Prominent fund manager warns: The AI party could end in a 50% crash

The "speed bump" that could stall the AI rally

Dan Niles, founder of Niles Investment Management and one of Wall Street's more prominent technology investors, is now warning of what he describes as an inevitable "speed bump" in the ongoing AI trade. According to Seeking Alpha, Niles believes the market's euphoric view of artificial intelligence is no longer grounded in economic reality.

The core of his argument centers on a fundamental shift in how companies use AI: from what he calls "token maximization" to "token minimization."

"The higher the AI bubble drives stock prices, the more painful the descent on the other side will be for investors." — Dan Niles

From growth at any cost to cost control

In the early phase of the AI wave, companies spent vast sums to maximize their use of advanced language models. Niles points out, among other things, that Uber reportedly burned through its entire AI budget within four months – a pace that is clearly unsustainable over time.

Now the trend has reversed. Businesses are increasingly routing their AI requests to cheaper open-source models, which according to Niles can cost as little as one-eighth the price of the most expensive alternatives. Microsoft, for example, is said to have asked its engineers to limit their use of Anthropic's Claude due to unmanageable token bills.

This transition could show up in companies' financial results as early as the September quarter, and Niles fears that downward revisions to earnings guidance could act as a catalyst for a price correction.

Semiconductors most overbought since the dot-com bubble

Niles directs particular attention toward the semiconductor sector. The semiconductor index has doubled, and Niles believes the sector is now more overbought than at any point since the 1995–2000 period – just before the dot-com collapse. He urges investors to become more selective and to stop treating the entire sector as one homogeneous group.

Agentic AI: greater demands, but differently distributed

One factor Niles highlights as particularly noteworthy is the rise of "agentic AI" – systems that perform autonomous, multi-step tasks without human intervention. These systems require far more computing power than traditional chat-based AI.

Paradoxically, Niles believes this could shift demand within data centers. The ratio of processors (CPU) to graphics processors (GPU) could move from the current approximate 8:1 toward 1:1 – potentially benefiting traditional processor suppliers such as Intel over the GPU-dominant players.

History's echo: the dot-com warning

Niles is not alone in drawing parallels to the dot-com bubble, but he is among the most specific in his forecast. He estimates that a "painful reset" could occur in early 2027, with potential price declines of 30–50 percent for the most AI-exposed stocks.

It is worth noting that such predictions are inherently highly uncertain. Niles has historically been a contrarian voice, and the timing of any potential corrections is notoriously difficult to predict. Investors should treat such scenarios as one perspective among many, not as a definitive answer.

For Norwegian investors, AI exposure is particularly relevant through the Government Pension Fund Global (GPFG), which holds significant stakes in companies such as Nvidia, Microsoft, and Alphabet – all heavily exposed to the AI theme.