Recently, we were speaking with a venture investor, reviewing a company they were considering putting money into. After we closed our fairly skeptical analysis, the investor paid us one of those compliments that we are told are common in the South: “We really appreciate your commercial view on this company”. Ok, but we had not been talking about its commercial operations, we had been discussing its technical merits, or lack thereof. It was at this moment, we realized that the VC had already made the investment, and we had not been giving them the positive feedback they were expecting. Needless to say, this company designed an AI accelerator. And we know we had not discussed the commercial side of the company because there are none.
We have written a lot of pieces on this theme, but this experience convinced us we may need to write a few more. By our count there are now 63 start-ups designing some for of AI accelerator chip in the US and Europe, and who knows how many more in China. We came to that count with a fairly simple Google search (this is a pretty good list), and we also have our own notes. There is some technical differentiation among these in so much as some are targeting ‘the Edge’ and some are targeting the data center with all that implies for chip design. A few list themselves as “RISC V” companies, but that is not a product definition, and almost all of those are building AI accelerators. A few of these companies seem to have differentiated architectures for their tensor cores. But for the most part, they are all making various trade-offs between memory bandwidth, performance, chip size and power consumption. In fairness, we have only dug deep into a handful of these, but very little of it strikes us as truly differentiated.
And then there are the commercial prospects. What is the market all these companies are selling into? A few are explicitly targeting low-power or IoT products and so may be able to peel off a niche market here or there. However, an awful lot of them are targeting the data center. This means they need to stand out from a few dozen competing start-ups, Intel, AMD, Qualcomm, and of course Nvidia. To say nothing of the hyperscalers’ own chips. What is left?
It pains us deeply to be critical of semis start-ups. We believe deeply that US needs more investment in early stage electronics companies, but bless all those investors’ hearts, this just does not seem to be the path to success.
Leave a Reply