Broadcom and the Rise of Custom AI Silicon

The artificial intelligence (AI) boom is redrawing the boundaries of the semiconductor industry.

While Nvidia Corp. (NASDAQ: NVDA) captures headlines for its record-setting graphics processing unit (GPU) revenues, the deeper structural transformation lies in the move toward custom silicon. Particularly important are AI-specific application-specific integrated circuits (ASICs).

Google (NASDAQ: GOOG), Amazon (NASDAQ: AMZN), Microsoft (NASDAQ: MSFT), and Meta (NASDAQ: META) are rapidly building their own chips tailored for inference and training. But in the background, Broadcom (NASDAQ: AVGO) has emerged as the critical enabler—quietly powering this transition through its dominance in custom ASIC development and AI networking infrastructure.

This article explores how Broadcom is capitalizing on the AI shift by serving as the silicon architect behind the hyperscaler arms race, and why its strategic positioning makes it one of the most underappreciated beneficiaries of the AI compute cycle.

Why U.S. Hyperscalers Are Developing Their Own ASICs

Broadcom’s role is best understood in the context of the changing chip strategy among hyperscalers. Rather than relying solely on Nvidia’s general-purpose GPUs, cloud giants like Google and Amazon Web Services are designing in-house AI chips for greater efficiency and control. Google’s TPU, Amazon’s Trainium, Microsoft’s Maia, and Meta’s Artemis are each custom-built for the specific needs of the platforms they serve. These ASICs are particularly well suited to inference workloads, which are repetitive and predictable, allowing optimization around power, latency, and cost.

While these chips help hyperscalers reduce reliance on third-party vendors, few of them have the internal capability to execute silicon development at scale. This is where Broadcom enters—serving as a design and fabrication partner, providing the glue logic, interconnects, and custom ASICs that underpin hyperscaler infrastructure.

Do Hyperscaler ASICs Threaten Nvidia’s Business?

Before evaluating the impact of custom chips on Nvidia, it’s important to understand the technical distinctions between traditional GPUs and the newer breed of task-specific ASICs now being deployed. These include LPUs (Language Processing Units), TPUs (Tensor Processing Units), and NPUs (Neural Processing Units). The following table outlines their core architectural differences.

According to Table 1, GPUs remain the most flexible processors in terms of general-purpose parallel computing, which is why they dominate AI training. However, task-specific ASICs achieve significantly better energy efficiency, latency, and compute density for inference workloads. LPUs sacrifice flexibility for speed in language tasks. TPUs are optimized for tensor operations, and NPUs emphasize tight integration between memory and logic for ultra-efficient edge computing.

According to Table 2, Nvidia is still dominant in training, controlling over 90% of that market in 2025. But ASICs have already carved out 20% of inference, and are expected to gain more ground in the years ahead. This shift supports the view that the AI chip market is bifurcating—GPUs for training, ASICs for inference—and Broadcom is positioned to capture the latter segment.

The Role of Broadcom and Marvell in AI ASICs

While hyperscalers develop internal chips, third-party players like Broadcom and Marvell are capturing a lucrative middle ground by supplying the custom silicon that powers those very ASICs. Broadcom has emerged as a foundational player in this ecosystem, building tailored AI ASICs and high-speed networking chips for the largest cloud firms. In 2024, Broadcom’s AI-related revenue reached $8 billion, up 77% year over year. Major customers include Google, AWS, Microsoft, and Meta—all of whom rely on Broadcom’s custom silicon for both AI accelerators and networking interconnects that enable large-scale AI compute.

The growth potential is staggering. Broadcom CEO Hock Tan recently disclosed that custom AI chip demand from three major customers could hit $60 to $90 billion by 2027. Among its major collaborations: Broadcom co-develops Meta’s MTIA training accelerators and has been contracted by OpenAI for two generations of high-end AI ASICs, using 3nm and 2nm process nodes with 3D SOIC advanced packaging. These partnerships cement Broadcom as a critical supplier in the AI supply chain, even as hyperscalers push forward with internal designs.

Marvell, meanwhile, plays a smaller but growing role. Its 2024 AI revenue stood at $1.2 billion. Although Marvell lags Broadcom in overall scale, it has made important strides through partnerships with Amazon and Microsoft.

In 2024, Marvell signed a five-year agreement with AWS to co-develop multi-generation ASICs, including silicon for the Trainium3 platform. This next-generation AI chip, built on a 3nm node, promises double the performance of its predecessor and will debut in AWS instances later in 2025.

Microsoft’s Maia 100, another custom ASIC supported by Marvell, further diversifies the market. Maia is optimized for LLM training and inference within the Azure ecosystem, underscoring the increasingly bifurcated AI chip landscape: Nvidia for general-purpose compute, and hyperscaler+ASIC for specialized, cloud-native optimization.

According to Table 4, the size of the gap between Broadcom and Marvell is not just in revenue, but in the breadth of customers and strategic impact. Broadcom is involved in core AI accelerator design and system-level integration, while Marvell serves more as a component supplier.

Nvidia and Broadcom: Partners in the AI Infrastructure Stack

While Nvidia and Broadcom both serve the AI data center market, they do so from opposite ends of the stack. Nvidia builds GPUs for training and inference. Broadcom designs custom ASICs for inference and provides the switches and interconnects that scale data center clusters. Their businesses are complementary, not competitive.

This distinction is clear when comparing data center revenue trends. Nvidia’s AI revenue exploded in 2023 and 2024, reaching nearly $40 billion in Q1 2025. Broadcom’s revenue climbed steadily from around $6 billion to just over $8 billion. The contrast reflects the difference between being the brand-name product versus the embedded architecture provider.

Chart 1 tracks this divergence in revenue paths. While Nvidia captures the GPU layer, Broadcom is embedded across the system—inside ASICs, across networks, and underneath hyperscaler clusters. It is the foundational supplier behind many of the AI chips that aren’t branded Nvidia.

Broadcom has positioned itself as the quiet powerhouse in the AI compute stack. It is not chasing headlines but building the architecture that supports both Nvidia GPUs and hyperscaler-designed ASICs. Whether AI workloads are trained on Nvidia clusters or inferred on custom chips from Meta or AWS, Broadcom wins either way.