Insights AI & Innovation

AI Is a Supply Chain, Not a Single Stock

Chatbots make AI look simple. Underneath is one of the most complex supply chains ever built, and its layers matter more than finding the one AI stock.

July 20, 2026 · 8 min read

Chatbots make AI look simple. Underneath sits one of the most complex supply chains ever built — and for investors, understanding its layers matters more than hunting for “the AI stock.”

Your phone is a small miracle you never think about. Behind the glass sits engineering of staggering complexity, and behind that, a global supply chain capable of producing millions of identical devices a year. Artificial intelligence works the same way. Typing a question into a chatbot feels effortless — but that simplicity rests on a full value chain of chips, buildings, power, models, and software that would have read as science fiction a decade ago. That chain has become one of the most important forces driving markets and the economy, and it’s the reason a serious view of AI has to go well beyond a handful of technology stocks.

So let’s walk the chain, layer by layer — with real data at each stop — and finish with how we’re actually positioned for it.

01 · The Value Chain

“AI” is a supply chain, not a single stock

The most important insight for investors is that “AI” is not one type of investment. It’s natural to think first of the model providers at the heart of the capabilities — OpenAI, Anthropic, Google, and others — but they are one piece of a much larger puzzle. Around them sits a full supply chain spanning semiconductors, data centers, the models themselves, and the software and businesses that put them to work. Each layer has its own economics, its own business models, and its own risks.

The four layers of the AI supply chain
Each layer has its own economics, business models and risks. Demand flows through all four.
01SemiconductorsGPUs, memory and networking that power training and inference
02Data centersBuildings, electricity and cooling that run the hardware around the clock
03AI modelsLarge language models trained across thousands of connected servers
04ApplicationsSoftware and businesses that turn the capability into productivity
For illustration only. Not a recommendation of any security or sector.
Four layers, four different businesses. Framework adapted from Clearnomics.

At the foundation is the semiconductor hardware, and it gets pulled on in two very different ways. The first is training — building the large language models themselves, a process that runs enormous datasets across thousands of connected servers for weeks or months. The second is inference — the everyday use of those models, where every single prompt consumes computing and memory resources to produce an answer. Together, they explain why demand and pricing for this hardware have surged, and why the companies supplying it carry the valuations they do.

AI
Training — Build the Model
Huge datasets processed across thousands of connected servers
Runs for weeks or months per model generation
Massive, lumpy, upfront demand for compute
Inference — Use the Model
Every prompt consumes compute and memory
Scales directly with adoption and usage
The recurring demand stream — and it’s still growing

Why does the distinction matter? Because a debate about any one layer — how much training compute the next model generation needs, say — is not a verdict on the whole chain. Inference demand, data center capacity, and software adoption each run on their own clock. Judging “AI” by one layer is like judging the auto industry by tire sales.

02 · The Buildout

The buildout is real — and it’s in the government’s own data

Hardware has to live somewhere. Picture a data center as a warehouse packed floor-to-ceiling with servers running 24/7 — demanding security, electricity, and industrial-scale cooling. Scaling AI means building these at a pace with few precedents, and you don’t have to take Wall Street’s word for it: it shows up directly in the U.S. Census Bureau’s construction data.1

Data center construction now exceeds all other office construction
Private construction spending, seasonally adjusted annual rate, Jan 2014 to May 2026. Data centers passed all other office construction in Nov 2025.
Data center constructionAll other office construction
$0B$20B$40B$60B$80B2014201620182020202220242026ChatGPT launch, Nov 2022$65.69BData centers$49.47BAll other office
$0B$20B$40B$60B$80B201520202025ChatGPT launch, Nov 2022$65.69BData centers$49.47BAll other office
Source: U.S. Census Bureau, Value of Construction Put in Place (C-30), latest revision (Nov 2026). All other office is total private office construction less data centers.
Private construction spending, seasonally adjusted annual rate. Source: U.S. Census Bureau, through May 2026.

The inflection is unmistakable. When ChatGPT launched in November 2022, the U.S. was building data centers at a $13.9 billion annual pace. As of May 2026, that figure is $59.3 billion — more than a fourfold increase in three and a half years. And in November 2025, a quiet milestone: for the first time since the Census began breaking out data centers in 2014, America began spending more on data center construction than on every other category of office construction combined.

