AI Stocks Hit by Selloff as Compute and Evaluation Draw Attention

Highlights

• Nasdaq 100 falls 6.4% in a single day as AI stock selloff deepens
• Moonshot AI reportedly in talks for Nvidia GPU compute in Europe
• Benchmark of 102 papers tests LLMs' review-writing ability

Market: AI stocks hit by one of the year's sharpest selloffs

Tech stocks fell sharply on the first trading day after the U.S. midterm elections. The Nasdaq 100 dropped 6.4% in a single session, and the Financial Times reported that $3.2 trillion in market value was wiped out on Wall Street that day, with AI-related stocks leading the selloff. (Source: FT Technology)

CNBC's Jim Cramer said after the close that the selloff in AI stocks may not be over, but he also argued it was different from the dot-com bust. He added that AI stocks are becoming increasingly hard to value, citing uncertainty around data centers as one reason. (Source: CNBC Technology)

Taken together, the two points show that the core of the market disagreement is the basis for pricing AI assets: there is currently no aligned reference point between the certainty of compute spending and the pace at which revenue materializes. That judgment could be overturned or reinforced by capex guidance and order data in the coming earnings season.

Compute: Chinese model company seeks European GPU resources

CNBC, citing four people familiar with the matter, reported that Chinese AI startup Moonshot AI held talks this year with several emerging cloud providers operating data centers in Europe about renting Nvidia GPU compute. (Source: kaopunews)

The report remains at the "talks" stage, with no disclosed amount, scale or whether an agreement was reached. But it points to a persistent structural problem: for leading model companies, where and how they secure compute is becoming as important a variable as model capability.

Research: LLMs can find literature but struggle to judge it

A new study on arXiv used a published review of "Spanish plume" weather patterns and its 102 cited papers as a benchmark, evaluating six AI tools across four stages: finding articles, interpreting them, synthesizing them and proposing future research directions. (Source: arXiv cs.AI)

The results showed that in the initial search, five tools together found 57 articles, with the best single tool finding 42. After access to institutional library permissions was added, Codex's discoveries rose from 42 to 82 papers, and Claude Code's from 15 to 79. Given all 102 PDFs, three tools correctly identified all 18 articles that substantively discussed the topic, but misclassified 40% to 79% of the 84 papers that mentioned it only in passing as substantive discussions. Synthesis reports blind-reviewed by two domain experts were generally accurate but offered limited insight, and contained unsupported claims, citation errors and exaggerated consensus. (Source: arXiv cs.AI)

The numbers show that LLMs are approaching usable performance at the "finding" stage of literature work. The bottleneck is judging how relevant a paper is and enforcing evidentiary constraints during synthesis.

Engineering: Agentic AI's resource needs are hard to predict

Another arXiv paper formalizes resource orchestration for Agentic AI systems as an "uncertainty-aware" problem: these LLM-driven multi-step pipelines cannot know total compute, bandwidth and end-to-end latency in advance when scheduling. Demand is generated at runtime by intermediate observations and stochastic sampling, producing heavy-tailed, multi-modal distributions. Existing schedulers require a known full pipeline graph, while edge resource-management methods rely on parameterized or stationary demand models. Neither fits. (Source: arXiv cs.AI)

This points to the same thing as the evaluation above: as LLMs move from single-turn Q&A to multi-step tasks, reliability and cost predictability become preconditions for deployment, not side issues.