table 1, The Information Network’s illustrative 2025 model assigns 7.2 million merchant GPU units eight HBM stacks apiece and 1.5 million custom ASIC units three stacks apiece. This produces 57.6 million and 4.5 million HBM stacks, respectively. These are modeled volumes, not independently verified worldwide shipments, and they exclude processors without HBM.

Table 1. Illustrative 2025 HBM Demand by Accelerator Architecture
2025 model Units (M) HBM stacks/unit HBM stacks (M) DRAM dies (M)
Merchant GPUs (Nvidia, AMD) 7.2 8 57.6 576
Custom ASICs (TPU, Trainium) 1.5 3 4.5 45
Total 8.7 — 62.1 621
Source: The Information Network illustrative model. Ten DRAM dies per stack assumed; wafer yields, die-area conversion and non-HBM processors excluded.

The model’s most useful lesson is that accelerator unit volumes do not determine memory demand by themselves. Each processor’s HBM content matters. It would be misleading to treat the model as evidence that a particular percentage of worldwide HBM consumption goes to training or inference, since both categories of processors can perform both workloads.

Why Rising ASIC Memory Content Matters

Newer custom accelerators may require substantially more memory capacity and bandwidth than earlier generations. Consequently, custom ASIC growth could increase HBM demand through both rising unit shipments and higher memory content per unit. The effect can be illustrated by holding merchant GPU shipments constant while doubling custom ASIC volumes and varying the number of stacks per ASIC.

According to Table 2, doubling ASIC shipments from 1.5 million to 3 million units raises modeled total HBM stack demand by 7.2% if each ASIC retains three stacks. If the new ASICs use six stacks, total demand rises by 21.7% relative to the original model. These are sensitivity calculations, not forecasts, and do not incorporate changes in memory efficiency or GPU configurations.

 

Table 2. HBM Demand Sensitivity to Custom ASIC Shipments and Memory Content
Metric 2025 base 2× ASICs; 3

stacks

2× ASICs; 4

stacks

2× ASICs; 6

stacks

Merchant GPU units (M) 7.2 7.2 7.2 7.2
Custom ASIC units (M) 1.5 3.0 3.0 3.0
ASIC HBM stacks/unit 3 3 4 6
Total HBM stacks (M) 62.1 66.6 69.6 75.6
Total DRAM dies (M) 621 666 696 756
Growth vs. base — 7.2% 12.1% 21.7%
Source: The Information Network model and calculated illustrative scenarios. Merchant GPU volume and configuration are held constant; ten DRAM dies per HBM stack assumed.

The sensitivity also exposes an important uncertainty. More capable memory architectures, larger on-chip caches and software optimization can improve the amount of inference delivered per unit of memory. Accordingly, neither ASIC shipment growth nor AI token growth should be used alone to forecast HBM demand. Investors must consider memory stacks per device, stack capacity, bandwidth and the number of DRAM dies in each stack.

The Memory Suppliers Have Their Own Opportunity

Micron Technology (MU) can benefit from HBM demand generated by both GPUs and custom ASICs, provided its products meet customer qualification, bandwidth and yield requirements. Its opportunity is linked to total memory content and manufacturing economics rather than allegiance to a particular accelerator architecture.

SK hynix and Samsung Electronics face the same fundamental drivers. Increasing accelerator complexity can support higher HBM content, but additional capacity, manufacturing yields and competing memory architectures influence the ultimate revenue opportunity. Processors that rely on conventional DRAM or large SRAM arrays will not create the same HBM demand as processors built around multiple high-bandwidth stacks.

The competitive advantage in memory will therefore depend on the ability to qualify new HBM generations, manufacture them at acceptable yields and deliver the bandwidth and capacity required by evolving accelerator platforms. That is a separate contest from the competition between Nvidia, Broadcom and Marvell.

Investor Takeaway

The expansion of custom data center AI chips creates a credible opportunity for Broadcom and Marvell, but it does not imply that Nvidia GPUs are becoming obsolete. Hyperscalers have compelling reasons to deploy different processors for different workloads, balancing flexibility against cost, power and performance. As a result, the AI semiconductor market can become more diversified even while demand for merchant GPUs continues to grow.

Broadcom’s established TPU and networking relationships provide exposure to large-scale deployments. Marvell’s broader Google agreement offers potential gains across inference accelerators, controllers and connectivity. Their financial outcomes will depend on how much silicon content they win, the timing of customer ramps and the profitability of those programs. Meanwhile, Micron and other memory suppliers can benefit from rising HBM content across both merchant and custom architectures.

The Information Network’s HBM sensitivity analysis illustrates the essential investment point: counting accelerators is not enough. Investors must understand which chips hyperscalers buy, how much memory and networking each system contains, and whether suppliers can convert that content into profitable revenue. In a market reassessing AI spending, those distinctions may matter more than the broad direction of AI capital expenditure.

Contact [email protected] for any questions or corrections.

Dr. Robert Castellano

Dr. Robert Castellano has over 40 years of experience analyzing the high-tech industries. He is president of The Information Network (www.theinformationnet.com). He earned a PhD degree in Chemistry from Oxford University (UK). His PhD thesis advisor, John Goodenough, won the Nobel Prize in Chemistry in 2019 for the invention of the Lithium Ion Battery. He writes with George Gilder, novelist, futurist, and economist, and his team for Eagle Financial Publishing.

All articles →