If you had told a trader in 2022 that by mid-decade, artificial intelligence would be both the single most important driver of equity valuations and the source of endless debates at the tee box and pickleball courts, you would have been accused of hype-chasing. Yet here we are. AI has gone from a story to a theme and from a theme to becoming a hard-wired component of the market’s reptilian brain. It’s not just venture capitalists anymore; it’s your neighbour trying out ChatGPT on their taxes, your kids pestering you about AI homework tools, your dentist casually mentioning NVIDIA. Few technologies since the smartphone have crossed the cultural–financial divide this quickly. The early acceleration of AI was breathtaking — each new model a leap forward, and each model demanding higher compute power. But speed has a way of sowing doubt.

On August 12th, the Solactive MAG index (which tracks the “magnificent 7” stocks) jumped more than 1% to a new record, as the chart above shows. It felt like last year, when the index bounced off its low near 16,000 in August to rally to close to 24,000 in December, but this rebound from a similar low in April (just below 16,000) surpassed that gain and took the MAG to a new record high north of 24,600. In the past week, cold water has been thrown on this fire and we have seen a retreat back to the 24,500 area. As the dust settles, we find ourselves at a curious stage - the AI plateau. Not a collapse and perhaps not even a stall, but a flattening of the trajectory.

As with any market story, there is more than one angle and the same is true here. On the surface, we have an explosive rally that has been fueled by stronger-than-expected Q2 earnings, a rekindled belief in rate cuts and a desensitization to tariffs, which I reviewed on this week’s conference call (click here for the recording link). Considering that September and October are just around the corner, it would be natural that some market participants would want to take some money off the table, rather than potentially lose a lot of the post-April gains. In this case, however, it is the AI sector that has punctured the balloon – or shall we say “bubble”.

Last week, Open AI CEO Sam Altman openly responded in the affirmative to a question as to whether he believed that AI is in a bubble. He remarked that investors are 'overexcited,' valuations of certain startups are 'insane,' and warned that 'someone’s gonna get burned.' In a Washington Post op-ed, he also suggested that an underwhelming GPT-5 release, combined with public skepticism, amounts to an unofficial slowdown. The piece warned that political backlash over jobs, energy, and water use could derail AI expansion without stronger governance and infrastructure. Now, in the same breath, Altman also reiterated OpenAI’s expectation to spend trillions of dollars on data centers in the near future—underscoring both the ambition and cost pressures facing the industry.

If that was it, then we would think the markets would take a momentary pause, yawn and then move. The problem is that there were other anecdotal conjectures made. Meta has temporarily paused AI hiring after a massive recruitment wave. And a recent MIT/NANDA study found that about 95% of enterprise generative AI pilots are not delivering measurable returns. This evidence of poor profitability has led many analysts to suggest a step-back phase is underway. The Financial Times has raised the question of whether AI is 'hitting a wall,' highlighting diminishing returns to scaling and underwhelming results from GPT-5.

The marginal leap from GPT-3 to GPT-4 was extraordinary; the leap to GPT-5 less so. The data is larger, the models more polished, but the wow factor is more muted. It feels like AI has read the library cover to cover, and while it can summarize with elegance, it’s not yet living in the world it describes. That presents a problem in terms of the how firms can monetize the AI platforms they are buying. In other words, can they make money or at least the money they and the analyst community thought probable?

This is how the market has tended to think of monetization. Use AI to create something that will make a company more efficient and lead to greater profitability. Peers in that sector will be forced to adopt the same practices or face extinction. It is an artificial Darwinism if you like. Yet, monetization also applies to the engines of AI landscape, from the power that feeds a data centre, constructed from concrete, steel, copper and fibre, to the chips, processors and racks.

When I was young, I would travel to England with my parents and among the many differences I observed between life there and in Canada was nestled in the pay phone booth. Back then, we could put a dime (soon to become a quarter) in the phone here and talk for as long as we wanted. Over there, I remember the frustration on my parents’ faces as they scrambled to shove coins into the box to keep a call going. AI is sort of the same, but instead of coins we refer to tokens. For every query and for every response, there is a cost.

Now, I don’t want to get too complicated in this commentary so let’s think of tokens as merely pieces of words that go into and come out of a large language model, the likes of which underpin ChatGPT. On average, 1 token is roughly equivalent to 4 characters of English text. For example, “Wealth management” is 2 words but represents 3 tokens (‘Wealth’, ‘ manage’, ‘ment’). Each model out there will have a maximum token limit per request and the longer the conversations and drafts, the more tokens are consumed. If you ever wonder why there is a free Chat GPT app and a monthly subscription, this is it. In fact, for OpenAI’s GPT 4 model, the approximate price for 1,000 input tokens is 3 cents. The cost for 1,000 output tokens is roughly 6 cents.

At this level, the monetization is simply charging for access to a model, which then becomes an input cost for the next level. For companies that require moderate access to models, then the external application providers like OpenAI and Anthropic are good solutions, though usage will see costs scaled linearly with token usage. For firms that will need heavier lifting, these costs could spiral to the point where developing in-house models might become more cost effective over the long term.

Ok, let’s step back from the token weeds for a moment. Once you can increase capacity and supply for an item, it will tend to lower the cost of that item. Invest billions into the infrastructure behind AI models and you will end up bringing token prices down and that’s happening today. If you are the end user, this is a great thing. If you are selling tokens, then the lower prices for the things you are selling against the cash flow drain from capital expenditures could eventually lead to margin compression and a less spectacular income statement. And this brings us back to our investment decisions.

In terms of the “squeeze factor”, the most vulnerable in this environment would be those companies that are selling AI features that are bolted on to existing applications. A number of companies, including Salesforce, ServiceNow and Adobe have come out with premium pricing for AI add-ons. As token prices fall, there will be pressure on them to reduce prices.

The next level affected would include the foundation model providers, like OpenAI, Anthropic and Cohere. Companies in this space are competing head-to-head on token pricing and unless they effectively move up into applications or vertical integrations, their margins could deteriorate. You can operate a chip truck and flip burgers with razon-thin margins every day, but that sounds like too much work.

The better-protected companies in our opinion are what are referred to as the cloud hyperscalers.  Think of Microsoft, Alphabet (Google), Amazon and Oracle. These firms actually control the infrastructure and make money through storage, processor rental and networking. Even if token prices fall, presumably demand for AI increases, resulting in stronger traffic.

And that leaves us with the fourth segment of the market which makes the guts of the system. Here we find Nvidia, Broadcom, Taiwan Semiconductor, Micron and Eaton to name a few. These companies build the 'picks and shovels' of the AI boom—GPUs, chip packaging, HBM memory, networking, and power systems. Token pricing wars do not affect them directly. In fact, cheaper tokens can drive more AI usage, which fuels demand for hardware. That’s not to say they have no risk. If hyperscaler capex slows, then demand for picks and shovels goes down and regardless of token price compression, there will be likely be a negative influence on stock valuations.

As we have said many times, when the buyer of your product doesn’t show up at your store, you might have to cut costs to maintain margins. This is definitely a macro theme facing the U.S. and other parts of the globe, but it applies to AI. Stocks in all four categories of the AI space that we discussed above are vulnerable to demand disappointment, over and above the commentaries from executives and journalists. And yes, that might also apply to the fifth category that we didn’t talk about, which is power generation (utilities, natural gas, LNG suppliers, etc). The recent setback in the AI space could still morph into a more substantive correction in September and October, but the long-term trend for AI adoption tells us that this is probably just a token pause and not a secular retreat.

On behalf of the Pyle Wealth Advisory team, have a wonderful weekend.   

Andrew Pyle