Also a key player in AI "chip sellers"! Last week, Broadcom conquered Wall Street

Authors: Gao Zhimou, Bu Shuqing

Source: Hard AI

In the "gold rush" sparked by large AI models, NVIDIA has always been the "seller of shovels" favored by capital. So, who will be the next NVIDIA?

Just as AMD and Intel are gearing up and eager to try, Broadcom, standing aside with hands crossed in front of its chest, had a faint smile on its lips that was not easily noticeable:

You guys are the sellers of shovels for AI, I will be the seller of shovels for you!

Last Wednesday, Broadcom held an investor conference with the theme "Enabling AI Infrastructure". At the meeting, Broadcom not only released a variety of AI infrastructure innovations such as switches, PCIe, copper connections, but also announced its third largest custom chip customer to a group of analysts—these news triggered a capital frenzy on Wall Street.

Subsequently, Broadcom's stock price continued to soar. In the past week, Broadcom's stock price surged nearly 8%; over the past year, Broadcom's stock price skyrocketed by 115.92%.

Overall, the information disclosed at Broadcom's investment conference last Wednesday was substantial, with many points worth delving into. Here are some highlights we have organized for you, enjoy~

You guys are the sellers of shovels for AI, I will be the seller of shovels for you!

Training large AI models requires a large number of GPUs, NPUs, LPUs, and other processors produced by companies such as NVIDIA and AMD—or XPUs, as well as network chips such as Ethernet switches and routers to connect these high-performance chips and transfer data between them—this is precisely Broadcom's expertise.

The following PPT is easier to understand:

This is an "XPU" infrastructure necessary to support AI workloads. The right side of the image shows Broadcom's technology, while the corresponding numbers on the left side reflect their positions in system packaging.

Next, we see a configuration diagram that expands upwards and outwards to numerous clusters. Through optical devices marked in red and yellow, Broadcom's technology achieves connections from the backbone network, top-of-rack switches to NICs, PCIe innovations, and other parts of the system, covering a market far beyond the scope of GPUs and CPUs.

During this conference, all of Broadcom's PPTs revolved around a key concept that the company has always emphasized—open, scalable, and energy-efficient. Charlie Kawwas, President of Broadcom's Semiconductor Solutions Group, stated: For providers responding to the growing demand for generative AI clusters, the key to success will be a network-centric platform based on open solutions and scalable at the lowest power consumption.

Our innovations have expanded our leading position in custom AI accelerators, Ethernet, PCI Express, and optical interconnect product portfolios. They are built on world-class foundational technologies such as SerDes and DSP, providing the best custom XPU and commercial network solutions to support AI infrastructure.

Furthermore, Kawwas emphasized that Broadcom has established good cooperation with a wide range of customers, including consumer internet companies, device manufacturers like Apple, Dell, HP, and hyperscale computing enterprises, with Nvidia being one of the most critical companies in the AI field. Kawwas stated:

"Broadcom has a good relationship with Nvidia, and Nvidia is its fastest-growing customer."

It is worth mentioning that during the meeting, Jas Tremlay, General Manager of Broadcom's Data Center Solutions Group, shared an interesting case about Nvidia Blackwell: while Blackwell's southbound connection uses proprietary NVLink, its northbound connection is open PCIe. This indicates that even Nvidia, which heavily relies on proprietary technology, has to consider the importance of openness in building its architecture. Tremblay said:

"Blackwell's southbound connection uses proprietary NVLink, while its northbound connection adopts PCIe."

After the meeting, TD Cowen analyst Matthew Ramsay upgraded Broadcom's stock rating to "outperform the market" and raised its target price from $1400 to $1500.

"Indeed, our rating has been on the wrong side for a while, as the stock has outperformed the broader SOXX index by about 50% over the past two years. Despite recent developments, we are confident in the sustainability and diversity of growth, as well as the potential upside of forward-looking consensus forecasts driven by AI computing/network and software growth/profit."

Research firm Baird's Tristan Gerra stated that Broadcom is a "world-class" competitor in AI networks, but the current valuation does not reflect this. He also gave Broadcom's stock an "outperform the market" rating and set a target price of $1500.

The Mysterious "Third Largest Buyer": ByteDance?

Of particular note, Broadcom disclosed its third largest custom chip customer, which has attracted great attention. While the company did not explicitly reveal who it is, many analysts believe that this "mysterious buyer" is likely ByteDance Citigroup's analyst team led by Christopher Danely believes that the third largest mysterious customer may be ByteDance. Broadcom has not responded to this yet. On Thursday, Citigroup reiterated its buy rating on Broadcom with a target price of $1560.

Broadcom expects that by 2024, 35% of its chip revenue will come from AI, an increase of 10% from previous forecasts, far exceeding the less than 5% in 2021:

By 2024, 35% of Broadcom's chip revenue will come from AI.

Undoubtedly, after the investor conference last week, Broadcom, which caught the wave of AI, seems to have become the "new darling of Wall Street." Against this backdrop, investors and analysts seem "unconcerned" about the less than ideal traditional networking and storage business revenue in Broadcom's financial report.

After all, this year's Wall Street has a consensus: where there is AI, there is everything else, and the rest can wait for now