AI is going “to the moon” – the recent rally in semiconductor stocks reflects this sentiment, and semiconductor stocks have risen 180% since the beginning of 2023. Are we in a tech bubble, or are these stocks appropriately valued due to revolutionary technological change?
This piece breaks down this question in three steps:
1. “Customers of Compute”: Estimates computation demand from now until 2030 if we optimistically believe AI is going “to the moon.”
2. “Country of Compute”: Maps the current supply chain and estimates how different players would cut up future demand into value today – if the entire semiconductor chip-making industry were a “country of compute,” how valuable would it be today? In its supply chain, how much value would different sector slices capture?
3. “Companies of Compute”: Show the sizes of companies in the supply chain to potential future value if AI goes “to the moon.” How much of that value they are currently priced to capture?
This topic is both incredibly broad and incredibly deep. As such, this piece only gives high level conclusions (while pulling out key assumptions). We are not forecasting future technological advancements, just making sense of how markets are evolving and reacting to fundamental changes. If we take the upper bound of what we can imagine to be the pace of future technological change – how overstretched do current valuations look on an absolute and relative basis?
Section One: “Customers of Compute” – if AI goes to the moon, how much demand for computing will there be?
The short answer is a lot, even more than some semiconductor industry experts* are predicting (~$1 trn USD revenues for the semiconductor industry in 2030.) Our “to the moon” optimistic estimate is close to $2 trn of revenues in 2030. We assume AI develops quickly where possible. Below we walk through the methodology:
The 2023 global semiconductor market is about half a trillion dollars. Customers of computation power (which needs a variety of different semiconductor chips) span the below uses (HPC is the piece powering AI computing models).
What does it mean for AI to moon? We broke this into a few key pieces that we estimated and summed up to arrive at our forecast for the path of total computation and chip demand (revenues, USD) until 2030. We assume that compute chips are 60% - 70% of all semiconductor chips (of all types) for HPC.
1. To Reach AGI: There are many well-funded true believers trying to create artificial general intelligence (AGI). We assume that large language models (LLMs) will keep scaling relentlessly until this happens, along with computation demand.
This roughly results in models becoming 1.7x larger each year**, although the specific path is impossible to predict – improving old methods slows down at the margin, while new breakthroughs will suddenly and massively scale. AI models doubled every 3.4 months from 2012-2018 (7x larger each year, per an OpenAI paper) and 1.8x per year from 2020 to 2023 (comparing size of models from 2020 to 2023). We optimistically assume that growth keeps up with the recent past even as things get harder.
We convert the computation demand into a USD demand by also estimating a cost curve for computation power (that assumes Nvidia’s previous pace of cost reduction, with a faster decline in costs 2026-2027 as more competitors catch up).
2. AI Ripple Effects (Applied AI): There will also be those who don’t care much about creating AGI but are very interested in profiting from commercialization of AI models (which has already started in 2023).
Grow and Improve Other Existing Compute – we assume that AI supercharges growth in non-HPC chip-reliant industries. They take advantage of AI advancements and double their rate of growth.
Substitute Labor Costs – one of the biggest areas that AI can add meaningful value is in replacing the service sector (~21 trn USD of GDP across US, Europe, and Japan) and lowering service costs. We assume that 15% of the service sector is gradually substituted by AI over the next ten years, starting off slow and increasing exponentially as changes take traction (see Appendix 1 for more on methodology).
There are likely other impacts we can’t imagine – however, we’ve been generous with estimating what we can imagine to compensate.
None of these assumptions are meant to be precise – this is an optimistic take for future AI-led computing growth. It takes more than just chips to produce computation power, but chips are the most important component. To simplify, we look at just the chip production piece to meet future computing demands.
Section Two: “Country of Compute” – how do we value today these future revenues?
Suppose the semiconductor supply chain producing chips to service computation demand were a small country, a “Country of Compute,” with different sectors. The “Total Chip Demand” revenues we estimated in the previous section would be the GDP of this “Country of Compute.” In this section, we estimate the total value of the “Country of Compute” and the value of the different sectors.