$59.3B
Annual Pace Of Data Center
Construction · May 2026
4.3×
Growth Since ChatGPT
Launched · Nov 2022
2,025
Year Data Centers Passed
All Other Office Combined

Two honest caveats. Not all of this is strictly AI — broader cloud adoption and automation, especially since 2020, were already lifting demand for computing resources. And construction spending is an input, not a profit. But as a measure of how much real economic activity the AI theme is generating right now, it’s hard to find a cleaner one.

03 · The Debate

The central question: will hundreds of billions pay off?

Here is the debate at the heart of AI investing right now: will the enormous sums being poured into AI infrastructure — particularly by the largest technology companies — eventually generate sufficient returns? Demand for computing power has been substantial, supporting everyone who provides hardware and data center capacity. But as models improve, they also become more efficient, potentially needing less compute for a given task. That tension is a big reason AI-linked stocks have delivered their gains with stomach-churning swings.2

The Magnificent 7 beat the S&P 500, with much deeper drops along the way
Cumulative return, Dec 31, 2021 to Jul 17, 2026
Magnificent 7, equal weightS&P 500 (price index)
−50%0%+50%+100%+150%202220232024202520262022 lowMar 2026+142.46%Magnificent 7+56.47%S&P 500
−50%0%+50%+100%+150%2022202420262022 low+142.46%Magnificent 7+56.47%S&P 500
Magnificent 7: Meta, Amazon, Apple, Alphabet, Nvidia, Microsoft and Tesla, equal-weighted at the start and not rebalanced, dividends included. S&P 500 is the price index. Source: daily closing prices via Yahoo Finance; Bull Run calculations. Indexes are unmanaged and cannot be invested in directly. Past performance does not guarantee future results.
Equal investment in the Magnificent 7 vs. the S&P 500, Dec 31, 2021 – Jul 17, 2026. Magnificent 7 with dividends; S&P 500 price index.

Since the end of 2021, an equal investment in the seven mega-cap technology leaders has returned roughly +140%, against +56% for the S&P 500 — but the path included a 48% drawdown in 2022, the tariff selloff of April 2025, and this year’s AI-efficiency scare, a roughly 19% peak-to-trough decline that bottomed on March 30, 2026. Each downswing had the same shape: optimism about the buildout, followed by doubt about whether demand would fill it.

The efficiency worry deserves a direct answer, and history offers one. It’s called the Jevons paradox: when a technology becomes more efficient and its cost falls, total consumption often rises rather than falls, because cheaper capability unlocks uses that were never economical before. Electricity didn’t stop at light bulbs. Computers didn’t stop at corporations.

Why cheaper AI can mean more demand, not less
When a technology gets more efficient, total use often rises instead of falling. This is known as the Jevons paradox.
Models get moreefficientCost per task fallsNew uses become viableTotal demand for computerisesJevons paradox
Models get more efficient
Cost per task falls
New uses become viable
Total demand for compute rises
Back to the start
Electricity went beyond light bulbs. Computers went beyond corporations. AI is going beyond chatbots. Concept: W. S. Jevons (1865). Framework via Clearnomics. For illustration only.
Why efficiency gains have historically expanded demand instead of shrinking it.

The counterweight is just as important: markets have a long record of overestimating how quickly new technologies turn into profits, even when the long-term potential is entirely real. The internet enthusiasm of the late 1990s took decades to fully play out — and vaporized plenty of capital along the way for investors who confused the theme with any particular stock. The lesson isn’t to avoid the theme. It’s to hold it with a broad view of the chain and a genuinely long horizon.

04 · Valuations

Valuations reflect high expectations

As AI has captured investor attention, valuations across the technology complex have moved steadily higher. The Information Technology sector currently trades at 21.4× earnings — elevated against both its own history and the broader market — and sectors like Communication Services and Consumer Discretionary, home to several of the same mega-cap names, show a similar pattern.3 These multiples aren’t irrational: they sit on top of genuinely strong earnings growth as AI demand builds. But they are demanding, and they leave less room for disappointment.

Here’s the discipline we’d emphasize: valuations are not a timing tool. They tell you very little about what markets will do tomorrow. What they are useful for is deciding the appropriate mix of assets in a portfolio — where expectations are stretched, where they aren’t, and how much of each risk you want to own on the way to your actual financial goals. Plenty of other market segments carry attractive valuations with solid expected earnings growth of their own, which is exactly why the AI theme belongs inside a portfolio, not instead of one.

