Chapter 1
When I started learning about AI, I assumed companies like ChatGPT and Claude were the AI industry. After spending time reading earnings calls, investor presentations, and industry reports, I realized they’re really just the products consumers interact with. Behind every AI application is an enormous amount of infrastructure, capital, and electricity that most people never see.
One of my goals is to build my understanding from the ground up. Rather than memorizing facts about individual companies, I want to understand how the entire ecosystem works, who creates value, who captures value, and where the biggest opportunities and risks exist.
The AI Value Chain
The easiest way for me to think about the AI industry is as a value chain. Every layer depends on the layer below it.
At the very top are the AI applications people actually use. Products like ChatGPT, Claude, Gamma, Perplexity, and hundreds of other AI tools help consumers and businesses become more productive. These are the products people subscribe to and interact with every day.
Those applications rely on foundation models, also known as large language models (LLMs). These models are trained on enormous amounts of information from publicly available sources, licensed data, and human feedback. Once trained, they generate responses through a process called inference, predicting one token at a time until a complete answer is produced.
One concept that initially confused me was tokens. The simplest way I’ve found to think about them is that they’re small pieces of text. Sometimes a token is an entire word, sometimes it’s only part of a word, and sometimes it’s punctuation. During inference, the model predicts the next most likely token based on everything that came before it. It repeats that process thousands of times in just a few seconds.
These models don’t operate on someone’s laptop. They run inside cloud infrastructure provided by companies like Microsoft Azure, Oracle Cloud, Google Cloud, and Amazon Web Services.
Supporting the cloud are massive data centers filled with servers, GPUs, networking equipment, storage systems, and cooling infrastructure. These facilities consume enormous amounts of electricity and operate around the clock.
At the heart of those data centers are GPUs, where NVIDIA currently has one of the strongest competitive positions in the industry. Although NVIDIA designs these chips, companies like TSMC manufacture many of them, while dozens of other companies provide memory, networking equipment, advanced packaging, and semiconductor manufacturing equipment.
Finally, everything ultimately depends on electricity. Without reliable and affordable power, none of the higher layers of the AI value chain can function. As AI demand continues to grow, electricity is becoming one of the industry’s biggest constraints.
Who Pays Whom?
One of the most useful frameworks I’ve learned is to follow the money.
Consumers and businesses subscribe to AI applications like ChatGPT and Claude because those tools create value for them.
The AI application companies then spend billions of dollars training foundation models and renting computing capacity from cloud providers.
Cloud providers invest heavily in building and operating data centers capable of handling enormous AI workloads.
Data center operators purchase GPUs, networking equipment, storage, and servers from companies such as NVIDIA, Broadcom, and many other suppliers.
NVIDIA designs its GPUs and relies on manufacturing partners like TSMC to fabricate the chips.
Finally, everyone in the chain depends on utilities to provide reliable electricity.
Following the money helped me realize that AI isn’t simply a software industry. It’s also a semiconductor industry, a cloud computing industry, a construction industry, and increasingly an electricity industry.
Why Power Has Become the Biggest Bottleneck
One of the biggest surprises during my research was realizing how important electricity has become.
Modern AI data centers require enormous amounts of power, with some campuses expected to consume more than a gigawatt of electricity. Unlike many industrial facilities, these loads operate nearly 24 hours a day, seven days a week.
Providing that level of reliable electricity isn’t easy.
Utilities must build additional generation, expand transmission infrastructure, and often invest in substations before many of these facilities can even be connected to the grid.
At the same time, communities and regulators are asking an important question:
Who should pay for this infrastructure?
If utilities recover these costs through higher electric rates, homeowners and businesses could end up subsidizing AI infrastructure. That has turned data center development into both an economic and political issue.
Power is no longer just another operating expense. In many cases, it has become one of the primary constraints on AI growth.
Who Has Pricing Power?
From everything I’ve read so far, the companies with the greatest pricing power today are those controlling the industry’s biggest bottlenecks.
NVIDIA is the obvious example.
Demand for advanced GPUs continues to exceed supply, giving NVIDIA significant pricing leverage. While competitors are emerging, NVIDIA’s combination of hardware, software, and ecosystem creates a substantial competitive advantage.
