Why growth is a problem for AI companies

AI does not scale like a software business, or a hit song

Author

George Pillari
March 06, 2025

A note about comments. I want to hear your comments. Unfortunately, email and web platforms force subscribers to go through a clumsy login and verification process when you want to leave a comment. It is all in the interest of reducing spam. So, please reply to this newsletter or email me at [email protected] with any of your thoughts or suggestions.

A typical rock song requires four takes to get it right. When Queen recorded Bohemian Rhapsody in 1975, it needed 180 takes to produce a final version. The group spent three weeks working in five different studios across England to create one of the most complex rock songs ever.

Studio costs were about $250,000 in today’s dollars and let’s add another $250,000 for ancillary personnel and other costs, which brings the total cost to create Bohemian Rhapsody to about $500,000.

While precise numbers are difficult to track, it is generally agreed that Bohemian Rhapsody is the rock single with the largest gross sales ever, with hundreds of millions of revenue. Add to that the $1.0 billion in sales from the Bohemian Rhapsody movie released in 2018 and it’s safe to say that this song was (and still is) an economic monster.

Movie poster from Bohemian Rhapsody

The financial beauty of a hit song is because of one key factor: it scales. When we say that something “scales,” what is it we are actually talking about?

The hit song required significant upfront investment risk but once the song was recorded, it could be burned on to CDs, downloaded, or streamed (forever) at almost $0 of incremental cost for each additional sale beyond that first sale – which required the initial $500,000 investment.

❝

This is the magic of a business model that scales: if you get it right and build a product that requires low incremental costs for each sale beyond that first sale, you really have built a better mousetrap. And a mousetrap that will attract capital at extraordinary valuations.

Ever hear of Microsoft Office or Office365? With this product, Microsoft created a hit song. There were tens of millions invested by the company to produce the iconic product that is primarily the bundle of Word, Excel, and PowerPoint.

Office generates about $70 billion in annual revenues and is used by more than 1.5 billion people worldwide. Across the whole company, Microsoft had $250 billion of revenues last year and more than $100 billion of operating profits.

A hit song if there ever was one.

With the glory days of Office in the rearview mirror, Microsoft, along with Google, Meta, Amazon, Nvidia and a few others are currently locked in a battle for AI infrastructure supremacy.

They each want to be the company that AI application providers like OpenAI and Anthropic use for running their massive AI models.

Google, Apple, Microsoft, Amazon, Meta, Nvidia. Couirtesy: Data Gravity

Popular media has dubbed Microsoft and their brethren as Hyperscalers since they are trying to rapidly increase the size and scale of their AI capabilities to keep up with projected demand from OpenAI and the application providers.

I believe the Hyperscaler tag is erroneous, and I have referred to these companies collectively as Big AI.

With the AI “boom,” Big AI is trying to convince people that as demand for AI-driven services continues grow, there will be a need for bigger and bigger data centers with more and more computing power. Most technologies do scale to some degree (eg, Microsoft Office), but AI does not scale for a couple of reasons.

When we think of a software application like Office, it is tightly defined and does not really change except for incremental upgrades and bug fixes. It is not unlike that recording of a hit song. Make it once, sell it a million times.

The problem is that AI does not fit this model. For example, today there are hundreds of millions of ChatGPT users creating billions of unique queries per month.

Every one of those queries is different and requires a significant amount of computing power (referred to as “compute”) to generate a response.

While it is true that the AI models learn from each query and become smarter, not much related to each query is repeatable or reusable, while, on the other hand, each sale of Office or each streaming of Bohemian Rhapsody requires almost no compute and does need the attention of teams of highly paid engineers working on the cutting edge of science.

Each query to an AI model is like the band going back into the studio to do another take.

This is Big AI’s problem. They are all-in with massive commitments to build out the infrastructure (in other words, to make enough compute available) for those billions of queries every month.

And the compute is not cheap. OpenAI and others running applications like ChatGPT on all of this infrastructure are running on the proverbial hamster wheel. Each query costs them much more than the revenue the query generated. The more popular OpenAI’s ChatGPT gets, the more money it loses.

Estimates peg OpenAI’s losses at $5 billion for 2024, $10+ billion for 2025, and $14 billion for 2026. At some point, buying more and more compute does not compute.

Courtesy: Techrights

Unless there is a major shift in the laws of physics that drastically reduces the cost of compute, the application providers will need to pivot to a business model that reduces reliance on Microsoft et al and does not use more and more compute as more and more people use their applications.

The AI crowd understands that growth is good. It attracts capital, especially when the promise of future profitability is plausible. Remember Amazon? It went nine years before turning a profit while burning through billions. Investors who bought the story and hung in there were rewarded with big returns.

If the “AI revolution” will take even five years, there is not enough venture capital and private equity to fund the outrageous losses of these aspirational business models.

AI has an even a bigger problem than the inability to scale. What happens when a cover band like China-based DeepSeek can produce an application that does not need the massive infrastructure created by Big AI?

Key Takeaways

  • Just because a business attracts a lot of capital and is growing, without profitability in the near term, its a wing and a prayer. I don’t know about you, but I have seen more than my share of business plan projections that suddenly become profitable in year 5, after blowing millions of investors’ dollars.

  • There is not enough news to satiate all of today’s media outlets. Any story that can dazzle attracts every blog, newscast, podcast, etc. and they all just repeat what the others are saying. Type “Hyperscalers” into a Google search and see what I mean.

  • Big AI is still in search of a use case for all that capacity it plans to bring online. It will be interesting to watch how these technologies will be repurposed as the AI boom fades and reality sets in.

Things I think about

A Boeing 777 aircraft has 150 miles of wiring running through its fuselage.

Fooled by Randomness
The first in Caleb’s series and the “Black Swan” book. All his books are excellent.

Secrets
First-person account of the release of the Pentagon Papers. Well written story about what it’s really like to be a DC insider.

OpenAI is a Bad Business
Ed Zitron is a technology contrarian and not afraid of, well, anything.

Launch Key
Weekly newsletter full of wisdom on how to launch a business.

The Psychology of Money
Lessons on money and life. I have given this book to a dozen people.

Hard Drive
Bill Gates and the making of Microsoft.

 Full list of recommended reading is here.

Reply

or to participate.