So, according to the stories, Broadcom’s earnings demonstrated AI’s troubles, sending their stock, and the NASDAQ down. Right? Sort of, but not in the way it’s usually interpreted. Was this an indication of an AI disaster in the making? Was it because of software, meaning VMware? We do need to look at Broadcom here, to dig out whatever truth lies behind the disappointing results.
The company’s earnings release is the best place to start. The company reported 48% gains in revenue, with operating margins of a record 67% and EBITDA another record at 69% of revenue. All of this was above guidance, but a bit shy of some estimates. This was the technical factor that caused its stock to fall; when hedge funds think a stock is strong in the long term but has a blip, they’ll sells short to drive down the price, the buy in to get a better cost position on top of the short-sale proceeds. You can see a bit of this in the markets this week. On Monday, Broadcom and AI bellwether Nvidia were both up, and this morning (Tuesday) futures are up for both. If this was the end of the AI hype cycle, that’s not what you’d expect. The Street loves a bubble, and while quantum computing is likely being groomed for the role, it’s a tougher sell since the technology is hardly populist.
With regard to AI, Broadcom doesn’t make GPU competitors, but rather the connecting/networking components. Its website currently has an AI solutions segment that’s focused on this, featuring its new network switch chip (Tomahawk 5) and optical elements. Thus, when we hear about their AI, we’re hearing about AI networking. Their AI story, and guidance, didn’t match the heady Street hopes, though they exceeded their own guidance. This is one of the problems with a bubble. A company tries to be realistic, and the Street tries to push the story to justify the hype. So, there’s inevitably a shortfall versus expectations, a blip to be exploited. They can hit Broadcom and other AI players. OK, the bottom line here is exactly that, the bottom line—for hedge funds.
For the rest of us, what can we learn? Truth be told, not necessarily a lot in a direct sense, Broadcom’s Tomahawk is the go-to chip for white-box networking products, so not all of the sales of the chip are even related to AI, nor is all AI networking based on it. Giant Cisco, the largest player in the market, has its own Silicon One chips, though rivals Juniper and Arista do use some of the Tomahawk family. DriveNets, arguably the most credible alternative supplier for large-scale deployments, also uses Broadcom. All of these companies are pushing an AI story, but all of them were selling switches before AI, and enterprises tell me that the number of true enterprise AI network deployments hovers at the edge of statistical significance.
That’s the basis for the real lesson here, or at least its foundation point. We are still seeing most AI deployments made by the hyperscalers. Enterprises are at most kicking self-hosting tires in the AI space, and that is the most important AI truth to be had at this point.
Almost all of enterprise IT spending from the 1950s to the present has been justified by “core business” applications. The “core” nature of these applications means the data they collect and process is critical to the companies, which in turn means its subject to governance. How much of a tech spending wave could be driven by applications that didn’t require governed data? Why would we expect the same enterprises who rejected “move everything to the cloud” hopes, in large part because of data security risks, accept “move a lot of things to cloud AI”? Enterprises will, to fully exploit AI, host more and more of it on their own resources, and as they do they’ll expand the need for AI-capable data center networks, not only to connect the clusters but also for data access. Eventually, this has to be the dominant opportunity, but for now it’s developing only slowly. Why?
Reason One is the same hype problem that I’ve already noted as a factor in Broadcom’s earnings miss. Enterprises overall tell me that they’re uncertain about self-hosting AI since nearly everything being talked about in the AI world relates only to the cloud model. Which, of course, is because the hyperscalers have spent billions on AI data centers and need to show Wall Street that it wasn’t a stupid move. The media loves the story because cloud AI is approachable; you can (and almost surely do) use it in your regular activities, even as a consumer. You almost surely don’t pay for it, though. Citizen AI has won the battle for good ink, which means that there’s little available online to bring confidence and comfort to enterprise IT organizations looking to make a business case for AI applications that use their core data.
A very few companies are driving the right bus. IBM is the company most mentioned by enterprises who have self-hosting AI project in process or being planned. Most IBM accounts are industry giants, and obviously they make up a small percentage of the total enterprise population. Not only that, they often find that AI is only adding to “horizontalization” pressures they’ve been facing from the time they started using virtualization to build efficient data centers. Thus, they are less likely to need “new” network technology. They’re also more likely to be Cisco customers, who won’t be using Tomahawk chips in their gear.
Broadcom needs self-hosting of AI to go mainstream for its “AI chips” to experience explosive growth. We’re seeing growth today, but driven by a small number of giants. Real AI success depends on democratizing hosting, and we’re not really addressing that. Where Broadcom is failing itself, IMHO, is there.
VMware is perhaps the dominant virtual-data-center technology, and Broadcom bought it. Some on the Street analyst side have been critical of the minimal growth they’ve seen in VMware revenue, but truth be told, that’s not the real problem Broadcom has with data center software. Giant data centers don’t spring up like mushrooms, and AI is not going to justify building new ones. Even hyperscalers are expanding their cloud data centers more than creating pure AI ones. What that means is that data center virtualization is a bit like IT overall, something that’s fallen into sustaining mode because new applications that could tap new benefits to justify new spending have not been emerging. One question is whether AI could change that. Another is whether Broadcom could drive it.
VMware has a larger account footprint than IBM’s strategic accounts represents. Could the sales teams then play, for VMware customers, the same role as IBM’s strategic account teams play? Why not? Will they? A lot is likely to depend on it, and enterprises so far say that VMware isn’t even trying to play the role of AI strategist.
The mainstream, as opposed to the tech and financial, media has not been shy raising questions about AI. Here’s an example from Axios, and one that makes a troubling connection to 2000’s dot-com market problem. As that piece points out, the big question now will be whether (to quote a source in the article) “Recent earnings reactions suggest that even outstanding growth isn’t always enough when expectations are stretched — a classic ‘priced for perfection’ dynamic.” There is no question that while the Street still loves the AI bubble, it’s wary about the bursting. The hype will end, as it always does, and for Broadcom the question is whether they’re prepared to step up and try to make that a PR event more than an economic one.
