Our big topic of the moment is the move towards more custom silicon (we have written about the topic here and here and here). Semis companies have always been able to build custom chips but they have been loathe to do so for largely economic reasons. Custom chips are expensive and risky. No one wants to design a chip for a single customer only to see that customer walk away at the last minute, after all the expenses have been incurred. But the temptation is always there, and for many companies custom chips are much closer to the norm than the exception. The best example of this comes from the world of analog chips.
We do not spend a lot of time talking about analog chips. The companies that make them build their brands on being reliable, stable and dull because they sell to customers who value those traits above bleeding edge hotness. That being said, there are some practices from the analog world that are likely to become more widely seen in the world of digital semis.
The first thing is that analog companies tend to be much closer to their largest customers. This is evident in their financial results. Analog companies typically have higher gross margins than digital companies, in the 60’s and 70’s for the former, 40’s and 50’s for latter, because they have more pricing power with many customers buying in smaller volumes. This evens out at the operating margin level, as the analog companies have to invest a lot more in customer service – all those sales engineers and application architects. (Yes, the big digital companies have much higher gross margins today, but as a rough rule this still holds).
That investment in customer service is a rough proxy for custom silicon. Here customers want a chip that meets their fairly specific set of criteria – say a 16-bit micro-controller (MCU) with a certain set of I/O ports and a little bit of embedded software. Another customer may want a 16-bit MCU too but with a different set of interfaces, and a third wants a 64-bit controller with the same set of I/O. The combinatorics quickly become a big headache. So the suppliers tend to find shortcuts – a line of 16-bit controllers with all the I/O options, and then they turn off the unwanted interfaces for each customer. The vendor can greatly streamline their production runs, and customers get the part they want. To name just one example, Indie Semi, an auto chip supplier, has really mastered this approach, and it is a big part of the reason they have survived in a very demanding market.
One of the key problems comes in the form of fab utilization. Each time a different chip runs through the fab, the line has to be taken down and re-tooled for the new chip. This is common across all manufacturing systems, and a lot of industrial history is finding ways to reduce this retooling and maximize economies of scale. Every time the fab is taken off line reduces its total utilization, and thus increases the cost of the fab. Even for fully-depreciated fabs, the re-tooling is a big opex expense.
So another way to manage the costs of custom chips is to produce every version of the chip (every SKU) that could be required in larger-than-demanded volume. Chip companies then have to manage all that inventory. This can be expensive in terms of working capital, but often those costs are minor compared to the benefits for other parts of the product’s margin profile. And the analog companies have a few tools for managing this cost. For one, they sell a lot of product through ‘the Channel’ – the vast network of distributors and channel partners. These typically have better pricing (customers are buying in smaller volumes) and serve as an escape valve for products that might otherwise have to be written off as excess inventory. (There are other, important benefits derived from the channel, which we will revisit some day.) And for many companies, having thousands of extra parts sitting in some warehouse is just a cost of business. We know many companies who will produce years worth of parts because they know they will always have a customer to buy them – that is the benefit of long-term design wins in industries like aerospace and automotive. As we noted earlier this week, some analog parts have product lives measured in decades. Part of the reason that analog companies are so boring is that much of their competency rests in the field of cost accounting and not in cutting edge design.
Will this work in digital semis? To be fair, digital companies already do this and have for a long time. While few produce truly custom SKUs, they already ‘bin’ their products, creating products that are identical under the hood (no fab re-tooling required) but are modified at assembly (or even in software) to generate differentiated SKUs. The leading edge digital companies will have a hard time stocking up on products for years because they become obsolete much faster. Moore’s Law does not always apply in the analog world, but is crucial in high performance digital categories like mobile, PCs and data center. But with the slowing of Moore’s Law, this practice will become more practical in digital, especially for sectors like automotive.
Ultimately, we think the digital chip companies are going to have to invest a lot more heavily in customer service and take the risk of designing truly custom chips for customers. This will be messy, and we think it will wreck many future quarters and more than a few start-ups, but if this risk can be offset with higher prices (which are coming anyway) it will end up being worth it.