In short
Whether the “AI spending boom” (genAI data centers and GPUs) can generate attractive returns for investors.
Guest backgrounds
No external guests. Host is Brian Nowak, Morgan Stanley’s U.S. Internet Analyst.
Key claims
Morgan Stanley estimates major cloud providers will spend over $1.4T on AI infrastructure next year; compute capacity could quadruple from 2025 to 2028 to about 120 gigawatts. The firm’s bottom-up work suggests 25%–50% ROIC across three AI business models.
Notable examples
(1) Renting compute: ~30% ROIC for a large next-gen data center; scenarios ~low 20% to nearly 40%. (2) AI lab owning infrastructure: ~40%+ ROIC and ~75% incremental operating margin via API access. (3) AI developers renting infrastructure: ~25% post-tax return potential and ~30% incremental operating margin. Returns depend on token pricing and infrastructure efficiency (tokens per GPU per second).
Written by AI. May contain mistakes. Listen to the episode to check what was said.
Chapters
Tap a time to open that second in VOAI's Rapid Integration and Infrastructure Needs
0:17 to 1:10
Discussion on AI's quick integration into daily life and the necessary infrastructure for growth.
“It helps people write software, research purchases, automate work, find information, among a myriad of other use cases.”
Analyzing the First AI Business Model: Compute Power Rental
1:10 to 2:10
Exploration of renting compute power as an AI business model and its financial returns.
“We see paths to roughly 25 % to 50 % return on invested capital, or ROIC, across three emerging AI business models.”
AI Labs Owning Their Infrastructure
2:10 to 3:02
Insights on AI labs owning their infrastructure and implications for operating margins.
“So despite the enormous cost of building and capital being deployed for these facilities, we think the economics here are quite attractive.”
Factors Influencing AI ROI: Pricing and Efficiency
3:02 to 3:58
Factors affecting returns on investment in AI, including pricing strategies and efficiency.
“These returns on invested capital are impressive, but what determines whether these returns can actually materialize?”
Third Business Model: Renting Compute Infrastructure
3:58 to 4:36
Exploration of AI developers renting compute infrastructure and its economic implications.
“our base case still produces roughly a 30 % incremental operating margin and 25 % post-tax return potential.”
Conclusion: Is the AI Spending Worth It?
4:36 to 5:02
Wrap-up on whether the significant AI spending will yield economic returns.
“If you enjoy the show, please leave a review wherever you listen and share thoughts on the market with a friend or colleague today.”
Transcript
Automatic transcript. May contain errors.0:00Welcome to Thoughts on the Market. I'm Brian Nowak, Morgan Stanley's U.S. Internet Analyst. Today, can the enormous investment behind Gen.ai actually pay off? It's Wednesday, September 9th at 9 a.m. in New York. AI has moved quickly into everyday life. It helps people write software, research purchases, automate work, find information, among a myriad of other use cases. But we need an infrastructure buildout of extraordinary scale to support all of this activity and more activity to come. In all, we estimate that the major cloud providers are going to spend more than$1.4 trillion on this AI buildout next year alone.
0:45But compute capacity is potentially going to quadruple from 2025 to 2028, reaching roughly 120 gigawatts. But all of this spending has raised a lot of questions for investors. One of the most common questions is, what kind of return on invested capital can these companies earn from all of these trillions of dollars of data center infrastructure investment? Well, our bottom-up work points to encouraging answers to this question. We see paths to roughly 25 % to 50 % return on invested capital, or ROIC, across three emerging AI business models. Now, ROIC is a useful way of measuring whether investments pay off.
1:28Think of it as how much after-tax operating profit can be generated relative to the capital required in the first place. The first business model we've analyzed is renting compute power. This is the infrastructure layer of the AI economy. Cloud providers build data centers filled with advanced graphics processing units, or GPUs, and rent that compute capacity to customers. In our base case, a large next-generation data center can generate a return on invested capital of roughly 30%, simply renting AI compute power. And even if rental prices move around, our scenarios still produce returns ranging from low 20 % to nearly 40%.
2:10So despite the enormous cost of building and capital being deployed for these facilities, we think the economics here are quite attractive. The second business model we've analyzed is where an AI lab has their own model, and they also own their own infrastructure. They give access to their model through an API to consumers and enterprises who then build upon it. They utilize the model. In some cases, they build applications using that model that can be future sources of productivity or efficiency for the economy. In this scenario, we think the economics can be even stronger. When the model developer owns their own underlying infrastructure, our base case generates a roughly 75 % incremental operating margin and a return on invested capital of 40 % plus.
