Recently, working with a group of investors, we met with a company designing a chip that offered 10 times better performance than the leading chip on the market. (This was a digital logic chip, you can guess the category, but we are going to leave it deliberately vague.) We then engaged in an hour long debate with these investors as to the attractiveness of this company. Much of the conversation boiled down to how much more money this company would need to reach escape velocity cash flow – to put it simply they needed a lot more money to get where they were going. Some thought that 10x better performance was all we needed to know, but the cash burn remained a sticking point, to put it mildly. The company was resistant to cutting its burn because they needed to invest heavily to bring their product to market, and what they really needed was to invest in software. And there was no easy way around that.
A few months back we pointed out that semis companies all need to bulk up their software skills. Semis companies would all like to make money selling software services alongside their chips, but as this company (and many others) demonstrate software is not easy, or cheap.
Semis companies really have three problems when it comes to software.
The first is obviously the cost – they need to hire a software development team, and even in this economy those are not cheap.Our rule of thumb is software teams can easily be as large as the chip design team, so roughly 50% of the engineering workforce.
Second, the company needs to know what software to build. This is particularly challenging and has wrecked many chip companies in recent years, especially in AI. Six months ago, no one was building an LLM or Transformer or GPT chip. Now we imagine there are a few, but who knows where the AI world will have moved in another six months. And even if GPT is the future, a chip for GPT v.3 is going to be very different than a chip for GPT v.4, let alone v.5. This is why Nvidia still has such a solid market position. There may be some purpose-built AI accelerator out there that performs better than a GPU for a particular task, but when that task changes, the performance evaporates. And this is true for many other categories of chips. The best way to know what end customers want in software is to design it directly with them, but in many cases that means competing existing customers and partners.
The third problem is that many customers already expect certain things from the chip vendors, including some form of software support. This is mostly true for low-level software or firmware, but there is no arguing that there are already software expectations built into semi sales. This means that software is not a product in its own right, but a feature in the hardware. Convincing customers to pay more for that is possible, but the market is competitive and in this context software/support are just one more competitive tool.
And this leads to the real conclusion which is that semis companies do not really want to be software companies. They want recurring revenue and they want software company valuation multiples. And what they actually need are forms of competitive advantage. Software can convey this, but there are other ways to achieve it. As software consumed the world over the past decade, we think many executives lost sight of this. Building software for the sake of software should not be the goal. Finding better ways to monetize their products and gain a sustainable advantage in the market is what matters.
You nailed it with this post.
We are at the end of the current semiconductor technology roadmap. No more semiconductors on the Periodic Table; can’t use light anymore to pattern. The general-purpose binary microprocessor has little room to advance…
which means the software that runs on this processor is where improvements in compute power live – and where value proposition lives.
The semiconductor industry needs only to look at the display industry to see what its future looks like if the semiconductor industry doesn’t embrace advancing algorithmic compute efficiency and performance. That is, growing revenues with shrinking profits…