AI slowdown call exposes a hidden portfolio risk

The warning was aimed at artificial-intelligence laboratories. The first people to feel it were investors.

On Monday, shares tied to artificial intelligence and semiconductors fell across Asia, Europe and the United States after Dario Amodei, the chief executive of Anthropic, called for companies to slow the development of increasingly capable AI systems.

Sam Altman, the chief executive of OpenAI, quickly backed the appeal, joined by other prominent figures in the technology industry. Markets that had spent years rewarding speed suddenly had to price in the possibility that speed itself might become a liability.

The sell-off was widely read as a judgment on the immediate economics of artificial intelligence: If frontier models are developed more slowly, the argument went, companies will buy fewer chips, build fewer data centers and spend less on the infrastructure that has powered one of the market’s most spectacular rallies. Nigel Green, the chief executive of the financial advisory firm deVere Group, says that interpretation misses the more important risk. The question is not only whether AI spending slows. It is how much of the market has come to depend on one story continuing without interruption.

“Markets are treating this as straightforward bad news for anything tied to AI spending,” Mr. Green said. “It’s the wrong lens entirely.”

The immediate backdrop made the executives’ intervention harder for investors to dismiss. Days before the market reaction, a researcher at Anthropic resigned while warning that the industry was building systems whose dangers were not being adequately addressed. Jacob Coxon, who had worked in pretraining research at Anthropic and OpenAI, said in a message to colleagues that advanced AI could pose a serious risk if the industry continued without stronger safeguards. His resignation brought an internal dispute into public view at precisely the moment when financial markets were most heavily committed to the assumption that development would keep accelerating.

Amodei’s appeal was not a prediction that artificial intelligence would disappear, nor was it a declaration that companies should abandon investment in the technology. It was a warning about the pace at which the most capable systems are being built. That distinction matters for investors, because the value of many companies associated with AI has been calculated on the assumption that more powerful models will arrive quickly, regularly and at enormous scale. A change in that timetable could force markets to reconsider not just individual companies, but the earnings expectations attached to an entire category of assets.

“Amodei and Altman are debating the speed of the frontier,” Mr. Green said. “Investors should be asking a different question: how much of their expected growth was ever anything other than one theme, dressed up as diversification?”

That question reaches far beyond specialist technology funds. The rise of a small group of enormous companies has changed the character of broad stock-market indexes. When companies associated with artificial intelligence rise sharply, their growing market values give them larger weights in index funds. Investors who buy those funds do not choose each company individually, yet their portfolios can become increasingly dependent on the same handful of businesses and the same economic narrative. Research from BlackRock, for example, has contrasted the roughly 33 percent AI exposure of the S&P 500 with about 8 percent in an S&P 500 fund that excludes the largest companies.

That is the paradox of passive investing in an era dominated by a few giant technology companies. An index fund can look diversified by name while remaining concentrated by driver. It may hold dozens or hundreds of stocks, but if many of the largest positions benefit from rising demand for computing power, data centers, cloud services and advanced chips, a single shift in expectations can affect them all at once. The diversification is real at the company level. It may be less real at the level that matters during a market shock.

“Somebody who has never bought a tech stock in their life can still be sitting on a concentrated AI position through their pension,” Mr. Green said.

The concentration can be difficult for savers to see because it is distributed across accounts. A workplace retirement plan may hold a broad United States equity fund. A personal investment account may contain a global tracker. A balanced fund may own shares in technology companies directly and also hold the same companies through an index. Each holding can appear reasonable in isolation. Together, they can amount to a large, unplanned wager on the continued expansion of the AI economy.

The problem is not that the companies involved are necessarily weak, or that the underlying technology lacks commercial value. The problem is that good businesses can still be poor portfolio risks when their prices reflect unusually high expectations. The market does not need to decide that artificial intelligence is a failure for the trade to suffer. It only needs to decide that the benefits will arrive later, cost more to capture or accrue to fewer companies than investors had assumed.

Monday’s market reaction showed how quickly a debate among executives can become a repricing mechanism. The Guardian reported that AI-linked stocks fell after leaders from Anthropic, OpenAI and other technology companies supported calls for a more cautious approach to development.