3:02These returns on invested capital are impressive, but what determines whether these returns can actually materialize? Well, two things matter a lot. The first is the price that developers are able to charge for tokens, which are the units of information that AI models process. The second factor that matters considerably is how efficient can this infrastructure process these tokens. This is why continued improvements in chips and software to drive higher token throughput or more tokens per GPU per second are critical to the long-term unit economics across this AI ecosystem. The third model we've analyzed is when the AI developers rent their compute infrastructure rather than owning it.
3:46So effectively, they are paying someone else for the data centers and the GPUs that they need. While this lowers their returns on invested capital because another provider takes a piece of the unit economics, our base case still produces roughly a 30 % incremental operating margin and 25 % post-tax return potential. So while the AI buildup requires enormous investment, the size of the spending alone doesn't tell the whole story about whether or not there are economic returns to come. What ultimately matters is the revenue and profit that the infrastructure can generate. And as more of the infrastructure shifts from training AI models to serving customers through emerging products and inference, we think we're going to get a much clearer answer to this question investors are asking today.
4:35Was all this spending worth it? Our research suggests yes. Thanks for listening. If you enjoy the show, please leave a review wherever you listen and share thoughts on the market with a friend or colleague today. The preceding content is informational only and based on information available when created. It is not an offer or solicitation, nor is it tax or legal advice. It does not consider your financial circumstances and objectives and may not be suitable for you. Thank you.
From the publisher
Big Tech is pouring more than $1.4 trillion into AI, prompting investors to ask: Is it worth it? Our U.S. Internet analyst Brian Nowak looks at three business models that could earn 25 to 50 percent returns for Gen-AI-enabled technologies.
Read more insights from Morgan Stanley.
----- Transcript -----
Brian Nowak: Welcome to Thoughts on the Market. I'm Brian Nowak, Morgan Stanley's U.S. Internet analyst.
Today, can the enormous investment behind Gen AI actually pay off?
It's Wednesday, September 9th, at 9am in New York.
AI has moved quickly into everyday life. It helps people write software, research purchases, automate work, find information, among myriad[s] of other use cases.
But we need an infrastructure build-out of extraordinary scale to support all of this activity and more activity to come.
In all, we estimate that the major cloud providers are going to spend more than $1.4 trillion on this AI build-out next year alone. But compute capacity is potentially going to quadruple from 2025 to 2028, reaching roughly 120 gigawatts.
But all of the spending has raised a lot of questions for investors. One of the most common questions is: What kind of return on invested capital can these companies earn from all of these trillions of dollars of data center infrastructure investment?
Well, our bottom-up work points to encouraging answers to this question.
We see paths to roughly 25 to 50 percent return on invested capital, or ROIC, across three emerging AI business models. Now, ROIC is a useful way of measuring whether investments pay off. Think of it as how much after-tax operating profit can be generated relative to the capital required in the first place.
The first business model we've analyzed is renting compute power. This is the infrastructure layer of the AI economy. Cloud providers build data centers filled with advanced graphics processing units, or GPUs, and rent that compute capacity to customers. In our base case, a large next-generation data center can generate a return on invested capital of roughly 30 percent simply renting AI compute power.
And even if rental prices move around, our scenarios still produce returns ranging from low 20s percent to nearly 40 percent. So, despite the enormous cost of building and capital being deployed for these facilities, we think the economics here are quite attractive.
The second business model we've analyzed is where an AI lab has their own model, and they also own their own infrastructure. They give access to their model through an API to consumers and enterprises who then build upon it, they utilize the model. In some cases, they build applications using that model that can be future sources of productivity or efficiency for the economy.
In this scenario, we think the economics can be even stronger. When the model developer owns their own underlying infrastructure, our base case generates a roughly 75 percent incremental operating margin and a return on invested capital of 40 percent plus.
These returns on invested capital are impressive, but what determines whether these returns can actually materialize?
Well, two things matter a lot. The first is the price the developers are able to charge for tokens, which are the units of information that AI models process. The second factor that matters considerably is how efficient[ly] can this infrastructure process these tokens.
This is why continued improvements in chips and software to drive higher token throughput – or more tokens per GPU per second – are critical to the long-term unit economics across this AI ecosystem.
The third model we've analyzed is when the AI developers rent their compute infrastructure rather than owning it. So, effectively, they are paying someone else for the data centers and the GPUs that they need. While this lowers their returns on invested capital because another provider takes a piece of the unit economics, our base case still produces roughly a 30 percent incremental operating margin and 25 percent post-tax return potential.
So, while the AI build-out requires enormous investment, the size of the spending alone doesn't tell the whole story about whether or not there are economic returns to come.
What ultimately matters is the revenue and profit that the infrastructure can generate. And as more of the infrastructure shifts from training AI models to serving customers through emerging products and inference, we think we're going to get a much clearer answer to this question investors are asking today.
Was all this spending worth it? Our research suggests: Yes.
Thanks for listening. If you enjoy the show, please leave a review wherever you listen and share Thoughts on the Market with a friend or colleague today.