About That AI Bubble

The AI Sandwich

One of the more popular pastimes in financial markets these days is worrying about the AI bubble. The specter of 2000 hangs over every conversation. The valuations are dizzying, the revenue is often speculative, and the whole thing can feel a bit ephemeral. Are we just bidding up the price of chatbots and abstract digital intelligence? Is there anything real holding it all up?

The worry is not unfounded. As John Thornhill recently pointed out in the Financial Times, the current boom has all the hallmarks of a classic speculative frenzy built on top of a legitimate industrial transformation:

Is the privately owned OpenAI, on course to burn $8.5bn of cash this year, worth $500bn...? Does it make sense that the data and AI company Palantir is trading on a forward price/earnings multiple of 225...? At best, such valuations embrace defiantly heroic assumptions about the long-term earning power of these companies. At worst, they resemble expensive lottery tickets on the future.

One way to check is to ignore the frothy filling for a moment and look at the two big, boring pieces of bread that make up the AI sandwich. On the bottom, you have the physical manufacturing layer—the actual silicon being forged in factories. On the top, you have the application layer—the point where a real business pays a software company to use this stuff to do a real job. If both slices of bread are solid, the stuff in the middle is probably more than just air.

And this week, both slices looked very solid indeed.

Let's start with the bottom slice: TSMC, the Taiwanese foundry that is the physical manifestation of the AI boom. They are the ones actually making the impossibly complex chips for Nvidia and Apple. On Thursday, TSMC didn't just report a record quarterly profit; it raised its entire annual revenue forecast, citing "robust demand" from its top AI customers. This is the direct financial result of real companies paying them billions of dollars, right now, to make tangible things. When the world's most critical factory is running at full tilt, it's a pretty good sign that the investment is real.

Now, one could argue that this just proves the spending is real, not that the spending is rational. Shovel sellers always get rich in a gold rush, regardless of how much gold is actually found. This is where the frothy bubble concern lives: what if all this capital expenditure is just being lit on fire by companies chasing a technological mirage with no real business model?

And that's what makes the top slice of bread—Salesforce—so important. For months, the story that terrified software investors went like this: the business of enterprise software has always been about selling seats. You hire a sales team, you buy 100 Salesforce licenses. Simple. But agentic AI threatens to wreck that per-seat model. Why pay for 100 seats when a single AI agent can just... achieve the outcome?

This wasn't just a fear of replacement; it was a fear about the logical endgame of AI itself. If you take the AI hype to its conclusion—that agentic AI will one day achieve true human parity—then what, exactly, is a software company selling? You're no longer selling a better tool for your sales team; you're selling their replacement. And you can't charge a per-seat license for a digital agent that just gets the job done.

This is why the announcement by Salesforce on Wednesday was so significant. Instead of being cannibalized, the company projected its revenues would exceed $60 billion by 2030—a huge leap from its roughly $35 billion in 2024—driven by the integration of AI into its existing products. The market seemed to breathe a sigh of relief. This signals that real-world businesses aren't firing their software providers; they're paying them more to embed AI into their existing workflows, proving that there is a profitable business model for AI in the enterprise software space. Salesforce is so confident in this future that it's buying back $7 billion in stock and making an $8 billion acquisition to bolster its AI capabilities.

So you have the foundational layer of the AI economy reporting record-breaking demand, and you have a key player at the final application layer forecasting a new era of AI-driven growth. If the people making the shovels are selling out and the people at the end of the line are successfully selling gold-plated digging services, it suggests there's probably a lot of gold in that hill.

Which doesn't mean there isn't a bubble, of course. It just suggests the bubble is being inflated with some very real, very expensive ingredients.

The Friday Download

Connecting the micro to the macro – the core of our analysis inside the ARPU Intelligence Desk

Software's AI Execution Phase

• The Trend: The enterprise AI race is shifting from hype to execution, with software giants like Salesforce proving tangible $100 million ROI from its own tools (Bloomberg), partnering with multiple model providers like OpenAI and Anthropic (Reuters), and Workday investing $200 million in dedicated AI centers focused on regulatory compliance (WSJ).
• ARPU's Take: These moves reveal a market that has matured beyond "what if" to "how to". Salesforce's $100 million savings announcement is a direct sales tool to prove AI's ROI, not just promise it. At the same time, its partnership strategy shows it's playing the role of an agnostic enterprise platform—a "Switzerland for AI models"—rather than a core model builder. Workday's investment in a compliance-focused hub in Ireland demonstrates the next phase: building AI that can navigate a complex global regulatory landscape.
• The Implication: The new bar for enterprise software is no longer just having an "AI feature". Competitive advantage will now be defined by three things: a clear, quantifiable ROI story; an open, model-agnostic platform that gives customers choice and future-proofs their investment; and a built-in "trust layer" that guarantees compliance with complex global regulations like the EU's AI Act. Pure-play model providers will win on performance, but enterprise software leaders will win on integration, trust, and business value.

Semiconductor's AI Supercycle Validation

• The Trend: The AI boom is accelerating, confirmed by bellwether foundry TSMC raising its revenue forecast on a demand "megatrend" (Bloomberg) and memory giant Samsung reporting its highest profit in three years on a surge in HBM chip sales (KED Global).
• ARPU's Take: These are synchronized signals from two critical, distinct parts of the semiconductor supply chain. TSMC's forecast is the ultimate forward-looking indicator for the chips that power AI. Samsung's results are a concrete validation that the demand for the specialized memory that supports those chips is equally explosive. When the foundry and the memory supplier both signal an accelerating boom, the supercycle is increasingly becoming a reality.
• The Implication: The debate over a potential AI bubble is being sidelined by tangible, massive order books and capital expenditure plans. This validates enormous spending on new fabs, solidifies the market power of foundational players like TSMC and the Samsung/SK Hynix memory duopoly, and makes the cost of competing at the leading edge almost insurmountable for anyone else.

These are select insights from the ARPU Intelligence Desk. The full platform provides the continuous signal tracking and comparative analysis required for professional work.

If this level of intelligence is a fit for your needs, we invite you to request a brief demo.