Electricity may become another area of pricing power.
If demand for power continues growing faster than supply, utilities and power producers could find themselves in an increasingly favorable position. That said, regulation will likely limit how much of that pricing power they can actually capture.
The key lesson I’ve taken away is simple: companies controlling bottlenecks often earn the highest returns.
The Strongest Bull Case
The bull case assumes AI adoption continues expanding across nearly every industry.
Consumers become more productive.
Businesses automate more work.
Developers build thousands of new AI applications.
Robotics, autonomous systems, and AI-enabled consumer products create entirely new markets.
If that happens, demand for computing power, cloud infrastructure, semiconductors, networking equipment, and electricity continues growing for years.
One thing that stands out to me is that almost every company involved in the AI ecosystem is investing as though demand will remain extremely strong for at least the next three to five years.
If they’re right, today’s infrastructure buildout may only be the beginning.
The Strongest Bear Case
The bear case isn’t necessarily that AI fails.
It’s that expectations become too optimistic.
The industry is investing hundreds of billions of dollars before anyone knows exactly what long-term demand will look like.
The biggest risks I currently see are:
- Data centers take longer to build than expected.
- Utilities can’t provide enough electricity.
- Power becomes significantly more expensive.
- Returns on invested capital fail to justify the enormous capital expenditures.
There’s also the question of pricing.
Today, many consumers happily pay around $20 per month for AI tools.
But what happens if that price doubles?
Demand for AI certainly isn’t perfectly inelastic. At some point, higher prices could slow adoption, particularly if businesses don’t generate enough productivity gains to justify the added cost.
Who Is Making Money Today?
Today, NVIDIA appears to be one of the largest financial beneficiaries of the AI boom.
Cloud providers are also generating significant revenue as demand for computing capacity continues to increase.
The companies spending the most money, however, are often the ones building AI infrastructure.
Foundation model developers, hyperscalers, and data center operators are investing hundreds of billions of dollars before knowing exactly what long-term returns will be.
That’s why capital allocation will be one of the most important themes to watch over the next several years.
Who Bears the Greatest Financial Risk?
At this point, I believe the greatest financial risk sits with companies building large AI infrastructure projects.
These businesses must make enormous capital investments today based on assumptions about future demand.
They need:
- Reliable access to GPUs.
- Affordable electricity.
- Sufficient financing.
- Strong customer demand.
- Returns that exceed their cost of capital.
If any one of those assumptions proves incorrect, returns could fall well below expectations.
What I’ll Be Watching Over the Next 12 Months
There are several themes I’ll continue monitoring.
First, I want to see whether demand for AI services continues growing at its current pace.
Second, I’ll watch data center construction and utility infrastructure to determine whether power availability becomes an even larger constraint.
Third, I’ll monitor GPU supply and pricing to see whether NVIDIA’s competitive position changes.
Finally, I’ll pay close attention to whether companies continue earning returns above their cost of capital despite the massive investments they’re making today.
Three Questions I Still Don’t Understand
One thing I’ve learned over the past week is that understanding an industry often creates even better questions.
Right now, these are the three areas I want to explore further:
- Where does each company create value—and where does it assume the most financial risk? Understanding the value chain is only the first step. I now want to understand the economics behind each participant.
- Will these massive investments earn returns above the companies’ weighted average cost of capital? Billions of dollars are being invested today based on expectations years into the future. I want to understand whether those investments are likely to create shareholder value.
- What happens if demand doesn’t meet today’s expectations? Every company seems to be planning for rapid AI adoption. If growth slows, who is most exposed? Who has the strongest balance sheet? Who can adapt?
These questions will shape much of what I study over the coming months.
My Biggest Takeaway
The biggest lesson from this first week is that AI isn’t just about software.
It’s an interconnected ecosystem built on semiconductors, cloud infrastructure, data centers, networking, utilities, and enormous amounts of capital.
The applications get most of the attention, but the infrastructure underneath them is what makes everything possible.
As an investor, that’s where I believe some of the most interesting opportunities—and risks—will be found.