Nvidia held its analyst day yesterday, and it was a pretty stunning event. Nvidia is one of the best positioned semis companies today, and this was on full display. One of their first slides was titled “Giant Wave of New Products” (see photo above), and that pretty much sums up the theme of the event. Nvidia now offers GPUs, AI accelerators (DPUs), their newly-launched Arm-based CPUs, a range of networking chips and the ability to design custom ASICs for others. Their addressable market is now essentially all of digital semis, except for memory – so a TAM of a few hundred billion dollars (they say it is $1 Trillion). We encourage reading through their presentation slides, they have a lot going on.
We think it is worth thinking through has this happened. Not so long ago, Nvidia was a “niche” vendor of GPUs – loved by gamers and bitcoin miners, viewed as a frivolous derivative product by much of the rest of the industry. Now they are the largest semis company by market cap and seem to be firing on all cylinders.
Obviously, a lot of the credit goes to Founder & CEO Jensen Huang and the management team. They have managed well and delivered a fairly grueling product roadmap consistently for a long time. But there is also an important trend in computing, a tide that they have ridden to get here.
At heart, the nature of computing is changing in some fairly important ways. The change is one of those things that everyone is aware of, but it is so large and so gradual that we tend to take it for granted. We do not even have a good name for it. Some attribute it to a “Post Moore’s Law” world. Others have called the de-throning of the CPU from its perch on top of the computing hierarchy. It could also credibly be called Heterogenous Computing, which is how we tend to think of it. Put simply, large computer systems used to be built around the CPU, since that was the chip doing the bulk of the work. Massive, $1 billion data centers were essentially designed this way for years. That is no longer true.
CPUs are still important, but they are no longer the only crucial chips in data centers. This all started around 2016 when Google launched their TPU. They did not single handedly start the trend, others were moving down this path already, but TPU really crystalized the shift. The big builders of large computing systems were starting to shift to AI workloads, and these required new kinds of chips. CPUs are workhorses, but AI workloads can be run much more efficiently on other kinds of chips. Google had its TPUs, but everyone else used GPUs – aka Nvidia’s core product. Since then, we have seen an explosion of special-purpose AI accelerator chips (also called DPUs, NPUs or many other names) in the data center. Nvidia dominated this market, in large part because of its CUDA software libraries. CUDA deserves a post all on its own, but our point here is that data centers started to change and Nvidia was able to dominate that shift.
Compounding this, data centers continue to evolve. As Moore’s Law slows, and CPUs are not advancing the way they used to, big chip customers have turned to a more diverse set of semis in their data centers. They are experimenting (and then some) with new types of CPUs, bringing in all those AI accelerators, trying all kinds of new networking chips, and designing their own special purpose chips to perform tasks specific to their software needs. All to maximize the capabilities of those data centers. And here is Nvidia delivering all of the above.
Nvidia has gone from leading vendor for a niche segment, to leading vendor for the entire data center. It is hard to think of any other chip company that can compare. Intel, in theory, can offer all these same products but falls short in many of its offerings, to say nothing of its existential crisis. Broadcom is still dominant in networking, but is investing approximately $0 in everything else, as it transitions into a software private equity firm. Qualcomm has very little in the data center, and nothing in networking. Marvell is a potential rival in many of these markets, but is an underdog and lacks GPUs. AMD is probably the closest peer, but lacks networking capabilities and is far behind Nvidia in GPUs. The one thing Nvidia lacks is FPGAs. In hindsight, they would have been better off skipping Arm and instead bidding on Xilinx which got sold while Nvidia was bogged down with the Arm acquisition. But even here, FPGAs are really not a crucial data center component. (As a side note: this should also make clear why all these companies were so opposed to Nvidia buying Arm. Nvidia is now competing with everyone.)
And we are just looking at the data center. Much of this holds true for some other important markets like robotics and automotive.
There is no question that Huang has earned his place in the Semiconductor Hall of Fame and deserves ample credit for his achievements. But the underlying force that made Nvidia’s rise possible shows how even large, static markets can change important ways and speaks to the possibility of new futures for the industry.