The market was not responding to a change in quarterly revenue alone. It was responding to uncertainty about the assumptions beneath years of investment: how many models will be built, how quickly they will improve, how much computing they will require and how much customers will pay to use them.

There is still a strong argument that slowing the creation of new frontier models would not eliminate demand for computing. AI systems already deployed must be run, maintained and upgraded. In many cases, using a model can require substantial computing capacity even after the expensive research and training phase is complete. Businesses are also experimenting with AI in customer service, software development, search, finance, medicine and industrial operations. If those uses spread, demand for inference – the process of generating outputs from trained models – could continue even if the race to build the next model becomes more cautious.

The physical infrastructure behind that demand is already imposing its own constraints. The International Energy Agency has projected that data-center electricity demand could rise substantially as AI workloads expand, with data centers accounting for less than 2 percent of global electricity demand in 2035 in one of its scenarios. Other estimates are higher, reflecting uncertainty over how quickly companies deploy systems and how much electricity each task requires. The disagreement itself is a reminder that investors are not simply buying a known stream of cash flows. They are buying into a chain of forecasts about technology, power, construction, regulation and customer behavior.

A more measured pace could therefore produce an uneven result. It might reduce the need for some speculative capacity while leaving demand for existing systems intact. It might shift spending away from training the largest models and toward software, efficiency, specialized chips or services that help companies use what they already have. It might also expose firms whose valuations depend on an almost uninterrupted increase in model size. A slowdown would not affect every part of the AI economy in the same way, but markets often discover that distinction only after prices have moved.

The comparison with earlier market cycles is unavoidable. In previous periods, a small number of dominant growth companies helped pull entire indexes higher before a change in expectations forced investors to reassess the group. The technology bubble of the late 1990s is the familiar example, though the details are different today. The companies driving the current market are larger, more profitable and more deeply embedded in the global economy than many of the speculative businesses of that era. Yet size and profitability do not make concentration disappear. They can make it harder to recognize, because the same companies are present in retirement plans, mutual funds and benchmark indexes that investors regard as ordinary.

That is why the relevant question for an investor is not whether artificial intelligence will matter. It almost certainly will. The question is whether the investor’s portfolio contains more exposure than intended, and whether that exposure is being counted repeatedly. Someone reviewing a retirement account should examine the largest holdings in each fund, not just the fund labels. A “global” fund may be heavily weighted toward the United States. A “balanced” fund may own several versions of the same technology companies. A portfolio that appears to contain stocks, bonds and cash can still be vulnerable if its equity holdings are concentrated in one group of companies.

This does not mean that investors should react to one volatile trading session by abandoning a long-term plan. Selling after a decline can replace one risk with another, particularly when taxes, fees and timing are involved. Nor does it mean that a call for caution from technology executives should be treated as a reliable market forecast. Corporate leaders may have strategic, political or reputational reasons for how they describe the future. Their warnings deserve attention, but they are not investment instructions.

The useful response is more prosaic: find out what you own. Investors can compare the overlap among funds, measure the combined weight of the largest technology companies and consider whether their portfolio would still meet its purpose if the AI trade went through a long period of disappointment. The exercise does not require a view on whether the next generation of models will be built next year or several years from now. It requires an honest assessment of how much of the portfolio’s recent success came from the same source.

“AI isn’t going anywhere, and neither is the demand behind it,” Mr. Green said. “But a portfolio that turned into a concentrated AI bet by accident needs to be reassessed on purpose, not after the next Monday like this one.”

The argument over speed will continue inside the laboratories and boardrooms. Engineers will debate safety, executives will debate competition and governments will debate oversight. Markets will keep translating those debates into prices, often before the consequences are clear. For ordinary savers, the most urgent decision is less dramatic than predicting the future of artificial intelligence. It is deciding whether the future they are already invested in was chosen deliberately.

“Reviewing that exposure doesn’t require predicting where AI development goes from here,” Mr. Green said. “It requires an honest look at how a portfolio got built, and whether the concentration inside it was ever a deliberate decision.”

Nobody needs to predict the next twist in the debate to do that work. They just need to look.

The post AI slowdown call exposes a hidden portfolio risk appeared first on Weekend Post.