Big Tech Selloff: Dan Niles on What Happens Next

Dan Niles sees the AI trade as increasingly vulnerable, though he does not regard the semiconductor index’s June 22 peak as proof that the market has reached a lasting top. AI investment is now about 3.5% of GDP, he said, more than the combined investment in highways, electrification and the internet. That scale raises the risk of eventual overinvestment and a major bust, but he likens current sell-offs more to the 1997–98 internet corrections. Six months earlier, he had forecast a 30%–50% AI sell-off by early 2027 without predicting when it would begin. He advises investors to stay nimble and judge each company by its data and stock reaction.

Financing and growth trends are adding pressure. Several neocloud deals, including those involving Firmus and Nscale, were delayed; CoreWeave debt traded at 15%, and credit default swaps rose for companies including Meta and NVIDIA. Niles also sees capital being spread across competing AI investments, with Oracle, OpenAI and Anthropic seeking money from the same pool. He cited a slowdown in Anthropic’s sequential revenue growth, from 145% in Q2 to 45% in Q3. Ramp Index spending on OpenAI and Anthropic rose about ninefold from the start of the year through early August, then flattened, even as underlying token volumes rose about 80% after early August. Competition from open-weight models and Google’s lower-cost frontier models is another pressure.

Niles expects Anthropic’s planned IPO to proceed, but possibly at a lower valuation and for less money than investors anticipate. A withdrawal, he said, would be “horrific” for the AI trade and would make him reconsider whether the market peaked in June. He views Anthropic as better positioned than OpenAI because of its enterprise focus; OpenAI faces competition from Anthropic in business and Google among consumers. Niles said about 5% of OpenAI’s customers pay for its service, while Anthropic has grown from smaller than OpenAI at the start of the year to larger and is profitable on an adjusted operating-profit basis. He sees enterprise adoption as early, moving from model training to chat and then agents, which can require 10 to 100 times as many tokens as a simple query.

Microsoft is positioned to sell businesses AI tools within a secure environment, Niles said. It has 450 million Microsoft 365 licenses, but fewer than 10% of users have Copilot licenses; he expects businesses to pay for Copilot and Microsoft to route tasks among models suited to their cost and capability needs. AI could also reduce staffing needs: Niles said he would not hire a research analyst this year because AI could handle work he would otherwise assign. But he worries that data-center capacity may not arrive as planned. He suspects Microsoft’s roughly $5 billion capex shortfall versus expectations could reflect a neocloud partner, possibly CoreWeave, failing to deliver. He cited local resistance to development in Texas and Virginia, and calls neocloud companies “toast,” while warning their stocks could still rally 40% or 50%.

Google has both exposure to AI’s disruption of search and a path to reach consumers through phones. Search and YouTube revenue each missed expectations, despite query growth reaching a five-year high; Google shares fell 12% in one day and were down 31% from their recent peak. Niles sees room to monetize AI Overviews by adding ads, and notes Android runs on 75% of the world’s phones. He thinks Google could have a strong year next year but does not know whether the stock has bottomed. Meta’s entertainment-focused business is less exposed to search disintermediation, and AI could improve efficiency or support a consumer agent, but Meta faces negative free cash flow and substantial funding needs. Niles also expects pressure on PC sales from higher semiconductor prices: Windows OEM revenue fell about 7% in the June quarter and 22% in September, with Microsoft guiding to a 35% decline in December.

Niles expects both Nvidia and ASIC makers to benefit, since GPUs are versatile while ASICs are cheaper and task-specific, but warns that even strong semiconductor stocks can suffer sharp drawdowns. Intel’s earnings miss was, in his view, the worst he could recall in 30 years covering the stock; server strength did not offset consumer weakness. He sees Intel’s foundry business as strategically important because 90% of advanced chips come from Asia, and expects possible agreements with major customers next year, though Intel could remain under pressure as AMD gains share. Apple, by contrast, generates substantial cash: about $30 billion in free cash flow in the June quarter and more than $130 billion in the prior 12 months. Niles thinks December-quarter expectations may fall as memory costs weigh on phones, but sees next year as potentially strong if a foldable iPhone and better AI features spur upgrades.

Niles says a future sell-off could be a temporary correction or a lasting peak; he cannot know in advance which. Cisco’s history illustrates both possibilities: its 1997 and 1998 declines were followed by gains, while its decline beginning in 2000 exceeded 80%. He warns against assuming every dip is a buying opportunity and argues that cash flow, funding capacity and business model matter more as capital tightens. He also sees risks from rising oil prices, which preceded 10 of the last 12 recessions, and expects the Fed to raise rates in December and early the following year.