Forget the AI Bubble Talk. Watch This Number Instead
Listen to the audio version of this article (generated by AI).
Why everyone’s arguing about an AI bubble when they should be watching a price tag… what two Big Tech CEOs are begging for… watch this “escalator”…
Before we jump in today, a reminder about tomorrow morning’s Breakthrough 2026 event at 10 a.m. ET with legendary investor Louis Navellier and TradeSmith CEO, Keith Kaplan.
They’ll be walking through TradeSmith’s Seasonality tool. It scans more than 5,000 stocks across decades of price history, hunting for one thing…
Windows of time when a stock has historically gone up – or down – with remarkable consistency.
Run through an 18-year backtest, trading only inside these windows produced 857% in total growth – more than double the S&P 500 over the same stretch, and the strategy still came out ahead even in 2007, the worst year in the test.
When you sign up to join Louis and Keith, you’ll get free access to the Seasonality Tool. Give it a spin in your own portfolio.
Then, tomorrow morning, Keith will lay out why he believes the period beginning around July 23 could mark an important shift in market leadership – and how to use the tool to capitalize. That ties into how Louis is using these timing signals with his own stock-grading system.
There will be plenty more, including some free stock recommendations. Just click here, and we’ll see you tomorrow morning at 10 a.m. ET.
Everyone wants to debate whether AI is a bubble
That’s the wrong question.
The question that determines who wins and loses – and what happens in your portfolio – is far more mundane…
What does it cost to buy an AI token?
To make sure we’re on the same page, a token is the basic unit AI models get billed by – roughly a few characters of text. It’s the meter running every time a chatbot answers a question, an AI agent completes a task, or a piece of software calls a model behind the scenes.
Think of token prices the way airlines think about jet fuel or manufacturers think about steel. They’re the core input cost of the AI economy.
Today, those costs are falling for a simple reason…
AI models are becoming dramatically more efficient, while competition among providers – including a wave of open-source alternatives – is driving prices down. Every major AI lab is racing to deliver more intelligence for fewer dollars.
Whatever you think about valuations, this is the number that ultimately decides who wins and who loses inside the trade.
And right now, it’s collapsing.
The number that’s already crashed
In March 2023, running OpenAI’s best available model cost roughly $30 to process a million tokens – roughly the amount consumed by a lengthy AI conversation or thousands of simple prompts.
Today, comparable-quality performance runs a few cents to a couple of dollars – a decline of 90% or more in a little over three years and still falling.
Now, despite that collapse, total enterprise AI spending hasn’t fallen. By most accounts, it’s tripled. Companies are using it dramatically more because it’s finally affordable enough to deploy across the business.
Chatbots have given way to autonomous agents that loop, recheck their own work, and call external tools dozens of times to finish a single task. Every loop consumes tokens. So, even as the unit price craters, total usage is growing faster than the price is falling.
This is exactly why AI infrastructure – mostly chip and compute demand – hasn’t cracked yet.
Prices down, but overall spending up. That contradiction is the whole ballgame.
But however much prices have crashed so far, two powerful AI CEOs think it needs to fall dramatically further.
AI is still too expensive
Last week, Palo Alto Networks (PANW) CEO Nikesh Arora went on CNBC to say, in effect, that the current price of an AI token is holding back enterprise adoption.
Here’s Arora to explain:
I think 54% is a good start… I think we probably need another turn at it.
That 54% was a reference to OpenAI’s claim that its newest model is 54% more token-efficient than its last one.
Arora’s point: a nice start, nowhere close to enough. Later in the same interview, he said it more plainly:
We need to see the pricing for AI come down.
Palantir (PLTR) CEO Alex Karp went further the week before, calling the token-pricing model broken outright:
I’m not throwing shade at them, but something has gone completely wrong.
The basic view among enterprises in this country is I’m going to chillax and waste my time with tokens.
In other words, customers don’t want to think about tokens. They just want AI that’s cheap enough to use everywhere.
Two CEOs, running two companies supposedly winning during the AI boom, are publicly complaining that AI costs too much to use at scale. That’s not noise.
These are two massive AI customers telling you where the ceiling is today – and where things are going tomorrow.
What our own Eric Fry is watching
Our global macro expert, Eric Fry, editor of Fry’s Investment Report, has been tracking a specific driver behind that price pressure: competition from open-source models, including Chinese labs like Z.ai, deliberately tuned to run on older, cheaper chips rather than the newest ones.
They’ve been getting results nearly indistinguishable from Western models several months more advanced.
Here’s Eric to explain:
Token costs have come down around 20% since the start of June, reducing what data centers can charge for computing power.
Reflecting that trend, shares of data center company CoreWeave Inc. (CRWV) have fallen roughly 40% in the past two months.
