The last decade has seen increased investments in hardware from major software players and VCs. The renewed interest in hardware is largely due to the rise of new technologies such as ML and cloud computing, and partly due to a general belief that hardware-software co-design helps build a competitive moat. As a result, the semiconductor industry, after many years of sluggish growth, has seen an influx of new entrants that hope to capture the expected growth in the broader market.
However, it’s hard to predict the extent to which these investments will bear fruit, because the hardware domain (specifically, the business of semiconductor chips) is fundamentally different than software in many ways. This post discusses some of the most challenging aspects of hardware today.
Scaling and lack thereof going forward
Anyone who spends some time talking to a hardware engineer will hear the words “the end of Moore’s Law”. It can sound ominous or exciting depending on your risk appetite.
Once upon a time, Gordon Moore predicted that the number of transistors on a chip would double every ~2 years. This prediction held true for about four decades, largely because we were able to shrink the transistor, the fundamental building block of a semiconductor chip. Shrinking the transistor is also referred to as scaling. To understand why scaling has been so crucial to the success of semiconductor chips, imagine a chip equivalent to the size of a nickel — such a chip would accommodate~2000 transistors in 1971, whereas a chip of the same size today can have over 50 billion transistors. That’s the power of scaling!
However after decades of scaling, the transistor has shrunk to the point that it’s very close to the limits imposed by physics for all practical purposes. In other words, Moore’s law will no longer hold true. This is the fundamental reason that hardware is getting harder to build. Essentially, a lot more effort is needed to extract meaningful gains out of hardware.
As non-intuitive as this might sound, hardware engineers mostly code just as software engineers do. The code that hardware engineers write represents the design of a chip, and bringing that design to life is called chip manufacturing or fabrication.
Chip manufacturing is mostly done in only a handful of companies in the world today — one of the largest is TSMC. The chip manufacturing process involves an army of chemical engineering experts and advanced specialized machines and results in billions of transistors and wires printed on a tiny silicon substrate. This mystic art of chip manufacturing is expensive and getting even more expensive with every shrink of the transistor. To put this in perspective, TSMC estimates that their future manufacturing facility will cost ~$20B. These costs intensify the cost pressures on chipmakers, and when added on to the point mentioned above — the scaling limits of the transistor — these costs invariably affect the ROI for chipmakers.
Long product cycles
This is perhaps the most bewildering aspect of hardware design for most of my software friends. Manufacturing a chip is like designing a city one brick at a time. Every neighborhood is designed for a special purpose, and every road has a purpose. Not only is building this city going to cost millions, but any improvements after it’s built will also cost millions. Hardware product cycles are generally 2 to 3 years long mainly because of this constraint — you can design a chip in stages, but you can’t incrementally manufacture it at every stage. It is unlike launching a website and fixing bugs post-launch. The biggest ramification of this constraint is that hardware must be designed to support the needs of the world 2 to 3 years in the future.
For companies like Intel that rely on selling hardware to exist, conducting business in the fast-moving software industry is getting increasingly difficult. For new players entering the market today, building expertise and developing processes that are required to compete with incumbents takes time. There are only a handful of players left in the hardware industry that can deliver a compelling product time and time again by the virtue of in-house hardware-software innovation (a.k.a. vertical integration).
Perhaps, a silver lining for hardware companies is that all of the aforementioned limitations lead to high barriers to entry for new entrants. These barriers keep the competitive space in check and will hopefully continue to do so, at least, until the next transistor is born.