Time to shine for custom AI hardware

What happened

Recent financial results from both Marvell and Broadcom showed similar dynamics: large positive earnings surprises for the AI-related business offsetting "legacy" products and upgraded business perspectives. Both stocks experienced substantial positive reactions.

Impact on our Investment Case

Life beyond training

With the recent AI boom, most of the market focused (rightly so) on Nvidia, due to the simple fact that the financial impact was already there and too big to ignore. Although we strongly benefited from this (the stock accounted for ~20% of the performance of the strategy last year), we recently reiterated our early-year view about the importance of inference: while training is of course an important part of the AI supply chain, it will ultimately be dwarfed by inference, at least when it comes to the volume of workloads.

Inference is the process of "using" AI, in opposition to training, which is setting up the model. GPUs provided by Nvidia can do both, but are primarily targeting the complex training tasks. Running inference workloads is perfectly doable but does not scale well due to being too costly - an analogy would be taking a race car to buy groceries. The logical consequence would be the proliferation of smaller, specialized chips for inference workloads, combined with networking solutions to synchronize the large number of chips required to enable large-scale use of AI.

Fasten your seatbelt

A balanced hardware allocation should encompass both training and inference players. For the latter, Broadcom and Marvell are the undisputed leaders, providing custom designs for the likes of Alphabet (already the 6th generation of TPUs) and Amazon. And indeed, this has already resulted in both stocks contributing substantially to this year's performance for our strategy. But the latest financial results show that this trend is just beginning and may surpass our initial expectations.

Indeed, it appears everyone on the Street clearly underestimated the market potential and its expansion velocity. Marvell, the more conservative of the two in its financial communication, is almost certain to see its AI-related targets being largely exceeded. The beat margin, according to some analysts, could reach up to 100%. On the other end, Broadcom is more communicative, and in this regard, its addressable market estimate adjustments are simply defying gravity: from a $15-20bn market opportunity in FY24 to $60-90bn in FY27, i.e., a ~62% CAGR over the next three years. Even more impressive is the fact that these estimates only include existing customers and could, therefore, extend beyond such numbers.

Our Takeaway

There is a life beyond Nvidia, and this future is already there. Custom AI chips already deliver substantial results and represent a massive market opportunity, which latest results from market leaders Broadcom and Marvell confirmed. We have dedicated a substantial allocation to those since early 2021, which has reached a total of ~10% on 12 December, and we expect their outperformance to keep delivering substantial contributions in the quarters to come.