How Is Broadcom Challenging Nvidia?

Broadcom this week launched Thor Ultra, a new networking chip designed to connect the massive-scale AI data centers of the future. The announcement came just one day after the company revealed a blockbuster deal to build 10 gigawatts worth of custom AI accelerator chips for OpenAI. Taken together, these moves represent a coordinated, two-front assault on Nvidia's dominance, signaling a major new phase in the war to build the infrastructure for artificial intelligence.

This isn't just about a new competitor entering the fray; it's a battle over the fundamental philosophy of how AI data centers should be built. While Nvidia has championed a closed, vertically integrated "all-in-one" solution, Broadcom is making a massive bet on an open, "unbundled" future, and it's using its deep expertise in both custom silicon and networking to make its case.

What is Nvidia's 'walled garden' strategy?

Nvidia's meteoric rise has been built on more than just powerful GPUs. The company offers a tightly integrated, proprietary system where its processors are connected by its own high-speed networking technology, like NVLink and InfiniBand. This "walled garden" approach delivers incredible performance but also creates significant vendor lock-in. Customers who buy into Nvidia's ecosystem find it difficult to mix and match components from other suppliers.

This strategy has been immensely profitable, but it has left the industry's biggest players—the cloud providers and AI labs—uneasy about their deep dependence on a single, all-powerful vendor.

How is Broadcom challenging the GPU monopoly?

On the first front, Broadcom is attacking Nvidia's core business: the AI accelerator chip. But it's not trying to build a direct GPU competitor. Instead, it is positioning itself as the premier design partner for tech giants that want to create their own custom chips, also known as ASICs (Application-Specific Integrated Circuits).

This strategy has already proven highly lucrative. Broadcom has generated billions by helping Google design multiple generations of its Tensor Processing Units (TPUs). The new deal to build custom chips for OpenAI—which comes after Broadcom announced an unnamed $10 billion customer in September—is a massive validation of this model. It offers the world's leading AI lab a path to develop bespoke hardware that is optimized for its specific needs, providing a powerful alternative to simply buying off-the-shelf from Nvidia.

Why is networking the second front?

A data center is more than just a collection of powerful chips; it needs a nervous system to connect them all. As Broadcom SVP Ram Velaga noted, "I'm not surprised that anybody who's in the GPU business wants to make sure that they are participating in the networking." This is the second front of the war.

Nvidia has pushed its proprietary networking solutions as the best way to connect its GPUs. Broadcom's new Thor Ultra chip is a direct counterattack, championing the open standard of Ethernet. Thor Ultra is the industry's first 800G network interface card fully compliant with the Ultra Ethernet Consortium, an industry group dedicated to creating an open, interoperable standard for AI networking. By offering a high-performance, open alternative, Broadcom is giving data center builders a way to break free from Nvidia's networking lock-in.

What is Broadcom's ultimate playbook?

Broadcom's two-front war is a bet on a different future for AI infrastructure. Instead of buying a pre-packaged, closed system from Nvidia, customers can pursue a more modular, "best-of-breed" approach enabled by Broadcom. They can design a custom accelerator with Broadcom's help and connect tens of thousands of them together using Broadcom's open-standard networking chips.

This "unbundling" offers customers more choice, greater control over their technology stack, and a hedge against being locked into a single vendor's ecosystem. With AI-related revenue already hitting 12.2 billion in fiscal 2024 and a target market of up to 90 billion by 2027, Broadcom is signaling that the battle for the AI data center is far from over. It's evolving into a fundamental choice between a closed, integrated system and an open, modular future.

The Reference Shelf

• Broadcom to launch new networking chip, as battle with Nvidia intensifies (Reuters)
• Broadcom Touts Ultrafast Networking Chip For AI Data Centers (Investor's Business Daily)