Battle of Exponentials — A Different Way to Think About Accelerating Semiconductor Growth

Recent increases in semiconductor growth may be explained by an ongoing battle between two exponentials: 1) the upward-sloping growth of data generation, which drives semiconductor unit demand, and 2) the downward sloping reduction in transistor costs driven by classic 2D Moore’s Law scaling, which drives ASP reductions.

Data have continued to grow exponentially, stimulating unit demand. But in recent years, the pace of transistor cost declines has moderated. This helps explain why semiconductor revenue growth has accelerated, as shown in the chart below.

Chart 1: Battle of Exponentials. Source: Applied Materials—Strategy & Market Intelligence
Chart 1: Battle of Exponentials. Source: Applied Materials—Strategy & Market Intelligence

A third exponential is network effects (more about these in a moment), which are also compounding and putting upward pressure on semiconductor demand.

For chipmakers, this battle of exponentials bodes well for revenue growth, profitability and the ability to invest for the future. It can also serve as a catalyst for new methods beyond classic Moore’s Law scaling that can deliver the cost reductions in computation needed to make it affordable to extract value as data multiply.

Data Exponential Amplified by AI

We estimate machine-generated data first exceeded human-generated data in 2018. Our proprietary data model suggests that by around 2025, machines will generate 99 percent of data created, fueled by a confluence of emerging applications including AI, the Internet of Things (IoT), blockchain, the metaverse and Web 3.0. Arguably still in their infancy, these applications create new, virtuous cycles, with increasing data generation requiring commensurate increases in data processing.  

Just as machines are generating more data than humans, they are also making more decisions. AI turns data into commercial insights, and it is being deployed at scale, including for applications that were previously considered too complex or specialized to be automated.

Rise of Network Effect Technologies

Blockchain, social media and gaming exemplify applications that increase in value in accordance with Metcalfe’s Law, i.e., proportionate to the square of the number of participants. Human brains are wired to think in linear terms and often struggle to comprehend exponential effects in advance of them playing out. The premium valuations now accorded to social media companies, gaming companies and the Bitcoin network are recognition that the financial markets are catching on to the power and value of network effects.

Beyond Classic 2D Scaling

When Moore’s Law was functioning like clockwork (through around 20151), each new generation of advanced fab technology simultaneously enabled a doubling of transistors per chip, greater functionality, higher performance and lower cost. Affordable transistors fueled growth for the electronics industry, and investments and profits were concentrated in technologies aimed at further classic 2D scaling. Specialty nodes for markets Applied Materials calls ICAPS (IoT, Communications, Automotive, Power and Sensors) did not receive a lot of attention or investment dollars. However, over the past 18 months of IC shortages, the world has become acutely aware of the importance of a wide range of semiconductors that rely far less on shrinking to deliver value. The industry is changing, with demand and investments for semiconductor manufacturing becoming more evenly distributed between leading nodes and ICAPS nodes.

In my next blog post, I’ll share additional data about semiconductor growth—and explain additional ways to interpret the growth using the “battle of exponentials” framework.

1 Computer Architecture: A Quantitative Approach, Sixth Edition, John Hennessy and David Patterson, December 2017

Receive updates in your inbox!

Subscribe Now

Comments

I think all of these insights are very valid, but I think there may be a little more to it. And have been thinking on the following two threads. All in the context of a material acceleration in technology trends that have catalyzed a durable, step-shift higher in demand growth for semiconductors:

Thread 1 - demand in trailing edge has picked up dramatically due to four primary drivers: 1) the great acceleration in technology trends alluded to in the prior paragraph 2) historical underinvestment in trailing edge capacity because a) there wasn’t enough perceived pricing power to justify the incremental material capex outlays (that calculus has totally flipped) and b) demand wasn’t racing ahead until the last 18-24 months 3) purchasers in trailing edge semis have suffered an acute lesson in what it means to their business when they can’t source trailing edge semis (that cost mere $1s and can hold up the sale of $50,000 vehicles for instance) which has thrown the idea of just in time semi inventories out the window (driving further demand) and 4) as alluded to in your post a lot of the products (vehicles, fridges, all manner of appliances) don’t have the same space constraints as a sleek cell phone or laptop and don’t require the intense compute power of those items either, but everything is getting “smarter” and needs more semi content.

Between those four demand drivers, I do believe that we’re in the early innings of a huge new wave of trailing edge semi demand that’s resulting in a knock-on effect of an enormous re-focus on building out more (and upgrading existing to make more efficient) trailing edge capacity. That said, while it’s only one company, TSM in its blockbuster capex forecast announcement this week stated it would be dedicating 70-80% of its unprecedented $40-$44bn capex budget to “advanced nodes.” Its very clear that materially stepped up capex spend from all major analog players (which are by definition focused on trailing edge nodes) and Intel’s huge foundry intentions (which will by definition be largely focused on trailing edge because INTC currently doesn’t have leading edge capabilities) will partially balance that out, but the cost of the competing at the leading edge (the capital intensity) is clearly going higher (as alluded to in your blog post).

Thread 2 - Is there a major demand driver here, especially for semicap equipment, that’s coming from increasing size of high performance computer chips. Look back to the decade of the 2000s where the semicap industry was materially, negatively impacted from the shift from 200 to 300mm wafers. That led to a huge step up in semi manufacturing industry capacity without the linear increase in need for semicap equipment. With HPC semis and chips for companies like Apple becoming dramatically larger, that’s taking up more space and leading to less chips coming off each wafer. With semiconductor unit growth needs only going forever higher as there are more and more technology implementations and many devices graduating to higher and higher compute capabilities, is this current dynamic of dramatically larger chips having the effect of cannibalizing existing semi manufacturing capacity and driving even greater need for enormous outlays for more semiconductor manufacturing capacity at the leading edge (almost the opposite effect of what happened in the 2000s)? Long term capex forecasts from TSM, Samsung and SKHynix provided in 2021 all are supportive of this thinking. And given TSM’s recent capex guidance and it’s confirmation at multiple points during its call that it wouldn’t be investing so heavily in capex if it wasn’t so confident in strong and long term improving demand dynamics for its business (with particularfocus on HPC where those chips are often the largest), I think this dynamic is only just starting to be understood by those investors focused on semis and not even on the radar of the broader market.

Want to join the discussion?

Add new comment:*

* Comments must adhere to our Discussion Guidelines and Rules of Engagement.

You can also fill out this form to contact us directly and we will get back to you.