Nvidia and Broadcom Shift from Chip Vendors to Data Center Architects

Nvidia and Broadcom are evolving from specialized chip designers into comprehensive data center architects, a shift that complicates the investment narrative for AI hardware. According to earnings data compiled by GN markets, both companies are moving beyond selling individual accelerators to controlling broader layers of the infrastructure stack. This transition means the competition is no longer just about which GPU or custom accelerator is faster, but which company can capture more of the value chain as cloud providers spend billions on both general-purpose and custom silicon.

The distinction between the two firms is blurring as Nvidia expands its CPU and software footprint while Broadcom deepens its reliance on custom silicon for major clients. Investors are now assessing who is better positioned to monetize the ongoing infrastructure buildout, with both companies leveraging their unique strengths to lock in long-term contracts with hyperscalers.

Nvidia Expands Beyond GPU Sales

Nvidia’s data center segment generated $89 billion in revenue during the second quarter, a 117% year-over-year increase. This growth is no longer driven solely by graphics processing units. The company’s Grace CPUs have already surpassed $5 billion in sales over the past year, and management projects that the upcoming Vera Rubin CPU platform will generate approximately $20 billion in revenue by the next quarter. These figures indicate a strategic pivot toward providing complete computing solutions rather than just accelerators.

The software layer is becoming a critical differentiator for Nvidia. The recent acquisition of Hugging Face for $12.9 billion signals an intent to dominate the model distribution ecosystem. By integrating CUDA, Nemotron, and other tools, Nvidia is positioning itself as the central hub for generative AI development. This move allows the company to capture value from the software side of the stack, reducing dependence on hardware sales alone and creating a more resilient revenue base.

Broadcom’s Custom Silicon Strategy

Broadcom’s growth is anchored in custom accelerators, which accounted for 73% of its $16.7 billion in AI semiconductor sales last quarter. The company’s strategy relies on designing specific chips for clients like Alphabet, Meta, and OpenAI, offering lower power consumption and cost efficiency compared to general-purpose GPUs. This approach allows Broadcom to secure deep, long-term relationships with a concentrated group of high-volume customers who require bespoke hardware for their specific model workloads.

Networking and infrastructure software are also key growth drivers for Broadcom. Infrastructure software revenue rose 29% year-over-year to $8.8 billion, a result of the transformative impact of the VMware acquisition. As hyperscalers build out their networks, Broadcom’s Ethernet switches and optics provide the necessary connectivity for these custom AI systems, creating a bundled offering that is difficult for competitors to replicate.

Hyperscaler Contracts Define Market Position

Nvidia’s business model is broadening to include a wider array of customers. The company expanded its relationship with Amazon Web Services to include two million additional GPUs through fiscal 2029. Additionally, Nvidia is co-funding a 12-gigawatt data center campus in Ohio with OpenAI and SoftBank Energy. The book of business from non-hyperscaler customers grew 138% year-over-year to $40 billion, demonstrating that Nvidia is successfully diversifying its revenue sources beyond the largest cloud providers.

Broadcom’s approach is more concentrated but equally significant in scale. Its custom chips are designed specifically for the requirements of its top clients, who are ordering by the gigawatt. This concentration creates a stable, predictable revenue stream but also increases dependency on a few key accounts. The choice between the two companies now hinges on whether investors prefer Nvidia’s diversified, full-stack platform or Broadcom’s specialized, high-volume custom silicon model.