Key takeaways
- Broadcom and Marvell are positioned to benefit from accelerating demand for custom AI chips and data center networking as major cloud providers pursue cost reductions.
- Catalyst: Industry expectations for rapid growth in custom AI chip shipments, alongside specific platform partnerships tied to hyperscaler AI infrastructure.
- Broadcom implication: Focus on AI ASIC design and its TPU-related role supports a growth narrative that the market is already rewarding.
- Marvell implication: A more modular ASIC approach and involvement across the TPU supply chain could strengthen its data center exposure, especially over the next few fiscal years.
Semiconductor investors are weighing two AI infrastructure beneficiaries—Broadcom and Marvell—as hyperscalers increasingly move toward custom silicon to improve performance per dollar. Both companies supply technology used in application-specific integrated circuits (ASICs) and data center connectivity, but the market’s expectations for their growth trajectories remain noticeably different.
What drove the focus on custom chips and networking
According to the article, the push toward custom AI chips is tied to a broader cost-and-efficiency race among cloud operators. While cloud providers have reported strong returns from earlier AI infrastructure investments, they continue to pursue further efficiency gains to improve economics.
The report highlighted that Alphabet has said the payback period for server investments built around its own custom AI chips can be cut from two years to one. That kind of accelerated return profile has helped spur hyperscalers to develop or source their own silicon rather than rely exclusively on general-purpose accelerators.
Custom chip demand is already forming a measurable market trend. According to Tom’s Hardware, custom AI chip shipments are projected to grow nearly 45% this year, and custom chips are expected to reach almost 28% of the data center chip mix in 2026.
How Broadcom fits into the AI ASIC build-out
Broadcom and Marvell are both described as key suppliers for ASIC development, with ASICs serving targeted workloads and typically offering improved power efficiency compared with more general programmable alternatives. The article also said ASICs are less flexible than programmable GPUs, but they perform narrowly defined functions efficiently and can lower operating costs over time.
Broadcom’s role, per the report, centers on AI ASIC design and high-value services that translate customer architectures into manufacturable chips. The piece cited Counterpoint Research’s expectation that Broadcom could maintain a 60% share of the AI ASIC market in 2026.
The article also pointed to Broadcom’s involvement with Alphabet’s Tensor Processing Units (TPUs), describing the TPUs as a primary growth driver. It further stated that Anthropic is expected to become Broadcom’s largest customer, based on an agreement in which Anthropic would buy Alphabet TPUs directly from Broadcom.
Beyond Alphabet, the article said Broadcom has contributed to custom chips for OpenAI and Meta Platforms, among others. It also referenced company guidance that Broadcom expects its ASIC-related revenue to double in each of the next two fiscal years and to reach $230 billion in fiscal 2028.
Where Marvell’s strategy diverges
Marvell is characterized in the report as offering a more modular, “à la carte” ASIC service model, with Amazon identified as its largest customer. The article also said Marvell previously helped Microsoft develop its Maia chip, reinforcing the idea that Marvell can support custom accelerator efforts across multiple hyperscalers.
At the same time, the report noted uncertainty around iteration-by-iteration leadership. It stated that Marvell is believed to have lost the lead position on future iterations of Amazon’s Trainium chips to AIchip, a Taiwanese semiconductor company.
Still, the article cited a significant win for Marvell tied to Alphabet: an agreement to provide a set of components within the TPU ecosystem, including AI inference accelerators, storage controllers, and network interface controllers. The report added that Marvell expects acceleration in its custom chip business in the second half of fiscal 2027 and into fiscal 2028, and it projected strong growth for its data center segment over that period.
Networking remains part of the AI infrastructure equation
In addition to custom silicon, both companies are presented as beneficiaries of data center networking demand. The article said Broadcom is a leader in AI data center Ethernet solutions, specifically citing its Tomahawk and Jericho chipsets. Marvell, meanwhile, is described as strong in connectivity, including optical interconnects.
According to the report, the optical interconnect market is gaining momentum as data center builders shift from copper wiring to fiber. It cited fiber’s advantages including bandwidth and lower energy costs, factors that typically become more important as network traffic scales alongside AI training and inference.
Market positioning and investor expectations
The article argued that the market’s valuation gap between the two names reflects differing expectations for growth. It stated that Marvell stock has outperformed Broadcom by a wide margin over the year-to-date period, while Broadcom’s gains have been more modest.
It also reported forward valuation differences using forward price-to-earnings (P/E) multiples. The piece said Broadcom trades at lower forward P/E levels than Marvell across fiscal 2027 through fiscal 2029 estimates, implying that Broadcom is priced for comparatively less “priced-in” growth and may have more room for upside if its TPU and ASIC momentum continues.
Bigger picture: what to watch next
For investors, the next datapoints will likely center on evidence of sustained custom-chip adoption at hyperscalers, the pace of deployment tied to TPU-related supply chains, and whether networking demand continues to track the expanding AI build-out. The report also points to the timing of expected acceleration—particularly for Marvell in the second half of fiscal 2027 and into fiscal 2028—so future quarterly updates and fiscal guidance will be key for validating those trajectories.
