Data center silicon is the an important topic and the focus of much industry attention. But as we noted in our previous post entering this market is very difficult from a sales perspective. So here we want to put some numbers behind this.
Sizing the data center semiconductor market is a bit tricky. If you sum up AMD’s, Intel’s and Nvidia’s Data Center segment, the total for 2022 was about $41 billion, but all these figures include a lot of products including CPUs, GPUs, accelerators, networking, FPGAs and probably some memory too. Then there is the question of homegrown silicon from AWS, Google and the rest. None of these companies have disclosed much on the size or cost of these devices. For our purposes here, let’s call it a $50 billion annual market, which is massive.
A key wrinkle in this market is that 10ish companies consume at least 50% of these products. As we noted previously this makes it very hard for new entrants to break into the space. The Valley is littered with the corpses of start-ups who followed the siren song of market size only to have their backs broken by their only customer who walked away just before the funding round closed. These are big numbers and even a small piece of it holds big appeal for any company, so there are likely to be many more attempts to navigate these rocky shoals.
Next, we need to break this down further into product categories, and again there is not a lot of clear market data. We found one market report that broke the market down as CPU 40%, GPU 20%, FPGA as 30% and ASICs as 10%. In our view, this overstates FPGAs, understates GPUs and overlooks networking entirely. So we would revise that to CPU 40%, GPUs at 25%, FPGAs as 15% ASICs and networking at 10% each. This is get out the ruler and measure unlabeled graphs, duct tape, napkin math, but as the bankers say it is directionally close.
We think this breakdown is going to prove critical for the long-term trend in the segment. In particular, the share for GPUs and ASICs is set to grow strongly in coming years. Nvidia is the best example of that – Data Center is now their largest segment, admittedly boosted a bit by their networking products after their acquisition of Mellanox. Regardless, with the surging growth of AI and Large Language Models (LLMs), the GPU share of data center spend is poised to accelerate.
So let’s assume GPUs go from 25% share of wallet today to 35% by 2026, ASICs go from 10% today to 20% over that same period, FPGAs decline to 10% and networking stays steady at 10%, and then CPU is the plug, whatever is left over, which by 2026 is 25%. If we then assume the market grows by 12.5% a year we get the following results.
A few important things to note. The growth of ASICs and GPU obviously stand out starkly. This is basically the reason that Nvidia trades at a multiple of roughly infinity. Secondly, if the market only grows at 10%, then by 2026 the CPU revenue figure shrinks. This is basically the reason that Intel trades like it is at risk of going out of business, even if it can fix its manufacturing process it then has to contend with basic stasis in its core CPU market.
Any piece of this market is attractive, but for new entrants the competitive and customer dynamics are painful. To launch a new CPU into this market likely requires a minimum $500 million investment and more likely a cool $1 billion. This is a big investment even for a big company (and their investment is to be 2X-3X bigger). For a start-up to spend a $1 billion they have to capture capture ~5% share of one of the segments above to reach sufficient scale to deliver a meaningful exit for their investors. Possible, but not easy.
Photo by Joshua Hoehne on Unsplash