Key Takeaways:
The AI chip market is splitting into two tiers — Nvidia owns training, but Broadcom and Marvell are capturing the $1 trillion inference buildout with custom silicon that costs half as much.
The shift from buying Nvidia Corp.'s fastest GPUs to lowering the cost of operating AI at hyperscale is creating a growing opportunity for Broadcom Inc. and Marvell Technology Inc., whose custom chips cost roughly half as much as Nvidia's most advanced racks.
"Hyperscalers continue to expand their investments in proprietary silicon to improve total cost of ownership and reduce dependence on merchant GPU suppliers," Morgan Stanley analysts wrote in a recent note.
Building one gigawatt of AI infrastructure using ASIC racks designed by Broadcom or Marvell costs between $6 billion and $11 billion, according to Milk Road AI estimates. By comparison, racks built around Nvidia's GB300 processors cost roughly $19 billion, and the figure jumps to $25 billion under Nvidia's next-generation Vera Rubin architecture. Google, Amazon, Meta Platforms Inc. and Microsoft Corp. are all pursuing custom silicon alongside Nvidia GPUs because every percentage point of efficiency matters when spending tens of billions annually.
The four hyperscalers are forecast to spend over $1 trillion on AI infrastructure next year, adding 19.5 gigawatts of incremental compute capacity — nearly triple the 6.7 GW they added in 2025. Google alone will add 6.8 GW, more than the hyperscalers combined two years ago. Broadcom, which partners with Google on its Tensor Processing Units and with Meta, and Marvell, which works with Amazon on its Trainium chips, benefit whether customers build proprietary processors or buy more off-the-shelf hardware.
The Economics of Custom Silicon vs. Merchant GPUs
The appeal of application-specific integrated circuits comes down to dollars. Nvidia's GPUs remain the gold standard for training frontier models, and its CUDA software platform provides a competitive moat that rivals have struggled to crack. But once models are deployed, inference workloads — answering prompts, generating images and powering AI applications — don't always require Nvidia's most powerful processors. They reward lower costs and higher efficiency instead.
That is where Broadcom and Marvell enter the picture. Rather than selling chips under their own brands, they help cloud providers design silicon optimized for their own software stacks and infrastructure. Google continues developing TPUs with Broadcom while still purchasing enormous volumes of Nvidia GPUs. Amazon follows a similar dual-track strategy with Trainium alongside Nvidia deployments. That diversification gives cloud providers leverage in negotiations while matching the right chip to the right workload.
Nvidia's Response and the Investment Case
Nvidia is hardly standing still. The company continues expanding beyond GPUs into networking, rack-scale systems and software, and has opened technologies such as NVLink Fusion to custom silicon partners — including Marvell itself. Still, every new custom AI accelerator designed by a hyperscaler creates another opportunity for Broadcom or Marvell. They are standing directly in the path of a massive spending wave.
For investors, the question is where the fastest incremental growth will occur. Nvidia commands the premium end of AI computing, particularly for training, and its data center revenue remains dominant. But custom silicon is becoming an essential part of every hyperscaler's long-term strategy. Broadcom appears best positioned today thanks to its deep relationships with Google and Meta, while Marvell continues strengthening its foothold with Amazon and other large customers. The AI infrastructure market is expanding so rapidly that multiple winners can emerge.
This article is for informational purposes only and does not constitute investment advice.
