Everyone believes AI will rewrite the economy, flooding in more money, detaching companies' valuations, and replaying the dot-com era (this time w/ GPUs and data centers instead of .com in companies' names) -- narratively, AI bubble.

I don't buy into that. Not because AI can't be overvalued (it definitely can), but because the "bubble" label ignores an important regime: reflexivity. So I tried a heuristic attempt in checking whether such beliefs have such an evidence of a positive looping feedback (at least for the near future). My feedback loop idea refers to the sequence: belief pulls capital, capital builds capacity, capacity creates usage, and usage becomes evidence that reinforces belief.

AI trend is real

One of the common critiques that is much discussed: "AI is hype and the fundamentals aren't real". To see if the trend is real, there are three important justifications that are expensive-to-fake: cash, revenue, and physical constraints.

A Goldman Sachs report earnings on Q3 2025 guided another increase in capex projections for the AI hyperscalers. Indicated by a capex spending of $106bn for hyperscaler in Q3 alone, the capex growth baseline is at +75% YoY. Such spending shows that large and mega cap firms are making irreversible commitments.

NVIDIA Q3 FY2026 also shows $57bn in revenue, with its data center segment generating $51.2bn (+66% YoY), showing demand from monetization.

International Energy Agency's Energy and AI report projects a +15% CAGR in electricity consumption for data centers alone by 2030, roughly doubling at 945 TWh. AI accelerated servers alone projected to grow at +30% annually.

The above are enough to show that the trend is real, and is supported by irreversible commitments. So the questions now should be directed to "who captures the value?" and "can the loop sustain itself?"

Picks and shovels win first

Another critique: "Only picks and shovels make money". In a gold rush analogy, most reliable profits go to whoever sells shovels. GPUs and data centers are the AI's picks and shovels.

I believe the critique is a misconception.

Michael Porter, in his book Competitive Advantage: Creating and Sustaining Superior Performance, breaks the concept of value chain into: inbound logistics, operations, outbound logistics, marketing and sales, and service. We'll use this as our rough lens.

Inbound logistics maps to our current bottlenecks: GPUs and data centers -- the raw inputs. As shown previously, they are already monetizing. For example, HBM suppliers like Micron have rallied significantly due to huge demand on the semiconductor chip and market's re-rating. This was discussed during Micron Fiscal Q4 2025 Earnings Call when management reported an annualized HBM revenue run rate of about $8bn.

The outbound logistics and the marketing and sales refer to the distribution on platforms that already own customers, where they can attach AI to their existing spend. As highlighted in Microsoft Fiscal Q2 2025 Earnings, their AI business has surpassed an annual revenue run rate of $13bn, +175% YoY. A similar monetization pattern shows up in Salesforce, another customer-owning platform: on its Fiscal Q4 2025, they ended the quarter at $900mn in Data Cloud & AI annual recurring revenue.

The operations and service refer to workflow owners, for which AI embeds into sticky business process with switching cost. Tim Tully et al. summarize that the infrastructure layer captured $18bn in 2025—the other half of all generative AI spending and up double from $9.2bn in 2024. The widespread adoption highlights the specific transformation of code-based inputs into functional software. For enterprises themselves, even if AI vendors compete in margins, buyers can still capture most of the productivity surplus. This is supported by Leonardo Gambacorta et al. in their findings: the use of genAI increased code output by more than 50% and is statistically significant at least among entry-level or junior staff.

So, a better implication should be "Picks and shovels win first".

Take Cisco Systems and eBay during the dot-com bubble. For simplicity, the chart I am showing compares their percentage price changes, reflecting the market's expectations at that time. As the inbound logistics of the early internet, Cisco rallied first. But after the bubble burst, the stock fell and converging into a similar outcome with eBay. At that time, as a customer-facing platform, it is arguable that eBay better represented the real economic value. In other words, the hype in "picks and shovels" deflated, moving it closer to what the market was supposed to pay (fairly) for businesses with more immediate.

You can still justify the above using other metrics such as earnings and cash flow, but the subtle point is: early concentration is actually normal during a shift. This point aligns with Gareth Campbell's study on how British Railway Mania of 1840s was amplified by leverage and speculation, yet the investments that built the transport infrastructure were proven (empirically) to deliver real productivity spillovers.

The question now is whether the value stays narrow or spread. We can use the reflexivity mechanism to find the answer.

