Recently, it seems to have become common wisdom that every tech company is going to build its own semiconductors for their hardware needs. Companies seem to be building chips for their own phones, data centers and networks. We have heard some analysts claim that the only way to compete in today’s hardware market is to have your own chip.
However, if we dig a little deeper, the reality is much more complex and far less revolutionary than the headlines imply. Admittedly, in many ways it has gotten far easier to design custom silicon than it has ever been before, but this is really only an option for a very small group of companies. Less than ten have the resources to really pull it off. For everyone else, the merchant market for chips is still the best, and likely only, path to market.
First off. If you want to design your own chip, you need a team of designers. Depending on the size of the chip you have in mind and the scope of its function, this team alone can cost $20 million to $50 million a year and up. And that is just the design team, actually getting a chip through a fab, or manufacturer, requires a team that can cost at least as much, to see nothing of the costs paid to the fab. Only a few companies can afford both the designers and the operations team required. A few others appear to be doing the design themselves, but then working with established chip companies for the operations side.
Having your own chip can be a competitive advantage, but it is important to understand the scope of that advantage. The whole point of having a chip is to accelerate your software. The closer that software is to the 1’s and 0’s of a chip’s physical layer, the better the performance of that software. For most software companies, there are already so many layers between the code and the electrical signals of the chips, that there are likely some far easier ways to accelerate the code. We are not saying these paths are ‘easy’, but they are likely far easier than hiring a semiconductor design team.
Also, it is important to remember that merchant chip companies can amortize their costs across many customers. For most customers that means the merchant solution is going to be far cheaper than a custom chip built in-house. This is what happened in the phone market ten years ago, and the PC market 20 years ago. Only companies with 50% market share in their market (or 50% of the profit pool) can really hope to compete with their own custom chips. This does not mean no one should or will design their own chips, but it does mean companies need to think very carefully about where their competitive advantage truly rests and be realistic about the costs of building a chip.
That being said, something has clearly shifted in the chip market over the last ten years. There are many more companies looking at building their own semis designs. Perhaps the most accurate way to describe this is that there is a pendulum. Once upon a time (going back to the 1960’s) all electronics companies had to build their own chips. Then merchant processors came and took over most of the volume markets. Gradually, it seems that the pendulum has swung away from that extreme.
As noted above, ‘building’ a chip really has two phases the design and the production. The world now has a much larger pool of talented chip designers to help companies with the design phase. We have also seen the emergence of a handful of companies to help with the production side. Importantly, a few of the merchant chip providers have also branched out into providing back-end or operations services for customers who buy their off-the-shelf parts.
Another important factor has been the changing economics of the fab market. A fab is the facility were a chip is manufactured, physically etching circuits into silicon. Today, very few companies own their own fabs. A leading edge fab costs something like $5 billion to build. However, not everyone needs to be at the leading edge. To compete in high volume electronics, having access to the smallest processes offers meaningful differentiation. But not everyone needs that. In many cases using a process that is two or three generations behind is perfectly acceptable. Production capacity at these fabs is far less expensive. If a product has power and room to spare, then this often works out to be fairly affordable.
One interesting wrinkle in this process is tied to the slowing of Moore’s Law. Put simply, it is taking longer for the industry to advance to each new manufacturing process. The old rule of thumb held that chips can double the amount of transistors they hold in the same area every 18 months. That has now stretched out to two years, and will likely keep going up. This means that everyone has stayed on existing nodes longer, which makes the older nodes ever more affordable.
One big example of all this can be found in the crypto market. As with all other forms of software, the better the chip the faster the software execution running on it. In crypto mining speed is usually a critical factor. With the explosion in crypto prices last year, the rewards for having the fastest chip on the market have brought massive gains to companies with access to those chips. The reward pools are so large, that many companies have begun designing their own custom chips for crypto mining. In crypto world these custom chips are known as ASICs, although for chip people this is a pretty anodyne term. There are now about a half dozen companies (mostly in China) that are building merchant crypto ASICs, as well as an unknown number of private, stealthy ASICs rumored to be out there.
Our point is that none of these crypto chips would have been possible ten years ago, but the diffusion of chip design talent has reached sufficient mass that companies are able to take advantage of the large reward pools for mining. Building your ow ship is possible, and certainly much easier than in the past, but this does not change the underlying economics. Building a chip design team can be expensive. For software companies, in particular, the benefits of doing so only make sense in a finite number of use cases. Instead of seeing some sort of broad change in semiconductor economics, the reality is that the largest companies are doing as they have always done, using their size to build bigger moats around their businesses. While over time we do expect the costs of building a chip in-house will come down, we should not confuse a profusion of easy venture capital with a revolution in the hard realities of building a chip.