As we discussed earlier this week, the market for non-chip companies building their own chips is growing, spreading beyond Internet companies to a growing number of industrial and automotive companies. Many people believe that ultimately every company that makes durable hardware or complex software systems will want to build their own chips. The semis companies are building out business lines to support these customers (in hopes of driving them to purchase catalog parts), and there is even the potential for Google to offer tools which let any software developer design their own chips. Are we moving to a world where everyone can build their own chip? Will there be a “Long Tail” of random, assorted chips designed by a huge sea of “makers”?
Probably not. We say this not because we are doomsayers or critics of a growing phenomenon (there is enough gloom already). Instead, we think it is important to explore the boundary of who should and who should not build a chip. Understanding this boundary will help inform the decision making process and hopefully lead to better outcomes for companies who do choose to go down this path.
A major constraint here is that semis are not software. Literally anyone can start writing software right now. The world is awash with free tutorials and coding tools. These cost nothing more than an Internet connection and a
good working computer. Instead, semis are software made corporeal . That last bit makes a big difference. The costs of converting the design of a chip (i.e. software) into a tangible device are immense.
Consider the surprisingly large number of people on YouTube laying out a chip with an electronic bread board and wires (lots and lots of wires). We have recently taken to watching live TikTok streams of one college student building a CPU this way, and his journey over the months can best be described as a descent into madness. Yes, anyone can build their own semis at home, but only the most zealous will.
So then we are dealing with true semiconductor manufacturers – the foundries. There are dozens of companies that will build any design, but these are not cheap. And there are really two buckets of expense to consider. The obvious one is just the cost of taping out that design (i.e. transferring a digital design and readying it for production). Want to tape out a chip with TSMC’s most advanced process – that will cost $50 million upfront, then $20,000 or so for each wafer. Mature, trailing edge processes are much cheaper – with tape out cost measured in tens or hundreds of thousands, with each wafer priced in the hundreds of dollars – more affordable but still significant. But there is also the cost of designing the chip itself – including hiring a team, buying all the required licenses and IP, and most crucially managing that process. By this stage, we have largely eliminated hobbyists and the vast majority of the population. But we have also removed most companies.
What this really leaves are companies that build a lot of software or hardware. This definition is still a bit squishy – where would banks fit in, for instance? Big retail banks have very large software teams, but it is hard to see them ever wanting to build their own chips. On the other hand, this audience is broader than just “technology” companies – John Deere is the obvious counter-example, but so are automotive companies and machine tool makers. What about mid-sized software companies? Again, for most of these companies, there is no point. Most software companies today do not even operate their own data centers, running instead on AWS and the public cloud. Before they design their own chips, they are probably better off first building their own data center capacity.
What this leaves are a pool of companies that make or operate some form of hardware, who also own their own software stack. This narrows down the pool considerably, but is still a fairly large base.
At this point, we imagine many people are thinking that all of this is a lot of trouble, and that only a select few companies – a few dozen – will see any point to designing their own chips. That being said, there is a strong force working in the other direction pulling companies towards this path – competition. If your competitor successfully designs their own chip it has the potential to strongly disrupt the market in their favor. Apple is currently the best example of this – the N Series chip for phones and M Series CPU for laptops deliver performance gains that drive meaningful share gains of phones and computers.
Now imagine Acme Refrigeration which makes refrigerators. Somehow, Acme invents a system that allows their products to consume 30% less electricity than other refrigerators. Acme also invents an algorithm that scans the items in the refrigerator and uses AI to determine every item inside, it can estimate the amount of milk and eggs in their cartons. Of course, this has been the promise of tech companies for as long as someone has needed a simple way to explain the potential of computers to their non-technical friends, but let’s just say it is possible, finally. The catch is that this system needs a fair amount of compute power. Acme goes around to vendors but finds that all they offer are full-powered CPUs and GPUs which are too expensive, a $100 per refrigerator. So Acme goes out and designs their own chip – a Refrigeration Processing Unit. The RPU has access to wired power and has ample space inside the door to work, so Acme can use a slightly older manufacturing process for their chip. Below is some vey rough math on how that works out:
For an additional $8.83 per unit, Acme now has a refrigerator that is meaningfully better than the competitors’ products, and better in a way that consumer can understand easily, so it will have a big impact on sales. How do Whirlpool, GE and Kenmore respond? Actually, the leaders in US refrigerator market share are Samsung and LG, and it is safe to assume we know how they will respond. Having the right chip can make a big difference, enabling things that were not possible before. This balance of forces means that every hardware maker has to keep one eye on the market and probably start hiring a few chip designers, just in case.
Ultimately, we do not think there will a “Long Tail” of thousands of companies designing their own chips, but there will be a few hundred, and that is still an important shift for the industry.
Photo credit comes from this streamer on YouTube. We cannot vouch for its content, but it gives a good impression.