AI GPU drives ASIC opportunities amid Broadcom, Marvell rivalry

The explosive growth of ChatGPT has triggered unprecedented demand for artificial intelligence (AI) computing power, leading to industry-wide supply constraints. While Nvidia maintains its stronghold as the premier AI GPU provider, major cloud service providers (CSPs) like Google, Amazon, and Microsoft are actively diversifying beyond Nvidia's solutions by accelerating their internal chip development initiatives.

For these large CSPs, customized application-specific integrated circuits (ASICs) have become increasingly attractive as they seek to optimize chips for their specific workload requirements. These ASICs deliver superior performance and energy efficiency for specific computing tasks, making them a strong alternative to Nvidia's general-purpose GPUs. Despite Nvidia's historical market dominance, the combination of high power consumption and premium pricing in their GPU lineup has created an opportunity for ASICs to establish a competitive foothold.

ASICs are particularly well-suited for AI inference and training tasks, offering significantly better performance-per-watt ratios than general-purpose chips. They also grant CSPs greater control over their technical stack. Within the ASIC market, Broadcom and Marvell stand out as key competitors, each employing distinct technologies and strategic approaches to meet the rising demand for AI-optimized hardware.

Marvell has strengthened its position in the AI chip market, notably through its partnership with Google on Arm CPUs, while maintaining Amazon as its primary computing power client.

Google's TPU v6e represents the most advanced AI ASIC among the chips developed by the four leading CSPs, coming close to the performance of Nvidia's H100. However, it still lags behind Nvidia GPUs by approximately two years.

AI GPUs vs. AI ASICs

Note: The four major CSPs collaborate with Broadcom, Marvell, and other ASIC design service providers to develop custom AI chips.
Source: DIGITIMES, January 2025

Marvell's collaboration with Amazon on the Trainium 2 chip has achieved performance levels between Nvidia's A100 and H100, setting the stage for substantial ASIC revenue growth in 2024.

During its earnings call, Marvell disclosed projections for significant ASIC revenue growth beginning in 2024 (fiscal year 2025), driven by Amazon's Trainium 2 and Google's Axion Arm CPU processors.

Marvell's Inferential ASIC project with Amazon is scheduled to enter mass production in 2025 (fiscal year 2026), while its Microsoft Maia project is expected to generate revenue starting in 2026 (fiscal year 2027).

According to Morgan Stanley, while Marvell's customized chip business is a key growth driver for its data center division, it also carries considerable uncertainty.

Short-term forecasts for Marvell's Trainium product line are positive, supported by TSMC's increased CoWoS capacity, Amazon's production expansion plans, and Marvell's confidence in market demand.

However, the long-term competitive landscape presents challenges. The emergence of companies like WorldChip Electronics in the computing chip sector may force Marvell to refocus on networking chips. Additionally, a potential decline in Amazon's Trainium series sales after 2026 means Marvell must secure new client projects to maintain growth momentum.

Strategic divergence between industry leaders

Broadcom and Marvell exemplify contrasting development strategies within the ASIC sector. Broadcom prioritizes large-scale integration and platform design, backing its approach with substantial R&D investments and sophisticated technological integration. In contrast, Marvell pursues growth through strategic acquisitions, having expanded its technological portfolio by acquiring companies such as Cavium, Avera, and Innovium.

These divergent strategies shape how Broadcom and Marvell excel in customer service, technology ecosystems, and overall market positioning.

Broadcom's competitive edge stems from its consistent R&D investments and advanced integration capabilities, enabling it to sustain leadership in network interconnectivity and ASIC design over the long term.

The company maintains its dominance in Ethernet switching chips and optical interconnect products through both technological superiority and market share leadership. Its economies of scale and efficient supply chain management further reinforce this competitive advantage.

While Marvell's technological foundation is comparatively less developed, it has rapidly established expertise in ASICs and data centers through targeted acquisitions, focusing particularly on ASIC solutions for AI server applications.

Marvell has secured its place among the top players in the global ASIC market, serving major cloud providers like Amazon, Google, and Microsoft. The company is ambitiously targeting a market share of over 20%.

Source: Morgan Stanley, compiled by DIGITIMES, January 2025