There was an interesting post on Techcrunch this afternoon, called “The Revival of Semiconductor Investing”. The title alone touched a nerve. It states something that I would very much like to believe is true, yet fear deep down inside is not. The author, Ilgiz Ahkmetshin, is a venture investor at SK Telecom (a South Korean operator) in their incubator. More than that, he actually crunched some data. What he showed is that even though the number of semiconductor-related venture investments has fallen sharply, in the past few years, the number of investments in new semiconductor companies has started to grow again.
This seemingly defies common “Valley wisdom”. Most VCs I know find the idea of investing in a semiconductor company deeply unattractive, if not downright laughable. Now that I am considering my next career steps, I have been watching such investment trends closely. Speak to most people in the semiconductor industry and there is a clear sense that we have entered a new age, one with little room for new, small chip companies.
For several years, I have been suggesting that there is a great need for chip start-ups, but the funding model has to change.
This is how I see the problem.
Designing hardware is much harder than designing software, and by ‘hard’ I really mean expensive. If you design a web site, the whole product can be built around the idea of iteration. The first version will not work perfectly, so plan for that, and fix it in upgrades. And upgrades on the web or in the cloud are very easy. In hardware, everything has to be right before you ship the product. And not just before you ship, but really before you even start to build it. Returns are prohibitively expensive, and hardware cannot be readily fixed on the fly or in the field. This problem is even more acute in semiconductors. First, those chips have to get designed into other things, and that means even longer preparation times. Secondly, the complexity of the semiconductor manufacturing process are hard for the human mind to grasp. The industry requires incredibly tight specifications, as you can imagine for circuits many times smaller than a human hair.
In terms of investment dollars, this has meant that the cost to go from ‘good idea’ to revenue-generating product is very large. There are several private chip companies out there that have raised north of $200 million and never shipped a product much less gotten to profitability. So if you are a venture investor, you can roll the dice on that kind of prospect or launch a photo-sharing messaging, cloud storage app which gets to revenue in hours. Not much of a choice, except there are already a lot of those apps out there, and a growing shortage in semiconductor innovation.
I have been arguing for years, that the problem is one of expectations and funds raised. Too often, chip companies were funded on the premise that they need to get to the point where they are shipping product. This has meant that companies needed to build a relatively large operation. WhatsApp did fairly well with something like 30 employees, most of whom focused on the core app and user experience. By contrast, for a chip company to get to revenue, they need to build a big operations team that can take clever designs and walk them through the all the steps needed to get manufactured. They need a large sales force. They need incredibly expensive design tools. The list goes on, and adds up to a lot of people and a lot of expense.
It is unclear what can replace that model. I think a better approach would be to break up the whole system and allow for some abstraction. VCs should fund promising chip designers, not on the expectation that they will get the product all the way to shipment, but to proof of concept, when the investment will be handed over to a company that specializes in transforming designs into prototypes. I think there is also room for clever incubators which pool some of the more expensive elements of chip design like software tools. This sort of ecosystem may be starting to emerge, but is still far from maturity. And none of this is easy. For chip designers to merely prove a concept, they would essentially be selling IP, and the market has not been that kind to IP companies like this. At their very worst, they end up becoming companies with more lawyers than engineers.
Nonetheless, I think a new model is starting to emerge. Many design tools are now becoming available on Software as a Service (SaaS), pay-as-you-go models. There are a few companies out there that can handle many of the complexities of translating designs into shippable products.
One of the paradoxes of the modern semiconductor market is that while starting a new chip company from scratch has only gotten more expensive, the cost of building a chip have generally come down meaningfully. I do not mean the physical costs of manufacturing, that’s a topic for a different day, but the upfront costs have come down in many areas. This is the reason that the big web companies like Google and probably Facebook, are now designing their own chips. I am working on another post which lays out the math behind Google building its own server chips, and it turns out that this is a pretty easy investment choice for them to make. Part of the reason for this has been the rise of ARM Holdings, which designs and licenses a key piece of chip IP. So in many areas, the abstraction process has already started, with a dramatic impact of chip design costs.
Now ARM designs only work for certain chip applications. There are many more areas which have little or no abstraction available, and some where it is not even possible. Nonetheless, I think there is room for a new model to emerge that would greatly reduce the upfront costs of starting a chip company.
So I am encouraged to see posts like Ahkmetshin’s with tangible proof that investments may be picking up in this space. We still have a years to go, but maybe there will be a day when Venture investments in semiconductors once again make sense.
Commenting from the perspective of someone who spent time at a company wending its way from FPGA to gate array to ASIC. Yes. Errors mean respins, which cost time and money. That said, I feel since the bar is higher and there’s more of a premium on quality before shipping the first time, there’s more room for black magic and differentiation. Not many make it through, but those that do can create value for market and shareholders. From my time at Rosum, MaxLinear and Microtune are two who made it through. Atheros is another, then’s there’s SiRF and Global Locate. There’s also Enuvis and DiBcom and Siano. Coming from that background, the speed with which a software company can get traction and market feedback is both amazing and frustrating. But there are fortunes to be had for those with the stomach for it.
Pingback: iCloud and Apple's Founding Myth | stratechery by Ben Thompson·