Market Outlook: 2025 Review and Beyond

The market continued its march to new highs with the S&P 500 up 3% in the fourth quarter and up 18% for the year. Incredibly, the S&P has compounded at a rate of 23% per year over the past three years, more than double the pace of its long-term average. The cost of such stellar performance is dangerously high valuations and, by extension, implied poor future performance. Depending on which valuation metric you prefer, the S&P is either at all-time highs in valuation (when looking at price / sales or price / book) or near all-time highs (when looking at price / earnings or dividend yield).   

The consequences of high valuations become evident when you graph the S&P 500 forward P/E against the subsequent 10-year annualized returns (see next chart). Current valuation levels have historically been consistent with near-zero annualized returns over the subsequent decade. The way this typically plays out is a crash followed by a recovery that nets out to a lost decade. It should be emphasized that valuations are pretty much useless as a market timing tool (it has no predictive power over the short term). However, its predictive power over a decade-long horizon should not be ignored.

While there is no telling when the current bull market ends (it must at some point), I think the ultimate trigger will occur along with a collapse in AI-related stocks. The artificial intelligence story has been the key narrative driving this bull market and will also be the key to its undoing. The debate over whether this is an AI bubble really shouldn’t be a debate at all. If we study history, we’ll find that every truly revolutionary technology, from railroads to automobiles and from radio to the Internet, has been accompanied by stock market bubbles. These aren’t coincidences. Revolutionary technologies and stock market bubbles go hand in hand. Only truly revolutionary technologies that are clearly world-changing to all observers can generate the scale of mass optimism and euphoria needed to produce stock market bubbles. Some people mistakenly believe AI can’t be a bubble because it is revolutionary. History shows that the norm would be for AI to be both revolutionary and a bubble. The question really shouldn’t be whether AI is in a bubble, but how long does this bubble last.

While the latter question is almost impossible to answer definitively, there are some observations we can make. The increasing scale of spending required to fund AI has brought about two unhealthy developments. The first is the increasing prevalence of circular funding deals. Bloomberg does a great job of highlighting these circular deals with the following illustration. And while the key AI companies all insist there is nothing improper going on, it is not a healthy sign for the sellers to be repeatedly funding the buyers. Famous short seller Jim Chanos makes a good point when he says, “Don’t you think it’s a bit odd that when the narrative is ‘demand for compute is infinite’, the sellers keep subsidizing the buyers?”

The second unhealthy development is the increasing use of debt and leverage to fund the buildout of data centers. With trillions of capital spending projected over the next decade, equity capital alone is insufficient. Companies are increasingly turning to debt to fund the buildout of data centers. Leverage is a double-edged sword that magnifies losses if future demand doesn’t fully materialize.

The problem is that we are now a full three years into the AI revolution and the return on investment remains unclear. The single most successful AI model company, OpenAI, remains deeply unprofitable. Microsoft, which owns about a third of OpenAI, reported in its Q3 2025 10-Q that losses from its share of OpenAI were $4.1 billion, up from $0.7 billion in Q3 2024. This implies that the entirety of OpenAI lost $12 billion in a single quarter. It also shows losses are getting significantly worse year over year despite growing revenues.

OpenAI has committed to spending $300 billion to buy cloud capacity from Oracle and a total of $1.4 trillion in infrastructure spending over the next decade. The problem is OpenAI doesn’t actually have the money. OpenAI is hoping to raise the capital it needs along the way. This is a bit like trying to lay tracks ahead of a moving train. It depends on both Wall Street remaining accommodating and figuring out an adequate way to monetize AI.

Although the S&P 500 reached new highs in the fourth quarter, there is trouble brewing under the surface. The following chart shows Oracle stock over the past 6 months. The stock initially popped in September following the announcement of the OpenAI deal. However, the stock has dropped 40% since then and has given up all the gains and then some.

Oracle, which already has $100 billion in debt, will likely need to borrow significantly more to fund hundreds of billions in capital spending in the coming years. So what you have is a company borrowing hundreds of billions to build data centers, a significant portion of which is for one large customer (OpenAI). OpenAI will in turn need to raise hundreds of billions of dollars to pay for its quarterly cash burn and for its massive future contract commitments to Oracle and other cloud providers. If we ever needed an example of what a house of cards looks like, I feel this is a pretty good candidate.

Oracle is far from the only example. CoreWeave is a company whose primary business is renting out GPUs (AI computing power) as a cloud service. Its largest customers include OpenAI and Microsoft. It is unprofitable and borrowing heavily to fund its capital expenditures. Its stock has fallen 60% since peaking in June. CoreWeave could very well be a zero if we get a bad bear market.

An example of what happens when funding is in question is Fermi, a pre-revenue, energy and data center development company that had its IPO in October of 2025. Its stock is down 75% in just three months. It had an impressive 46% single-day drop after an “investment-grade” tenant (rumored to be Amazon) terminated a $150 million construction funding agreement.

One area where OpenAI has been most impressive is with consumer adoption, as ChatGPT remains the leader in AI applications. However, even in this area, its shine is just a bit less bright. The search term “ChatGPT” had been growing 200% year over year in Q2 and Q3 of 2025. In Q4 of 2025, the growth decelerated to 90%. While still phenomenally strong in absolute terms, the significant deceleration in just one quarter could mark the start of slowdown in ChatGPT adoption.

Beyond ChatGPT, OpenAI’s Sora app has proven to be a flash in the pan. While initially hugely successful, interest quickly dropped off. Search interest for “Sora” has dropped 80% since the peak in early October. Mobile app downloads similarly fell off as Sora went from the #1 downloaded app in the iPhone app store to #70 recently. It turns out that watching yourself in fantastical settings is cool initially, but gets old fast. It also turns out that Sora has low social engagement as we have no desire to watch endless AI slop about other people. OpenAI’s failure to sustain engagement with Sora shows that AI does not automatically equal winning. It also suggests that the endless AI demand needed to sustain trillions of dollars of data center spending is not a sure thing.

We are now entering a dangerous part of the cycle where stocks could keep running higher as mania begets more mania and the fear of missing out creates one last speculative hurrah. However, we could also be forming a multi-year peak at any moment. Navigating the deflating of the AI bubble will be key to wealth preservation and growth over the coming decade. This is top of mind for me as I construct our portfolio to retain our equity positions while limiting exposure to the worst of the AI mania.