Given all increasingly competitive nature of the market for edge AI semis, someone recently asked us what it would take for a start-up to thrive in this market. In our mind, the best approach to this market would be for a company to offer not just a chip product but a “complete solution” bundling software and hardware. This is wildly different from how most start-ups have operated, but in our view offers a healthy mix defensibility and market growth upon which to build.

As a refresher, the edge inference market is highly fragmented. Much of this market will go to companies like Qualcomm who already have leading market share in ‘edge devices’, albeit hampered by customer expectations which make it hard to really monetize these capabilities. That means large swathes of ‘edge’ devices are already spoken for, by some fairly large companies.

That being said, there are still large swathes of opportunity. These includes well-known areas like video cameras and flying drones as well as emerging markets in industrial systems, robotics, and the big one – automotive.

So let’s say a start-up wants to develop a solution for this market, and just to make it more tangible let’s say this is an industrial system. Our company, Acme Industrial Automation, wants to design a chip that takes in a range of sensor data, processes that data and makes adjustment to the machinery it controls. This data could include temperature, humidity, machine speeds, positioning data on the equipment and power consumption. In theory, there is a market for a chip that processes this data, but most customers would not know what to with it. So Acme instead builds a software layer that runs on the data from its chips and makes adjustments to the machinery and provides notifications to operators. For instance, the machine could ‘learn’ that as temperatures rise, the machinery moves faster. Acme’s software could slow down the machinery to accommodate this and save power. Or it could alert the company’s ERP systems that output might be higher in this period. Or it could alert the HVAC system to turn up the air conditioning to cool the machinery and thus reduce wear and tear. Of course, there are systems that do this already, but Acme’s solution has “AI” and so learns over time what works best. It can also operate in real-time reducing the need for human interaction and provide much finer-grained tuning to the factory. We are not experts in industrial systems, and so there are probably all sorts of other features possible, but this seems a reasonable starting point.

To accomplish this, Acme has to overcome a number of challenges. First, it has to have an AI model that can actually do all this. We imagine there is a fair amount of ‘boot-strapping’ possible where an industrial engineer can design some basic functionality. This may even be an advantage for Acme as it broadens its deployments and captures more data, the model improves. We know of several companies who have already built this into their business, Tesla not least among them. Another big problem is that selling into this market requires all kinds of partnering, typically with very large companies who have very long planning and design cycles. Not insurmountable, but definitely an important factor. Finally, Acme also needs a robust software platform to sit in between their silicon and the customer’s machinery, to say nothing of the other systems like ERP, HVAC, etc. This means they need to be able to build and maintain a solid set of APIs which is fully documented and incorporates a process for continuous upgrades and bug fixes.

Note that almost all of these are software problems. In some senses, the chip is the easiest part. We can actually think of a few companies who started down a path like this and ended up as software-only companies, eschewing hardware entirely. Nonetheless, if Acme really can provide a superior chip, then supplying both provides a much larger revenue pie, critically one that includes recurring revenue, every company’s dream.

Moving from the hypothetical to the generalized, one thing we have been looking for later are semis companies that talk software from day one. We have seen a lot of start-up pitches this year, and most of them now include at least some mention of software, which is already a big change from just a few years ago. We think the next step in this evolution will be a much more prominent role of software in these pitches. In particular, we would like to see a semis company describe themselves as “software first”, or some terminology along those lines. What we have found is that many start-ups hire their chip design teams first and then run into a software team after a few months, but hiring software talent is incredibly hard right now, especially for candidates with “AI” in their resumes. So a good clue will be “Team” slides which clearly identify the software-focused members of the founding executive team.

Admittedly, there are good reasons why these companies do not really exist today. Building both hardware and software at the same time is challenging and expensive. That being said, we think the market is heading in this direction, and the ability to raise capital for this is improving. We think we will see many more companies like this in coming years.