Reports indicate that Nvidia is in the process of forming a specialized business unit to offer its intellectual property (IP) and design services to industry giants such as AWS, Microsoft, and Meta.
This corporate maneuver comes in response to the trend of numerous cloud providers and hyperscalers developing their own custom alternatives to Nvidia’s GPUs for tasks like AI and other accelerated workloads, according to a Reuters report citing sources familiar with the matter.
Among the early adopters of custom silicon was Amazon Web Services, which introduced its Graviton GPUs over five years ago and has since expanded its offerings to include smartNICs and AI accelerators. Similarly, Google’s tensor processing units (TPUs), introduced in 2017, provide an alternative to GPUs for AI training and inference workloads.
More recently, Microsoft and Meta, two significant consumers of Nvidia GPUs for generative AI, have begun deploying their own custom silicon. Meta unveiled its latest inference chips, slated for deployment across its datacenters to power deep learning recommender models, while Microsoft introduced its Maia 100 AI accelerators, designed for large language model training and inferencing.
While these chips are customized for specific cloud provider workloads, they often incorporate intellectual property from companies like Marvell or Broadcom. Nvidia, with its extensive portfolio of IP spanning parallel processing, networking, and interconnect fabrics, aims to emulate Broadcom’s strategy of licensing out its technologies.
According to reports, Nvidia executives see an opportunity to provide its IP to companies like Amazon, Meta, Microsoft, Google, and OpenAI for developing custom chips based on Nvidia’s designs. Nvidia has also extended similar offers to telecom, automotive, and video game customers.
Particularly noteworthy is Nvidia’s pursuit of Google, especially in light of rumors suggesting Google’s potential departure from Broadcom technologies.
Nvidia has been approached for comment regarding its IP licensing plans, with updates pending.
While several cloud providers are exploring custom silicon solutions, it appears unlikely that they will fully replace Nvidia, AMD, or Intel hardware in the near future. Despite efforts to make custom silicon available to the public—such as Google’s announcement of the performance-tuned TPUv5 AI accelerator—GPUs remain dominant in generative AI tasks.
Meta’s rollout of custom inference chips is significant, yet GPUs will likely retain their importance for various workloads. Meta’s deployment of H100 chips, aiming for 600,000 units by year-end, underscores its commitment to artificial general intelligence research.
Notably, other companies, like Microsoft, are also combining custom silicon with large-scale GPU deployments. Microsoft’s plans include leveraging AMD’s MI300X alongside Nvidia’s H100s for generative AI services. sx sx sx sx sx sx sx sx
AWS, meanwhile, announced a substantial deployment of Nvidia Grace-Hopper super chips alongside its Graviton CPUs and Trainium AI accelerators last fall.