Dan Niles: Be Nimble - 30-50% AI Crash By 2027

Dan Niles, founder and portfolio manager at Niles Investment Management, anchors the current AI cycle to the public launch of ChatGPT at end of 2022, comparing it to the Netscape Navigator moment for the internet era. He places the cycle closest to 1997 and 1998 rather than 1999 or 2008, noting that in 1997 the S&P 500 fell 11 percent intra-year but finished up 31 percent, and in 1998 fell 19 percent intra-year but finished up 27 percent. He used this framework on March 31 to call a market bottom during the Iran conflict, arguing man-made crises are more resolvable than the structural ones that caused those mid-cycle drawdowns.

The formalization of agentic AI on January 30 is identified as a distinct and major new driver of the cycle. Agentic AI requires 10 to 100 times more tokens than standard chat queries, and token growth in the two months following January 30 exceeded 120 percent compared to roughly 20 percent in the prior two months. Corporations are only beginning to adopt agentic AI because it fulfills the original promise of replacing human labor with agents capable of running continuously on extended projects, a structural demand driver Niles says had no equivalent during the 1990s internet buildout.

The shift to agentic workloads changes hardware architecture requirements, moving the GPU-to-CPU ratio from roughly 8-to-1 toward 1-to-1, making CPUs more strategically valuable. Niles recommended Intel and AMD in March on this basis, arguing Intel could win large foundry customers including Nvidia, Apple, and Tesla by year-end. Intel's stock only recently returned to its year-2000 price level and he considers it undervalued relative to AMD and TSMC on an enterprise value to sales basis. Broadcom, memory, and optical components are also identified as beneficiaries of the architectural shift.

Hyperscaler capital expenditure growth estimates for 2026 rose from roughly 30 percent at the start of the year to 70 percent after the most recent big tech earnings round. Google Cloud revenue growth accelerated from 48 percent year over year in Q4 to 63 percent in the March quarter, which Niles calls the standout result among mega-cap tech. AWS grew from 24 to 28 percent year over year. Microsoft Azure moved from 38 to 39 percent, which he characterizes as the weakest of the three and evidence Microsoft is losing share due to its dependence on OpenAI rather than owning its own large language model. Niles identifies Google as the best-positioned mega-cap for the next three to five years, citing ownership of the full technology stack including Gemini, proprietary chips developed internally for over a decade, advertising cash flows, and the Pixel phone as a distribution device for the Android ecosystem, which accounts for over 75 percent of smartphones sold globally.

Niles predicts a 30 to 50 percent crash in AI-related stocks sometime in early 2026, measured from wherever prices reach before that point, anchoring the expected deceleration to roughly two years after the January 30 agentic AI formalization date. Semiconductors are currently at their most overbought levels since 1995 or 2000, both of which preceded severe corrections, and a 15 to 20 percent short-term correction in the sector would not surprise him. He warns that investors treating semiconductors as no longer cyclical are making a serious mistake, pointing to Micron posting four consecutive quarters of negative gross margins after COVID as a recent example of that error. Despite long-term valuation concerns, Niles sees short-term upside momentum in stocks, drawing a direct parallel to 1999 when he believed stocks were expensive coming out of 1998 yet they exploded higher regardless. He expects current extraordinary growth rates to remain supportive until they are anniversaried at the beginning of 2026, after which comparisons become much harder.

OpenAI and Anthropic together account for approximately half of hyperscalers' backlog, making either company's financial stress a systemic risk to AI infrastructure spending. Anthropic's annualized revenue run rate reached approximately 4.5 billion dollars by mid-2025, up from 1 billion at the start of the year, while OpenAI grew from roughly 6 billion to 24 billion annualized over the same period. OpenAI at one point projected 1.4 trillion in capital commitments over eight years against revenues running at 20 billion at end of last year, with cash burn expected to worsen before improving. Approximately 34 percent of Mag-7 profits in the most recent earnings period came from revaluation of non-listed assets such as stakes in OpenAI and Anthropic rather than from free cash flow. The anticipated IPOs of OpenAI, SpaceX, and Anthropic, likely at trillion-dollar-plus valuations each, will force fund managers to sell existing holdings to fund purchases and will enter indexes faster than normal, mechanically reducing the weightings of currently held names.

Niles identifies three simultaneously elevated asset classes as unsustainable: oil prices up roughly 60 percent on the year, the 30-year and 10-year Treasury yields hitting their highest levels of the year, and stock prices at all-time highs, stating that one of those is wrong. He notes that 10 of the last 12 recessions were preceded by a sustained spike in oil prices and says if oil remains in the mid-90s for one to two quarters a recession becomes likely by year-end. The 30-year Treasury hitting 5 percent is framed as a competing guaranteed return that could draw capital away from equities.