There have been multiple reports over the past few weeks on Nvidia, potentially buying ARM from its parent company Softbank. (here, here, and here)

ARM generates a majority of its revenue from licensing processor core designs to every smartphone processor designer, like Qualcomm and Apple. One can argue that the value of ARM is that it is a neutral IP provider to everyone. Hermann Hauser, the co-founder of ARM, expressed that idea in a recent interview: "One of the ARM business model's fundamental assumptions is that it can sell to everyone. If it becomes part of Nvidia, most of the licensees are competitors of Nvidia and will, of course, then look for an alternative to ARM." While Nvidia is an ARM licensee, it mostly uses ARM core design in its Tegra SoC for the automotive and game console, and no business in the smartphone market. However, if one particular semiconductor designer owns ARM, even it's not directly competing with most of the customers, it still introduces a significant conflict of interest.

Softbank bought ARM in 2016 for $32 billion and has increased spending and headcounts by a meaningful amount since the purchase. In the most recent fiscal year, ARM generated $1.9 billion in revenue and ~$275 million of EBITDA. Therefore, it is also highly unlikely for Nvidia to cut spending and increase profitability in such an acquisition. If Nvidia were to pay anywhere close to the $32 billion initial purchase price, it would be at over 100x EV/EBITDA. So with all the complications and potentially a very high price tag, what could be Nvidia's game plan here?

ARM Net Sales in Recent Fiscal Years

Close

ARM Net Sales in Recent Fiscal Years

We think one possibility would be to increase ARM-based server CPU adoption in the data center and supercomputers. Intel has dominated the data center computing market for the past decade, but it is turning into a three-horse race over the last five years. Nvidia successfully penetrated the market with GPU accelerator for machine learning, and AMD has gained meaningful share from Intel in the traditional x86 CPU market. Comparing to the two competitors, one clear disadvantage Nvidia has is the lack of CPU. AMD already has a decent product portfolio covering both GPU and CPU, and Intel will soon have its discrete GPU for the data center later this year, assuming no further delay. Having both elements under the same roof could meaningfully help system performance through tricks like having both CPU and GPU access the same memory to reduce latency and boost performance. We believe that is one of the critical reasons AMD was able to secure big wins for the exascale supercomputers, including El Capitan and Frontier.

But here comes the question, why not just license ARM server design? You do not need to buy a farm to have fresh milk every day. We struggled with this question for a few days. Then we realized Nvidia has already partnered with ARM, a few ARM server licensees, including Marvell, Ampere, and Fujitsu for a GPU accelerated ARM-based server reference design. In this vein, we started picturing this somewhat crazy scenario, a much more tightly collaborated Nvidia GPU and ARM CPU design, with CPU licensed to partners, and put them all into a system like DGX. Instead of fighting alone, Nvidia will be competing alongside a group of ARM server licensees, collecting a licensing fee, and strengthening their position in data center GPU when the competition is increasing. It also helps that one of the more successful ARM server licensees is also a significant customer for Nvidia called Amazon. Or even crazier, license the whole thing.

Of course, this is all just speculation and nothing more than a thought experiment trying to make sense of all the news. In the meantime, we will be eagerly waiting for the actual game plan.