This CNBC article just snuck in on the last day of the year, and it took two authors to write it, but it is my candidate for the worst semiconductor industry article for 2024.
It's so bad it took me three attempts at reading it to get all the way through it.
The article implies the chips Broadcom does the analog designs and foundry management stages for the cloud computing companies are Broadcom designs, which is nonsense. Not that the analog work isn't important, it is critical, but the wording of this paragraph implies these are Broadcom products. They aren't. Broadcom is a subcontractor, albeit a very important one that clearly adds a lot of value.
Not only is this description technically incorrect, but it shows the authors have no idea what they're talking about. And then they quote some industry analysts who don't know what they're talking about either.
Ridiculous. You can go to Google Cloud's webpages and see that anyone can lease time on a Google Cloud TPU system:
And you can see from the broadcom.com web page that Broadcom lists exactly zero AI products. No XPUs at all. What these authors don't comprehend is that what Broadcom is doing is subcontracted work in selected phases of chip development, not developing chips. It is obviously a very good business, and that makes sense, because if you're really doing very few chip designs per year, like the cloud vendors, you're probably not going to keep your analog and manufacturing interface people busy enough to justify the cost of in-house groups. Is this business relationship really that difficult to explain?
But even an article this bad has one quote that was worth reading farther down for:
Seriously, Ms. Holthaus? Admitting Falcon Shores won't be wonderful, but you're spending a boatload of R&D funds on it? And you expect customers to want to buy it? Amazing. I couldn't make this stuff up.
It's so bad it took me three attempts at reading it to get all the way through it.
The article implies the chips Broadcom does the analog designs and foundry management stages for the cloud computing companies are Broadcom designs, which is nonsense. Not that the analog work isn't important, it is critical, but the wording of this paragraph implies these are Broadcom products. They aren't. Broadcom is a subcontractor, albeit a very important one that clearly adds a lot of value.
What’s exciting Wall Street is Broadcom’s role working with cloud providers to build custom chips for AI. The company’s XPUs are generally simpler and less expensive to operate than Nvidia’s GPUs, and they’re designed to run specific AI programs efficiently.
Not only is this description technically incorrect, but it shows the authors have no idea what they're talking about. And then they quote some industry analysts who don't know what they're talking about either.
Cloud vendors and other large internet companies are spending billions of dollars a year on Nvidia’s GPUs so they can build their own models and run AI workloads for customers. Broadcom’s success with custom chips is setting up an AI spending showdown with Nvidia, as hyperscale cloud companies look to differentiate their products and services from their rivals.
Broadcom’s chips aren’t for everyone, as only a handful of companies can afford to design and build their own custom processors.
“You have to be a Google, you have to be a Meta, you have to be a Microsoft or an Oracle to be able to use those chips,” Piper Sandler analyst Harsh Kumar told CNBC’s “Squawk on the Street” on Dec. 13, a day after Broadcom’s earnings. “These chips are not meant for everybody.”
Ridiculous. You can go to Google Cloud's webpages and see that anyone can lease time on a Google Cloud TPU system:
And you can see from the broadcom.com web page that Broadcom lists exactly zero AI products. No XPUs at all. What these authors don't comprehend is that what Broadcom is doing is subcontracted work in selected phases of chip development, not developing chips. It is obviously a very good business, and that makes sense, because if you're really doing very few chip designs per year, like the cloud vendors, you're probably not going to keep your analog and manufacturing interface people busy enough to justify the cost of in-house groups. Is this business relationship really that difficult to explain?
But even an article this bad has one quote that was worth reading farther down for:
Late next year, Intel will release a new AI chip that it codenamed Falcon Shores. It won’t be built on Gaudi 3 architecture, and will instead be a GPU.
“Is it going to be wonderful? No, but it is a good first step in getting the platform done,” Intel interim co-CEO Michelle Holthaus said at a financial conference held by Barclays on Dec. 12.
Seriously, Ms. Holthaus? Admitting Falcon Shores won't be wonderful, but you're spending a boatload of R&D funds on it? And you expect customers to want to buy it? Amazing. I couldn't make this stuff up.