Artificial Intelligence & The Market: What's the Situation Today

Daniel Lee
Jul 1
8 min read

Artificial intelligence (AI) is now the single largest driver of equity returns.

A small group of AI-related companies has accounted for the substantial majority of market gains, which means client portfolios — whether or not AI was ever a deliberate allocation — are now heavily exposed to sentiment around this one theme.

The case for and against AI is usually framed as a debate to be won: either the technology is real, or it is a bubble. That framing is wrong. Both sides of the ledger are true at the same time, and the tension between them is the most important thing for an investor to understand right now.

The exhibit below sets the two forces side by side: genuine, exploding demand on the left, and the structural problems of who is paying for it on the right. The sections that follow elaborate on each, reconcile them, and set out the implications for how we position.

I’ll explain these developments in 5 parts

the growth & upside
the downside and the risks
the reconciliation of the upside and downside
what this means to the market
what this means to us as investors

The Growth: How Much AI Is Being Used

The strongest argument the optimists have is that AI usage is not hype in share prices — it is real-world activity growing at an extraordinary rate.

The clearest measure is the “token”, the small unit of text (roughly three-quarters of a word) that a model processes on every request, which has seen tremendous growth over the past few years,

Google reported processing 9.7 trillion tokens per month two years ago, about 480 trillion a year ago, and over 3.2 quadrillion per month as of May 2026 — a roughly seven-fold jump in a single year, on top of a fifty-fold jump the year before. (A quadrillion is a thousand trillion.)
OpenAI, the maker of ChatGPT, said its developer system was handling more than 15 billion tokens every minute by March 2026, up from 6 billion per minute roughly six months earlier.
Usage is global and broadening: China's daily token consumption reportedly rose from around 100 billion at the start of 2024 to roughly 180 trillion per day by early 2026, driven heavily by AI-generated video.

This usage is converting into revenue at a pace with no precedent in enterprise software, as seen from the revenue of OpenAI and Anthropic, which has also supported the demand for an enormous capital commitment of hyperscaler infrastructure spending.

That spending, and the revenue growth it is chasing, is the substance of the bull case.

The downside: Who Pays & Whether it Pays Off

Each of the three downside legs is worse on inspection than the headline suggests.

On provider economics: The cash burn is structural, not a rounding error. While the revenue appears spectacular based on historical standards, the bottom lines of such providers show a totally different story.

OpenAI’s internal projections point to losses of roughly $14 billion in 2026 alone, with profitability not expected until around 2030, and analysts estimate it may need well over $200 billion in additional funding to get there.

Anthropic is the apparent counterexample, projecting its first profitable quarter in Q2 2026 — but the shape of that profit matters: roughly $559 million of operating income on $10.9 billion of revenue is about a 5% operating margin on a ~40% gross margin.

That is real, but it is nothing like the 70 to 80% software economics that would justify the valuations being paid, which is precisely why the company targets a 77% gross margin only by 2028.

Even the “profitable” provider is profitable in a way that does not yet look durable.

On the financing: The equity invested in the AI labs does not come close to covering the cost of the build-out, so the gap is being filled by debt. The strain is no longer theoretical; it is visible in prices.

Oracle, the load-bearing builder for OpenAI’s data-centre programme, entered the build-out with roughly $19.8 billion of cash against a $124 billion debt load, is raising $45 to 50 billion in 2026 through debt and equity to fund capital spending that runs toward $95 billion in fiscal 2027, and the market responded by sending its five-year credit default swaps to record highs and knocking the stock down nearly 9% on the spending news.

Further down the credit ladder it is starker: xAI raised bonds and loans whose fixed-rate portion carries a 12.5% coupon, and when a major bank tried to syndicate roughly $38 billion of data-centre debt it met weaker demand as it sold the loan down - a sign the financing is testing the market’s appetite, not sailing through it.

This debt has also migrated into supposedly safe places: through asset-backed structures, the chain runs from the AI lab paying the cloud provider, to the provider servicing the bonds, to those interest payments landing in ordinary bond funds and retirement accounts whose holders never knowingly made an AI bet.

On demand: The most uncomfortable finding is that the buyers largely cannot yet show a return.

An MIT study found that 95% of enterprise generative-AI deployments produced no measurable profit-and-loss impact, and the result is corroborated from independent angles: S&P Global found 42% of companies abandoned most of their AI projects in 2025 (more than double the prior year), IBM put initiatives delivering expected ROI at just 25%, and Morgan Stanley found only 21% of S&P 500 companies could cite any measurable AI benefit at all.

