A trader works on the floor at the NYSE, July 14.Brendan McDermid/Reuters
It seems that all investors worry about these days is whether there is an AI-driven bubble in the U.S. market, and if so when it will burst. But the more interesting question is, assuming this is the case, who among the AI companies are going to be the winners and who will be the losers? How can you avoid being the greater fool at the top of it?
Is there a way to determine this? We think there is.
Think of AI not as a product but as a stack of dependent layers, the way the early internet had its layered map (from fibre cables and Ethernet cables, to TCP-IP protocols, to HTTP for websites and SMTP for e-mails).
At the bottom of the AI stack sits infrastructure: power, the electrical grid, cooling, land, and critical minerals. Above it, the chips that turn electricity into computation. Above those, the data that AI models learn from. Above that, the models themselves, which produce raw intelligence. Then agentic execution: agents and robots turning AI into action. And at the top, the applications people actually touch, such as ChatGPT or Claude Code.
It is the layers in the middle of the stack – the models and the brand-name labs that have absorbed the most capital and the most attention – where durable value is hardest to defend. The price of running a model of a given capability has fallen roughly a thousandfold in a few years.
Opinion: What can Canada do as a future looms of AI dependence on the U.S.?
Two forces make this middle layer worse than it looks.
The first is that the frontier is increasingly being legislated into a corner. The most capable systems, including Anthropic’s Claude Mythos 5 and Claude Fable 5, and OpenAI’s GPT-5.6, are running into a thicket of liability and safety rules, as we witnessed a few weeks ago when the U.S. State Department briefly shut down Mythos and Fable by making it illegal to sell those models to non-Americans. These restrictions make it commercially difficult for companies to deploy their best work at all. A model you are not allowed to sell at full capability cannot earn its cost of capital.
The second is that several free open-source models, the most capable of which were developed in China, now have capabilities similar to those of expensive models. For a large and growing share of users, the free option is good enough. When a near-substitute for your product can be downloaded at zero marginal cost, your pricing power is capped no matter how brilliant your best engineers are.
Now, not all the frontier labs can fail. A model that lives inside a diversified giant – the way Gemini sits inside Google or Grok sits inside a constellation of Elon Musk’s balance sheets – can be subsidized through any amount of price deflation.
A pure-play frontier lab has no such cushion. And it is precisely the pure-plays – the ones whose names are synonymous with the boom and who are now the leading candidates to go public through IPOs – that are the most financially fragile. They are single-product companies selling a commoditizing product into a market with free open-source models, financed by recurring capital needs that assume a profitability they have not yet shown. The most celebrated companies in the cycle are also the ones most capable of going to zero, and they are the ones lining up for the exits with Anthropic and OpenAI both filing for IPOs in June.
Analysis: IBM’s Black Tuesday exposes a new economic fault line
Moreover, as long as the marquee AI companies stay private, there is no genuine price discovery. You cannot short a private company. Pessimists have no way to express their view, so the only price that gets set is the one optimists are willing to pay in the next funding round.
An initial public offering fixes this – going public finally lets the skeptics in. But in today’s world of massive indexers, and with the largest IPOs getting indexed right away, it may not. When a company stays private until it is enormous and then lists, index-inclusion rules can pull tens of billions of dollars of passive, entirely price-insensitive money into the stock within weeks of the debut. None of that capital expresses any view on whether the valuation is right; it buys because the index says to buy.
The pessimists who were locked out of the private rounds arrive to find the optimists’ price already set in concrete by automatic flows, with the Canadian saver’s index fund quietly providing the exit liquidity for the insiders who got in early. This means the price you pay at the IPO is to a large extent set by the insiders selling their shares, and as such it is likely too high.
We are not suggesting avoiding AI. But we prefer the durable ends over the deflating middle: the energy, grid, specialized manufacturing and critical-mineral choke points at the bottom, and the workflow-locked, data-rich applications at the top, over the brand name model labs in between whose product is commoditizing and whose best work is regulated into a corner.
Opinion: When the financial industry fails you, AI has your back
This means the software firms that own the system of record inside mission-critical workflows – such as Constellation Software, Kinaxis or Descartes in the Canadian market; ServiceNow, Intuit or Veeva globally; or Thomson Reuters, whose Westlaw sits on proprietary case law inside every litigator’s daily routine. For firms like these, commoditizing intelligence is a cheapening input into an entrenched product – the same deflation that squeezes the middle layers widens margins at the top. Be wary of the fragile pure-plays that can fail, rather than the subsidized model efforts inside diversified giants that cannot.
This matters more to Canadian savers than you may realize. Most of us will not buy these marquee AI labs directly; we will own them through index funds, pension plans and the default options in our retirement accounts. When the largest AI names finally list and index rules pull them in, the Canada Pension Plan, the big provincial funds and millions of individual index portfolios will become buyers automatically.
The prudent response is not to flee, but to know what you own and to favour low-cost broad exposure over the fashionable concentrated bets that crowd the top of the cap-weighted index precisely when they are richly priced.
Do it right and you can own this era without being its greater fool. Alternatively, you are funding someone else’s exit.
George Athanassakos is a professor of finance and holds the Ben Graham Chair in Value Investing at the Ivey Business School. His latest book is “Value Investing: From Theory to Practice”. Christian Dippel is an economist at the Ivey Business School.