Welcome to note #2. I want to dig into some questions that keep coming up before energy shocks took over all our attention. First, is AI going to kill software? And second, are we in an AI bubble? The short answer to both is "probably not, but it's complicated," but the longer answer is more interesting.
The SaaS-pocalypse
There's a growing narrative that AI is going to eat the software industry. The logic is: if an AI can generate code, automate customer support, build workflows, and handle the tasks that SaaS companies charge monthly subscriptions for, then what exactly are you paying for? The thing that makes the SAAS companies valuable is the ability to offer a function or service better and cheaper than you can do it yourself. If that can now be replicated or replaced by an AI model, then their pricing power approaches zero. And once pricing power goes, all you're left with is a race to the bottom in terms of cost.
This is a pattern we've seen before. Every time a general-purpose technology matures, the companies that make the technologies that enabled it eventually become victims of commoditization. Quality equalizes, features converge, and price is all that matters. We saw this with memory, microprocessors and PCs. Whole new categories of products were created, but chips and board and computers became cheaper and interchangeable. We saw it with the telecoms that built the internet. Outside of high-end smartphones, we’re basically there with mobile phones.
But is AI really a replacement for SaaS? I'm not so sure. One critical distinction is: AI is not software. With traditional software, increased users don't materially increase resource demand. If you have more users for Office365, for instance, programmers aren’t in a factory banging out new copies of Excel. You build the product once and then revenue for additional users rises much faster than costs. With AI, it's fundamentally different. More users mean you’re using more “compute” for inference. Inference is basically how LLM’s do the things they do. From coding to helping with your wardrobe and planning vacations. More compute means more data center usage. Cost and revenue increase in lockstep. This kind of fixes your margin, which is very different than SaaS.
Also, lowering the barrier to doing more things in-house is lower doesn’t mean it’s zero. There are intangibles like subject matter expertise, as well as new features that take advantage of best practices and new regulations. I think that in some cases mega caps may find it cost effective to leverage AI to bring some things in house just like many of them self-insure. If you get big enough the cost savings eventually make it worth the trouble. But for the rest of the economy that ROI becomes less clear. AI may become a viable, even preferable, option for some, but it won't enjoy the same economics that makes software so lucrative. So, will AI kill software? I don’t think so. I think AI will do what services like Squarespace and Wix have done. It lowers the barrier to entry for people doing side hustles and passion projects. There are more websites because of services like this, but Fortune 500 companies aren’t building their sites on Squarespace. They are still hiring design firms and consultants.
The bubble question?
As a rule, bubbles are about investor psychology more than hard financial numbers. Think of tulip mania. In short, bubbles are like Fight Club. The first rule of bubbles is that no one thinks it’s a bubble. The fact that so many people are worried about there being a bubble generally argues against it. So that’s the pithy answer to the question. “Irrational Exuberence” by Shiller and “Manias, Panics, and Crashes: A History of Financial Crises” by Kindleburger are good books on the subject if you want to go down a rabbit hole.
A more complete answer requires looking at previous technological revolutions.
Container shipping has changed our lives. It enabled global supply chains, just-in-time inventory systems, changed the way we shop. Globalization would have been impossible without it. While building out container shipping infrastructure is the foundation, it is the companies that leveraged this innovation that became household names. The real winners were Walmart, Amazon, and especially China as it became the factory for the world.
The telecoms tell the same story, arguably more dramatically. In the five years after the Telecommunications Act of 1996, telecom companies invested more than $500 billion, mostly financed with debt, into laying fiber optic cable, adding switches, and building wireless networks. The growth in capacity vastly outstripped the growth in demand. Twenty-three telecom companies went bankrupt, capped by WorldCom — at the time the single largest bankruptcy in American history. By 2001, an estimated 95% of the fiber installed during the boom was "dark fiber." There simply wasn't enough internet traffic or revenue to justify the investment. Companies like Global Crossing, Qwest, and Level 3 Communications were devastated. Equipment makers like Cisco, Lucent, and Nortel fell hard from their peaks.
Sound familiar?
General purpose technologies follow an installation-then-deployment arc1. The powers during the installation phase rarely maintain leadership in the deployment phase. Your internet service provider isn’t front of mind when you’re online. It’s the content and app makers and hosting that lives on top. These are the second and third wave winners. YouTube, social media, and online banks are perfect examples of deployment phase winners.
I believe in the transformative power of AI. I also believe the financial projections being built around it have gotten out over their skis. These two things are not contradictory. AI infrastructure investment is expanding at a rate that bears little relation to actual economic uptake. What makes this potentially problematic is how the financing structure has shifted. During the early part of this “revolution”, hyperscalers funded most of the capex from their own cash flows. That created a natural speed limit on both growth and risk. As costs soared, liabilities are being pushed onto special purpose vehicles, infrastructure funds, private credit. Depreciation schedules, collateral values, and cash flow assumptions are yet to be validated. And the token economics are brutal. Token costs are falling more than 70% per year as efficiency improves.
Source: “The AI Bubble: Hidden Risks and Opportunities”, Man Group 2/2026
When I pay my annual subscription for Claude/Gemini/ChatGPT, I'm basically renting resource usage/priority so it can answer my questions. It does this via a process called inference. Inference uses tokens. Each input and output is made up of tokens (bits of text). The more complicated your questions and answers the more tokens you use. Output tokens are priced higher than input tokens. The price per token has fallen faster than the prices of chips and bandwidth from previous revolutions. Tokens are always getting cheaper as people figure out more efficient ways to train LLMs and use compute. Making things more efficient is good because the less it costs per token, the more tokens you can process for the same electricity bill.
At the same time, more data centers are being built. Enough are being built to be a significant driver of GDP growth. This further increases the pool of available tokens. So, we have a growing pool of tokens that are constantly getting cheaper. Here we come to the central question of AI economics. Is there enough demand to profitably sell all the tokens and is that demand growing fast enough to raise the price on an expanding supply? Right now, the answer to both questions is: ¯\_(ツ)_/¯.
So where does that leave us? I think this part of the cycle will eventually end. Is that months or years or decades? I have no idea, but my guess is that the real revolution hasn’t started yet. The companies that will leverage AI to change the way we live, and work will build on the foundations being laid down today. Just like the internet was built on the overbuilding of telecoms.
To sum up: I think AI will change the world, but I think it’s going to take a lot longer than the timelines people are speculating about now. Do I think the current investment and high valuations will eventually spell trouble? I do. When that happens is anybody’s guess. But we’re keeping watch.
1: Concept taken from Perez, C. (2002). Technological revolutions and financial capital: The dynamics of bubbles and golden ages. Edward Elgar Publishing