Broadcom Is not Trying to Beat Nvidia, It May Be Building the Infrastructure Around It

I have been noticing deeper into Broadcom ($AVGO), and I think the most interesting part of the story is being missed. Broadcom does not need to beat Nvidia in GPUs. Its bigger opportunity may come from what happens after hyperscalers realise they cannot economically depend on Nvidia for every AI workload.

Broadcom’s recent growth already shows how quickly this is developing. Q3 revenue reached $29.6 billion, up 86% year over year, while AI semiconductor revenue jumped 221% to $16.7 billion. Management guided Q4 AI semiconductor revenue to around $21.7 billion. In other words, Broadcom’s AI semiconductor business has moved from roughly $8.4 billion in Q1 to an expected $21.7 billion in Q4. That is no longer ordinary semiconductor-cycle growth.

The reason is custom silicon. GPUs are incredibly powerful because they can handle many different workloads, but once companies like Meta, Google, OpenAI and other hyperscalers start running similar workloads across enormous AI clusters, flexibility becomes less important than cost, power efficiency and performance per workload. At that scale, designing your own accelerator can make economic sense. Broadcom sits directly in that transition by helping hyperscalers design and manufacture custom AI accelerators.

This is why I don’t think the correct Broadcom versus Nvidia debate is “Who wins?” Nvidia can continue dominating general-purpose AI compute while Broadcom benefits from the hyperscalers trying to reduce their dependence on expensive general-purpose GPUs for specific workloads. The bigger Nvidia becomes, the stronger the financial incentive for companies spending tens of billions on AI infrastructure to optimise part of that spending through custom silicon.

There is another layer that may be even more important: networking. A giant AI cluster is not useful if thousands of accelerators cannot communicate with each other efficiently. As AI models get larger, the bottleneck is increasingly moving beyond raw compute into memory bandwidth, networking, optical connectivity and power. Broadcom already has a major position in Ethernet switching, SerDes, optical connectivity, PCIe and other technologies required to connect enormous AI clusters. So Broadcom can potentially make money from both the accelerator and the infrastructure connecting those accelerators together.

That creates an interesting second-order AI trade. AI demand grows, hyperscalers buy enormous amounts of GPU capacity, infrastructure costs explode, companies increasingly develop specialised accelerators, Broadcom helps build those accelerators, and then Broadcom also provides much of the networking required to connect them. Broadcom therefore doesn’t necessarily need to own the entire AI compute market. It just needs AI infrastructure to become larger, more customised and more interconnected.

VMware adds another dimension that I think is still underappreciated. Broadcom is not only exposed to hyperscale AI infrastructure. VMware gives it a potential position inside enterprise and private AI as well. Banks, governments, healthcare organisations and large companies may not want every AI workload running in a public cloud. Many will want models and agents operating closer to sensitive corporate data, with stronger governance and security controls. If VMware becomes an important platform for private AI infrastructure, Broadcom could eventually participate across custom AI compute, AI networking and enterprise AI infrastructure.

The financial model also deserves attention. Broadcom’s customers are spending billions building data centres, buying power infrastructure and deploying massive AI clusters. Broadcom does not have to finance most of that physical infrastructure itself. It sells some of the highest-value intellectual property, silicon and networking technology going inside those systems. That gives Broadcom a potentially attractive position in the AI capex cycle without carrying the same infrastructure burden as the hyperscalers.

But the stock is far from risk-free. Expectations are already extremely high. When investors begin expecting triple-digit AI growth, simply producing strong numbers may not be enough. Customer concentration is also important because a relatively small number of hyperscalers can account for very large programmes. If one customer delays a custom accelerator, changes architecture or brings more development internally, the impact could be meaningful. Competition from Marvell and internal hyperscaler silicon teams will also intensify, while investors need to watch whether rapid AI-system growth changes Broadcom’s margin profile.

This is why I think the Broadcom thesis is more interesting than simply calling it an “AI chip stock.” Nvidia created and still dominates one of the most valuable compute markets in history. Broadcom may be positioning itself around what happens next: hyperscalers designing their own silicon, enormous clusters requiring increasingly sophisticated networking, and enterprises building private AI infrastructure.

So my concern is not that Broadcom replaces Nvidia. It is that AI infrastructure gradually becomes a mix of GPUs, custom accelerators, high-speed networking and private AI platforms, and Broadcom manages to capture value across several of those layers at the same time.

Concern for investors now is whether $AVGO is still being valued like a semiconductor company with strong AI exposure, or whether the market has already priced it as one of the core infrastructure platforms behind the next phase of the AI buildout.