Many chip companies are looking for or are being told to look for “services” businesses to sell on top of their chips. This makes a lot of sense. Companies that sell complex systems which bundle hardware and software capture better economics and build sustainable barriers to entry. There is just one problem – no company in the history of semiconductors has ever managed to pull this off. (We exaggerate, please send counter-examples, but we guarantee it is not a long list)
In a discussion on Twitter last week about the waning fortunes of AI chip start-ups Graphcore and Mythic Naveen Rao claimed that selling chips should not be the goal. Instead, he advocated for semis start-ups selling services.
Rao is the founder and CEO of Nervana, which was acquired by Intel for $350 million in 2016, one of the more successful AI exits to date. So his thoughts carry real weight.
He is not alone. None other than Jensen Huang, the CEO of Nvidia, laid out a vision for something very similar at Nvidia’s latest Analyst Day. The company calls this the “Omniverse” (Huang should also win an award for coining the best phrases in semis.) This highly ambitious service is billed as a “platform for creating and operating metaverse applications” and is probably best-labeled as a public cloud for Infrastructure as a Service. Hsu shares the big picture on this recent interview with Ben Thompson. And it is an exciting concept.
That being said, the difficulty Omniverse will face, and more broadly every semis company seeking to offer services, is that all of this competes with other companies seeking to offer bundled hardware and software services. These putative competitors are also known as “customers”. This is most evident when digging into what Nvidia is planning for the Omniverse. Much of what Nvidia envisions looks a lot like what AWS and Azure do every day. True, it has an angle around graphics and data that allows for specialization, but if that niche is big enough to merit the attention of the hyperscalers it will present a real dilemma for Nvidia. Who gets chips first? Who controls the roadmap? To name just two.
Another major problem with this is the idea of specialization. Investors, at least in the US, really favor companies that do one thing really well. This is especially true for start-ups, but is also a big part of the reason that the US shed its conglomerates 50 years ago. More specifically, a company that wants to build services on top of its chips really needs two engineering teams – one for the hardware and one for software. Not only is this expensive, it is incredibly hard to coordinate. A big part of the reason that Apple Silicon has been so effective is the company’s ability to integrate that team into the rest of the company’s roadmaps. Very few companies have ever been able to do this well. True all major chip companies today have large software teams, but this are always run as part of the chip design team, not stand-alone stakeholders, and so have a far more limited role in determining company priorities.
Reading through Rao’s thread. He makes the point that the big chip companies will not be able to accomplish this. They are structured to do hardware really well, adding software teams to that mix will present the same challenges that every large company faces when trying to incubate new teams internally. In this, we think Rao is probably right. Very few chip companies will be able to build this combination, but start-ups have a much better chance of doing so successfully. If they start from scratch with a systems approach then that structure gets baked in and the company builds around it. Start-ups will still face the problem of competing with some customers, but with the right business plan, they may be able to navigate those shoals.
Photo by Benjamin Davies on Unsplash
Semi provider selling services/data that competes with their customers = dumb.
Semi provider selling functionality and deep parametric data from the chip that helps customers gain operational insights = smart and monetizeable.
Makes sense. But the target customers for the latter is not that big.