What a beast Broadcom has become. Its meteoric march sees it firmly inside the top 10 publicly traded entities globally, custom silicon customers lining up out the door, and a top-tier prioritization resulting in deals with some of the highest-spending enterprise clients in the world.

At the center of its explosive top-line acceleration, though, is VMware. When news broke that the semiconductor giant was going to spend $69 billion on the cloud software vendor in late 2022, a few people scratched their heads at the idea. But for all the customer consternation, Broadcom has reaped the rewards, with revenue for its infrastructure software unit – which contains VMware – accounting for 32% of the company’s most recent quarterly revenues.

VMware quickly became Broadcom’s cash cow, with revenues that allowed its parent to court OpenAI, Anthropic, and Google Cloud to meet their growing compute demands.

But now Broadcom’s custom silicon units are self-sustaining behemoths in their own right, capable of fueling their own client expansions.

And so, heads turn once again to what’s next. If Broadcom really wants to challenge the lofty heights of Nvidia’s market cap, this editor believes the next big play is Arista Networks.

Hear me out

Nvidia’s moat isn’t the graphics processing unit (GPU). It’s not the CUDA software stack, either. It’s their tightly bundled, vertically integrated package. GPUs like Blackwell or Vera Rubin plus NVLink plus BlueField and Spectrum-X all tied together with software overlays aplenty.

If you look at Nvidia as an example, it bought Mellanox back in 2019 not to boost its marketing but to own the interconnect, thereby forcing customers to stay in its ecosystem and preventing them from mixing and matching.

Arista fits Broadcom in sort of the same way. Broadcom's networking silicon lines like Tomahawk already power the open Ethernet-based alternative to Nvidia. But what it doesn’t have currently is that software layer to turn its silicon into a fabric hyperscalers want to actually deploy at scale. And thus the beauty of Arista’s EOS and CloudVision lines.

The logic isn’t necessarily that networking is the next big thing in AI, so Broadcom should double down. It’s more an argument that if Broadcom wants to go from being a chip vendor to a vertically integrated fabric player, ala Nvidia, Arista is its Mellanox.

The other angle is that such a hypothetical would give Broadcom a direct route into the enterprise and campus networking markets. Nvidia only really plays the data center space.

Arista, meanwhile, is actively pushing to grow both its AI and campus segments. So Broadcom buying the company would give it a foothold in a market where Nvidia’s moat argument is ultimately irrelevant.

And the ultimate twist? Arista’s campus push is being built in part upon its acquisition of the SD-WAN platform VeloCloud from Broadcom as part of VMware-related divestitures. So rather than build out a campus-relevant asset itself, Broadcom sold it off and could then realistically buy it back further developed under Arista.

The final minutiae for why such a deal seems plausible: There has never been a better time regulatorily for mergers and acquisitions.

For all the alleged behind-the-scenes shenanigans, Hewlett Packard Enterprise’s (HPE) Juniper Networks acquisition opened the floodgates to a potential wave of tech industry consolidation.

In the weeks and months since, we’ve seen Nvidia, the king of AI compute hardware, admit it wants to buy Hugging Face, the open-source software and community layer everyone depends on. Not an apples-to-apples comparison, but one that’s a lot more abstract than two companies dominating a similar market coming together. And yet, it feels a lot more plausible than when it tried this last time with Arm.

So where could things all go wrong?

In the event such a hypothetical could ever come to light, the most obvious reasons for why a Broadcom bid for Arista would fail are often the most overlooked.

First up, the VMware deal that provided the initial spark for this brain fart? It led to some serious concerns being voiced repeatedly by industry groups about the deal. Add to that the various disgruntled customers upset about the various changes and cost rises.

Any future Broadcom acquisitions will likely face a fair bit more scrutiny as a result. Like Roz in "Monsters Inc.", the regulators would likely be watching … always watching.

The other is the simple fact of the potential cost.

VMware cost Broadcom $69 billion. At the time of writing, Arista’s market cap is a cool $233 billion. Not a move from Hock Tan’s usual playbook by any means. Any such move would represent a massive expense for a company that’s already juggling AI capex-fueled expectations.

And there’s the likelihood that any potential deal would risk upsetting some of Broadcom’s biggest customers.

Sure, it already supplies silicon to Arista, but it also sells silicon to everyone. Not sure how Cisco and the other white-box companies would feel when they find out their chief supplier now owns one of their biggest rivals. Like when Ben & Jerry's long-time largest supply chain partner, Dreyer's, owned Häagen-Dazs, these types of agreements never rest on solid ground.

You might say, well, Nvidia sells to every one of these vendors too. Correct. But converting neutral goodwill built up from being the neutral supplier into direct competition with its own customers – that’ll just force Cisco and others potentially even further away, maybe even into the arms of someone like Marvell. Hyperscalers might even push deeper into custom switch silicon … which could also benefit Broadcom … wait a minute … .

Ultimately, this is all just speculation and editorial hyperbole. And there are compelling arguments for Broadcom to go for firms like Celestica, DriveNets, or even Arrcus to fuel the software layer to its own networking empire.

But Broadcom has struck gold in scorched-earth M&A before. What’s stopping lightning from striking twice?

UPDATE: This op-ed has been updated to reflect that Nvidia has confirmed it intends to acquire Hugging Face.