05 · Our Positioning

How we own the theme: the Equity Innovation SMA

Everything above is the market’s evidence. Here is our answer to it. The BRIM Equity Innovation SMA is our concentrated portfolio of what we believe are the fastest-growing companies across the AI, software, and cloud-infrastructure landscape — the supply chain in this note, expressed as an actively managed strategy. This is not a theme we discovered this quarter; it’s the strategy’s entire design, and it has a track record:

Equity Innovation vs. the Nasdaq-100 and S&P 500
Cumulative return before fees, Jan 1, 2023 to Jun 30, 2026
Equity Innovation, before feesNasdaq-100S&P 500
0%+100%+200%+300%+400%2023202420252026+338.31%Equity Innovation+182.18%Nasdaq-100+105.11%S&P 500
0%+100%+200%+300%+400%2023202420252026+338.31%Equity Innovation+182.18%Nasdaq-100+105.11%S&P 500
Source: Bull Run Master Database, daily. Before fees; after-fee returns are lower. Indexes are unmanaged and cannot be invested in directly. Past performance does not guarantee future results.
Cumulative return, gross of advisory fees, Jan 1, 2023 – Jun 30, 2026. Daily data from the BRIM Master Database; benchmarks gross.

And with Q2 in the books, here is the same record scored against the complete universe of non-leveraged ETFs — every fund in America with a full track record over each period, using YCharts total-return data. Through June 30, 2026:

YearInnovation SMAS&P 500Percentile RankFunds in Universe
2023+98.4%+26.3%Top 1.18%3,479
2024+39.3%+25.0%Top 3.10%4,061
2025+35.2%+17.9%Top 8.12%4,854
2026 YTD+17.1%+10.3%Top 17.34%5,969
Since Jan 2023+337.4%+105.2%Top 0.75%3,479

That cumulative +337.4% since January 2023 ranks #27 out of 3,479 non-leveraged ETFs — the top 0.75% of the universe. (All figures gross of advisory fees; 2026 is a partial year through June 30.)

Don’t take our word for it — verify it yourself, live →The full comparison table is public: every fund, every calendar year, every rank, powered by the same database that runs our client reporting. We believe in showing our work.
Equity Innovation against every non-leveraged ETF since January 2023
Before fees. Jan 1, 2023 to Jun 30, 2026; 2026 is year to date. Universe: 3,479 non-leveraged ETFs with a full record for the period.
+338.31%Equity Innovation cumulative return, before fees
Equity Innovation annualized52.60%
Rank among 3,479 non-leveraged ETFs#27 (top 0.75%)
S&P 500, same period+105.11%
Calendar-year returns
Equity Innovation, before feesS&P 500
0%20%40%60%80%100%98.63%26.29%202339.34%25.02%202435.29%17.88%202517.05%10.21%2026 YTD
0%25%50%75%100%2023202420252026 YTD
Source: Bull Run Master Database, daily, before fees; after-fee returns are lower. ETF total-return data via YCharts. Indexes are unmanaged and cannot be invested in directly. Past performance does not guarantee future results.
Gross of advisory fees; 2026 through June 30, 2026. ETF total-return data via YCharts.
Cumulative return of every non-leveraged ETF since January 2023
Before fees, Jan 1, 2023 to Jun 30, 2026. Each point on the curve is one fund, ranked from best (1) to worst (3,479).
+338.31%Equity Innovation, rank 27 of 3,479
Top 1% cutoff+297.9%
Top 5% cutoff+173.5%
Median fund+57.6%
S&P 500+105.11%
0%+250%+500%+750%+1,000%+1,250%11,0002,0003,000S&P 500Equity Innovation
0%+250%+500%+750%+1,000%+1,250%11,0002,0003,000S&P 500Equity Innovation
Source: Bull Run Master Database, daily, before fees. ETF total-return data via YCharts; the fund curve is traced from Bull Run's original chart. Past performance does not guarantee future results.
Each point on the curve is one fund. Gross of advisory fees, Jan 1, 2023 – Jun 30, 2026.

One thing we want to be clear about: a concentrated growth strategy is the highest-octane way to own this theme, and it comes with the drawdowns to match — our own 2026 round trip, chronicled in our July note, is Exhibit A. That’s why it lives alongside our Growth, Core, and Low Volatility strategies inside client portfolios sized to each household’s goals and risk tolerance — the AI theme inside a portfolio, not instead of one.

A note for financial advisors

All four BRIM equity SMAs — Innovation, Growth, Core, and Low Volatility — are available to outside advisors through the Charles Schwab Managed Account Marketplace. If your clients are asking how to own the AI theme without buying a grab-bag ETF, the live rankings table above is exactly what your due-diligence process wants to see. Reach out and I’ll walk you through it.

Download this Insight as a PDF

Schedule a Call Today