Statistic	Figure	Source
Organizations using AI in at least one function	88%	McKinsey, 2025
Organizations capturing significant value (high performers)	~6%	McKinsey, 2025
Generative-AI deployments with no measurable P&L impact	95%	MIT Project NANDA, 2025
Enterprise AI projects that fail to deliver value	80%+	RAND, 2024
Infrastructure & operations AI use cases that fully meet ROI	28%	Gartner, 2025
Organisations scaling an agentic AI system	23%	McKinsey, 2025
Developers using or planning to use AI coding tools	84%	Stack Overflow, 2025

This is now showing up as hard cost decisions rather than survey sentiment: reporting in May 2026 cited Microsoft cancelling most of its licences for one AI coding tool over cost, while Uber said it burned through its entire 2026 AI budget in four months as adoption of the same class of tool jumped from 32% to 84% of its engineers.

The through-line across all three legs is that the stress has stopped being a forecast and it is appearing in credit spreads, syndication difficulty, cancelled contracts, and abandoned pilots, in real time.

Reconciling The Upside & Downside

The two upside and downside factors do not actually contradict, and that is the key insight.

Token growth measures activity, not profit. A quadrillion tokens flowing through a system tells you usage is exploding; it tells you nothing about whether the companies paying for those tokens are earning a return, or whether the companies selling them are making money - and on both counts, they largely are not yet.

The right mental model is a single financed loop: enterprises buy AI; providers book the revenue but burn cash; providers commit hundreds of billions to compute; cloud builders borrow to construct data centres; that debt is serviced by provider lease payments; which depend on provider revenue; which depends on enterprises seeing enough value to keep spending.

Every link holds only as long as the next one does, and the weakest link — the 95%-no-return finding — sits at the very start of the chain. If enterprises conclude the spend is not paying off, demand softens and the entire financed structure feels it.

Two qualifiers keep this honest, because the bear case can be overstated as easily as the bull.

First, the “zero return” finding is real but contested in interpretation. It may partly reflect organisations measuring the wrong things too early, much as email and the internet took years to show up in profits. The technology being early is not the same as the technology being worthless.

Second, most of the debt does not sit on the AI labs’ balance sheets but on the hyperscalers, whose existing, highly profitable businesses can fund this in a way the speculative start-ups of the dot-com era never could. So “it is all debt-funded vapour” is wrong.

The accurate, narrower version is this: a meaningful and growing slice is debt-funded, the marginal borrower is being charged a rising risk premium, and that risk has quietly seeped into instruments most people regard as safe.

The danger is not that AI is fake; it is that today’s price has already paid for a level of monetisation that the evidence says has not yet arrived.

What This Means For The Market

The market’s price level is the part most exposed to this gap.

US index concentration is at a half-century extreme with the ten largest companies now making up more than 41% of total index value and valuation gauges such as the Shiller P/E sit above 40, comparable to the dot-com peak, which means a great deal of the bull-case future is already in the price.

Tellingly, the debt market is already discriminating where the equity market has not: lenders are charging a measurable credit-spread penalty to companies classified as AI “adopters” versus “enablers” - in effect pricing the risk of spending without proof of return.

Credit usually identifies trouble before equity does. When equity catches up to what credit is signalling, the result is a fast repricing of the most AI-exposed names, which is precisely what occurred in the final week of June, when an “AI bubble” scare triggered a sharp global sell-off in technology and semiconductor shares.

The honest framing is that current prices are supported by genuine, exploding demand and by sentiment that has run ahead of demonstrated returns, and the two are very hard to separate from the outside.

Sentiment is doing much of the load bearing, and sentiment is what cracks first.

What This Means To Us As Investors

The analysis above resolves into a small number of practical implications for how portfolios are positioned and what we'll have to watch moving forward:

Diversified investors are now structurally long this theme - by default, not by choice. Because a handful of AI names dominate the index, a plain “buy the market” position now carries far more concentrated AI exposure than it did three years ago. That has driven strong returns and concentrated the risk. The most useful first step is not to call the top, but to understand how much AI exposure sits inside a portfolio’s “diversified” index holdings and decide whether that is the intended bet.
The downside is no longer priced as a base case, which is what makes it dangerous. The market has largely accepted the bull narrative. If the weak link gives, be it in the form of enterprises pulling back for lack of return or a high-profile financed builder stumbling, it would strike a market that has stopped hedging for it, tending to produce sharp moves rather than gentle ones.
Token growth and share prices measure two different things; conflating them is the trap. Exponential usage growth is real and is the genuine bull case. Whether that usage turns into profit, both for the providers selling it and the enterprises buying it, is the open question, and it is the question today’s price has answered optimistically.
“Safe” allocations may carry AI risk. Because data-centre debt has been packaged into bond funds and similar vehicles, exposure to the AI build-out is not confined to technology equities. Reviewing what sits inside fixed-income holdings is part of understanding total exposure.
The right question is whether you are being paid to own it. The investor’s task here is not to bet on whether AI is transformative - it probably is, eventually - but to ask whether the expected return adequately compensates for owning it at today’s valuation, with today’s concentration, funded by today’s debt.