Positive feedback loop

The reflexivity theory was developed by George Soros in his book Alchemy of Finance. The important assertions were: markets are always biased in one direction and markets can influence the events that they anticipate.

I mapped those assertions into the loop chain I mentioned previously:

The real trend of AI rejects the "story about stories", pulling capital to build capacity. For a higher price, overvaluation does not automatically imply a bubble condition. In fact, it accelerates the fundamentals, especially for both the capacity and the usage adoption. But it needs evidence to keep running, i.e., usage must translate into observable monetization.

The constraints are on the links connecting capital to capacity and capacity to usage. There are only two possibilities: constraints are manageable or they become binding and expensive. The latter implication: ROI is getting killed, weakening the usage evidence; hence, making the loop unsustainable (reversing it).

The former possibility shows up when constraints create scarcity and pre-commitments. In a sense, no idle capacity and collapsing demand are being indicated. Questions can then be phrased as "are markets tight?" and "are developers building against signed demand?".

JLL's NA Data Center Midyear 2025 report points out a large share of new projects are already preleased -- the construction pipeline being ~8GW and 73% of it is preleased. This indicates cashflow visibility and capacity is being financed on the committed demand.

CBRE's Global Data Center Trends 2025 report states that the global weighted-average data center vacancy fell to 6.6% in Q1 2025 (-2.1% YoY), indicating demand to continue to outpace the new job supply.

Both the scarcity and pre-leasing keep capital guarantee and utilization intact, supporting the reflexivity, rejecting the bubble idea.

To be clear, "not a bubble" doesn't inherently imply "no blow-ups". The remaining risks still lie on the buildout that might be poorly planned, e.g., data-center developers overbuild in the wrong places. As an early lead, some indicators that are worth monitoring according to the aforementioned constraints would be: rising vacancy rate, preleasing weakening, hyperscalers capex cuts, and financing stress.

Additional indicators from the "picks and shovels" would be wafer demand (through TSMC's CoWoS capacity guidance) and AI end-demand (through Anthropic and OpenAI's financials: top-line quality and cash flow trajectory).

References

Gareth Campbell, "Railway Mania: Not So Great Expectations," VoxEU (CEPR). https://cepr.org/voxeu/columns/railway-mania-not-so-great-expectations
George Soros, The Alchemy of Finance.
Goldman Sachs, "Why AI Companies May Invest More than $500 Billion in 2026," Goldman Sachs Insights. https://www.goldmansachs.com/insights/articles/why-ai-companies-may-invest-more-than-500-billion-in-2026
Leonardo Gambacorta et al., "BIS Working Paper (work1208)," Bank for International Settlements. https://www.bis.org/publ/work1208.htm
Michael Porter, Competitive Advantage: Creating and Sustaining Superior Performance. 1985.
Tim Tully et al., "2025: The State of Generative AI in the Enterprise," Menlo Ventures (Perspective). https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/

← How I Use AI to Get Closer to Ahyeon

1

10

11

12

13

14

15

16

17

18

19

2

20

2020-Fall

2021-Fall

2021-Spring

2021-Summer

2022-Fall

2022-Spring

2023-Fall

2023-Spring

2024-Spring

20K-smooth

20K-smooth (no image)

21

22

23

24

25

26

27

28

29

3

30

31

32

33

34

35

36

37

4

5

6

7

8

9

appendix

appendix

appendix

archive

BIO1001-General-Biology

bonus

bonus

bucketsort

CHM1001-General-Chemistry

collections

Competitive-Programming

components

cpu

CSC1001-Introduction-to-Computer-Science-Programming-Methodology

CSC1002-Computational-Laboratory

CSC3001-Discrete-Mathematics

CSC3002-Assignment-1-src

CSC3002-Assignment-2-src

CSC3002-Assignment-3-src

CSC3002-Assignment-4-src

CSC3002-Assignment-5-src

CSC3002-Assignment-6-src

CSC3002-Introduction-to-Computer-Science-Programming-Paradigms

CSC3050-Computer-Architecture

CSC3050-Project-1-assembler

CSC3050-Project-2-simulator

CSC3050-Project-3-verilog

CSC3050-Project-4-CPU

CSC3100-Data-Structures

CSC3150-Assignment-1

CSC3150-Assignment-2

CSC3150-Assignment-3