I’ll note that CoreWeave’s slide has more than one storyline behind it – reports of Meta (META) building its own compute-for-rent business have gotten most of the mainstream press attention.
But Eric believes falling token prices are the underlying force behind that story – an alternative explanation.
Running against the escalator
Now, what’s the implication for investors?
Well, let’s understand the landscape first with an analogy – trying to run up a descending escalator.
Usage growth is like you climbing up that escalator. Falling token prices are the escalator moving down beneath your feet.
Right now, you’re climbing faster than the escalator descends, so you’re still making progress toward the top – total AI spending keeps rising. This is a win for AI infrastructure companies, and somewhat of a win for companies that want cheap AI.
But if the escalator speeds up (prices fall faster) or your legs tire (usage growth matures), the escalator wins, and you get carried down instead of up. Not a win for AI infrastructure companies, but a big win for companies that want cheap AI.
Right now, usage is dominating – it’s growing faster than prices are falling, which is why infrastructure demand stays strong even as the per-unit economics erode underneath it.
Our growth investing expert Louis Navellier, editor of Growth Investor, provided evidence of how real that “usage is winning” phase still is…
Yesterday, IBM (IBM) issued an unexpected preliminary Q2 earnings report and a profit warning, triggering the company’s worst single-day stock decline in its history.
But this wasn’t a problem for Louis. Here he is explaining why:
Earnings season is off to a very good start. I know International Business Machines missed, but they missed because they’re losing out market share to data centers.
So, that’s good for us because guess what we own?
Lots of data center-related stocks.
This is Louis – one of the best analysts in our industry – who knows exactly where the money is flowing today and is successfully running up the escalator.
But as Louis knows, and will eventually factor into his recommendations, this same escalator will ultimately win out.
In other words, at some point in the future, token prices will fall far enough, or usage growth will mature enough, that the balance will flip – and when it does, the AI trade will reach a key inflection point.
To join Louis in Growth Investor so you can navigate that transition with him, click here to learn more.
Here’s the big-picture version of who’s on each side when that happens
Exposed: the companies that built expensive, specialized infrastructure to rent out compute by the unit. We’re talking chipmakers, neocloud data-center operators, and any hyperscaler selling raw processing power, because their pricing power depends on scarcity.
When compute stops being scarce, that pricing power goes with it.
Helped: What Eric calls “AI Appliers” – companies that adopt AI as a tool inside their existing business rather than sell compute as the product, expanding their own margins every time their AI bill shrinks.
And the biggest pool of value sits one layer further out: ordinary enterprises across finance, healthcare, retail, and industry, whose AI costs turn from a budget headache into a rounding error, unlocking productivity growth that shows up as real earnings growth rather than a bigger tech bill.
Eric has already recommended specific AI Appliers in Fry’s Investment Report– companies positioned to catch that margin expansion before the rest of the market catches on. To discover what they are, click here to learn about joining him.
One category that I won’t even pretend to have a clean answer on: traditional software-as-a-service companies. These are the companies that suffered the “SaaSmageddon” earlier this year.
Cheap AI should, in theory, let legacy software platforms bolt on powerful features without blowing up their margins. But cheap AI also lowers the barrier for customers – or nimble competitors – to build capabilities that software vendors used to charge dearly for.
Whether SaaS incumbents end up net winners or casualties of this same price collapse is genuinely unresolved, and it likely depends on the specific company, not the sector.
We’re watching this one as closely and will keep you up to speed.
How to monitor all this
Nobody knows exactly when this escalator drama will hit its inflection point, and we’re not going to claim that we do.
Instead, watch whether the major AI labs start reporting expanding margins even as prices keep falling – like, for example, OpenAI reporting that its cost to serve a query dropped faster than the price it charges for one. That would mean efficiency gains are outrunning price cuts, a bad sign for hardware-heavy names like Nvidia (NVDA).
Also, watch whether the software companies riding on top of all this start actually showing cheaper AI in their own reported margins, not just their marketing.
The bottom line
The token collapse isn’t a one-time event, and it isn’t finished.
It’s already happened, it’s still happening, and by Arora’s own math, it needs to happen substantially more before AI adoption really breaks wide open.
What’s unresolved isn’t the direction – it’s how long usage will outrun the price collapse, and when it ultimately reverses, what will happen if the infrastructure trade is still priced for the demand side to win forever.
This is the issue that decides who wins and loses here. It’s not whether “AI is a bubble.” It’s not some abstract multiple on some chip stock everyone already argues about on TV.
This issue requires more analysis and work, which is exactly why it’s the one most people watching from the sidelines won’t wrestle with…
But it’s the one you must wrestle with if you have money in AI.
We’ll keep you updated.
Have a good evening,
Jeff Remsburg