Below is a (simplified) overview of the process by which semiconductor chips are designed, created, and packaged to customers to turn into supercomputers and other devices. To accurately evaluate what constitutes reasonable sector market shares and profit margins, we first need to deeply understand the varying degrees of difficulty for numerous inputs, and the complex collaboration between different players. For those already familiar with semiconductor design and fabrication, feel free to skip this section.
We break down the semiconductor chip production chain into 6 main “sectors” and estimate each sector’s share of overall profits depending on the importance and difficulty of their contributions – this reflects how much of the overall “Country of Compute” GDP each sector can capture.
We looked at historical revenue share from an April 2021 SIA - BCG report and then adjusted to align with the relative contribution we’d expect.
We took the weighted average profit margin by sector.
We estimate a current valuation of ~$7 trn USD for the entire semiconductor industry. Below are our assumptions:
We use the 2030 revenues for the “Country of Compute” from section one, which assumes explosive growth.
We split this between the different sectors, then use recent weighted average by-sector profit margins to estimate 2030 profits (some monopoly companies may have up to 2x the average profit margins, but over the long term we’d expect competition to keep the average roughly in line and shuffle winners).
Since the semiconductor industry is cyclical, we use a long-term average industry P/E to estimate a 2030 “Country of Compute” valuation.
We use the risk-free rate to discount the 2030 valuation to today.
The valuation seems very high, but it’s an optimistic scenario for AI going to the moon. Maybe this is the $7 trn that Sam Altman was talking about.
Given this estimate of current valuation for the overall semiconductor industry and each sector, the next step is to put individual companies in perspective. How much of future production are companies priced to capture? How much of available liquidity can keep buying?
Part 3: “Companies of Compute” – are these companies overvalued?
In this section, we look at how overstretched equities in this industry look – both from a value perspective and from a liquidity perspective.
First, we take the biggest ~50 listed companies relevant to the semiconductor supply chain (out of over 100+ total companies involved). These 50 biggest companies account for > 90% of total semiconductor industry revenues. Since many are conglomerates, we calculate a “semiconductor market cap” for each company based on the share of total company revenues that come from its semiconductor production business.
Next, we look at these companies by sector to see how current market caps (only the “semiconductor market cap” portion) compare to our projected sector value in an optimistic AI growth scenario. Our main takeaway is that if you do believe AI is going to the moon, valuations are roughly reasonable, although some sectors are more saturated than others (possibly reflecting higher sustained pricing power or faster expansion to other business lines). At the individual company level, we are not here to give specific investment suggestions so will refrain from commenting. However, we think the data is interesting.
Relative to available US and global developed world liquidity, market caps for the biggest tech companies look stretched. Of these, half are big in the semiconductor supply chain (marked blue), and 9/10 are involved in the new AI and computing wave.
Looking at the biggest 10 components of the NASDAQ, their combined market caps are 70% of USD M2 (a metric to measure money supply and liquidity) and 30% of total developed world M2. Even at the peak of the 2000 tech bubble, the biggest 10 components of the NASDAQ were a smaller % of both USD and global developed world liquidity. This is not to say that a crash is imminent or that the fundamentals today necessarily mirror the 2000 tech bubble, but it would take very sizeable inflows to push prices up even further.
These are some personal observations and do not constitute investment advice. Investing is risky and make decisions at your own risk.
Appendix Section
* ISSCC (International Solid-State Circuits Conference)
** Since it takes considerable resources to join the race, we estimate ~5-6 main contenders as of 2023, scaling to maybe double that in 2030. We assume AGI zealots splurge and maximize peak computation power to train and query their models. We then project peak computation necessary for these models through time.
Appendix 1: Services Substitution
We take 50% of services sector GDP that gets substituted away to “AI services” and attribute to chip revenues – the rest is attributed to a combination of 1) non-chip inputs 2) AI service providers and 3) decreasing costs from substitution productivity gains to incentivize customers to make the switch.
Disclaimer: Does not constitute investment advice, represents author’s own research and opinions. Make investment decisions at your own risk.