One trend that has been accelerating for a couple of decades is turning fixed costs into variable costs. Often this is what is behind outsourcing some capability. Sometimes it is driven purely by lower variable costs (let's hire a team in Shanghai) or core-competence considerations (we don't really need to run our own cafeteria). However, often it is driven by a desire to switch an inflexible fixed cost for a variable cost. The biggest of these trends in our industry is the foundry/fabless model. Instead of a semiconductor company building their own fab (fixed cost), they buy wafers from foundries (variable costs). Fabless and Foundry There are two big problems with large expensive fixed costs like a fab. One is that it is expensive, and so it ties up a lot of capital that could be used for something else. But the fixed cost usually puts in place a fixed capacity of some sort, some number of wafer starts per month, in the case of a fab. But that capacity always risks being mismatched to what the market needs. If there is not enough demand, then the fab will not be at capacity, and the cost of the unused capacity, called fab variance, has to be absorbed into the financials. This is usually done either by adjusting the wafer price for the wafers actually manufactured, or simply adjusting the overhead value used in the financials. On the other hand, if the demand is higher than the fab capacity, you can't make enough product. You can actually run a fab over capacity (say 110%) for a time, by things like deferring maintenance, but you can't do that forever. Memory, where volume is in increments of a fab, is notorious for either being in feast (not enough capacity, so high prices) or famine (too much capacity, cash flow is everything). Right now, DRAM prices especially are very high and account for a lot of the recent growth of the entire semiconductor market (see my recent post Semiconductor Rankings: Thanks for the Memory ). Before the era of the fabless semiconductor company, an IDM (or just a semiconductor company, as they were called back then) had to have a variety of product lines so that they could fill the fab. If one product line was ramping early in its life, others were at a peak. If one business line was weak, another was strong. One of the under-recognized aspects of the fabless/foundry model is that, not only do you not need to build a fab, you don't need to have a variety of product lines to fill it. Most fabless semiconductor companies start with a single product, and grow it into a single product line, before worrying about doing something else too. Cloud Services In the software/internet world, you no longer need to build a high-capacity server farm. Amazon's AWS will sell you server time and disk space on a purely variable cost basis. This is a huge advantage when you are small, since the costs are very small, too. Another big advantage is that you can ramp fast. When Pinterest started to hit it big, they had 18 million visitors one month, compared to 12 million the month before. Try doing that with a physical server farm. I can't find a picture online, but I remember it from a presentation at Mobile World Congress, that when Pokemon Go launched, they had a level of downloads they expected, and a maximum level beyond their wildest dreams. The actual numbers were over ten times the dream level. It was just a tiny blip for AWS to absorb. When companies get big, they don't necessarily switch. Famously, Netflix runs entirely on Amazon's AWS. No fabless company, even the ones who are now very big, has ever switched and built a fab. When Dropbox announced they were switching from AWS and going to run their own cloud infrastructure, it was news just because it was so rare. If you look at the companies that supply the service, they are in the business of taking fixed costs and turning them into variable costs. A foundry builds a fab for $7B or whatever, and sells wafers for a few thousand dollars each. Amazon builds lots of enormous datacenters, and then packages access up in various variable cost services. However, there are other companies that we interact with as consumers that are similar. Airlines largely have planes (fixed costs, although they often turn them into something closer to variable costs by leasing them) and sell you plane tickets. Mobile networks like Verizon or China Mobile have a huge number of base stations, which is largely a fixed cost, and they work out how to charge variable costs to each customer for accessing them. It is not always obvious what the best model is, and that answer can change over time. For example, mobile networks used to be about voice, and they threw some data in for free. Now they are all about data and they throw voice calls in for free. EDA Until the early 1980s, semiconductor companies largely had their own internal EDA groups. They created their own tools. If you go back further, they created their own manufacturing equipment. I was at VLSI Technology, for example, and we had our own design tools. We created circuit extractors, DRC, layout editors, schematic capture, later synthesis, and place & route. Why? Because you couldn't buy them when we started. The beginnings of the EDA industry supplied circuit simulation and polygon layout only. When the Japanese semiconductor companies decided they need to do more than memory (which they could do with simulation and layout), they didn't really have the in-house capability to develop their own tools. Some of the early investors in SDA, one of the forerunners of Cadence, were semiconductor companies who needed tools but didn't have the capability to develop them. Gradually the EDA industry as we know it today came into existence. Cadence invests nearly 40% of its revenue in engineering, so in very round numbers $750M per year of its $2B revenue (you can find the exact numbers in the 10-K and 10-Q reports...now I'm starting to sound like I'm on an earnings call). I doubt any one company uses everything we produce, especially in the IP area, but even 10% of our engineering cost would be $75M per year. No wonder companies don't develop their own tools anymore. And that's without considering the expertise and the difficulty of assembling a development team that could do it. In some ways, we are in the business of taking a fixed (almost) cost of a huge highly skilled engineering team, and turning it into variable costs by selling licenses. Cadence Cloud Now, with Cadence Cloud, we are even more in the mode of turning fixed costs into variable costs, leveraging AWS, Microsoft and Google's clouds, along with cloud-ready EDA tools to make it possible to scale to a huge scale without requiring the fixed costs of a huge server farm. For smaller companies, they can survive without any server farm at all, for established companies with existing server farms, they can "burst to the cloud" for peak needs, and when their farms need to be refreshed have the option of gradually migrating into the cloud. I have heard that the total cost of ownership (TCO) of servers in the cloud is an order of magnitude lower than on-prem datacenters. I don't believe that it can be that big a difference, but even a factor of two would be compelling to a CFO. Of course, the elephant in the living room is whether EDA will switch to a SaaS model, with either metered use or more of a subscription model. For now Cadence has a hybrid model, with the same sort of time-based licenses as we have had historically, and with EDACard for flexibility to handle changing requirements without needing to go through a full formal purchasing cycle. Sign up for Sunday Brunch, the weekly Breakfast Bytes email.![]()
↧