Why You Should Wait Out AI’s Super-Spending False Start

Janusz Marecki argues that large language models may be approaching fundamental limits, citing a data ceiling, diminishing returns from more compute, and ongoing hallucination and probabilistic error issues. The piece is a cautionary insider view on AI hype versus reality rather than a company-specific development, so near-term market impact is likely limited. The main takeaway is a more skeptical read-through on the pace and durability of LLM scaling.

Analysis

The key market implication is not that AI stops growing, but that the marginal dollar of investment likely shifts from frontier model training toward infrastructure, distribution, and workflow software that can monetize existing models. If scaling laws are flattening, the winners are those that turn inference into recurring enterprise spend: cloud compute, data plumbing, orchestration, and application-layer vendors with sticky seats and usage-based pricing. That is a more selective trade than the prior “own anything AI” regime, and it should compress the premium on pure model narratives while supporting picks-and-shovels beneficiaries.

A second-order effect is competitive pressure inside the AI ecosystem: if frontier progress slows, open-source and smaller specialized models become relatively more attractive because performance gaps matter less than cost, latency, and control. That creates a headwind for the companies forcing heavy capex into ever-larger training runs, but a tailwind for firms selling inference optimization, model routing, and on-prem deployment. It also raises the bar for monetization timelines: the market may begin to demand evidence of revenue per token and gross margin durability rather than just usage growth.

The main risk to this thesis is timing. A near-term breakthrough in multimodal reasoning, agentic workflows, or a new data acquisition method could re-ignite the capex cycle and re-rate the leaders within months, while regulatory or copyright friction could slow data access and reinforce the scarcity argument over years. The contrarian read is that this may not be a “death of AI” moment but a normalization moment: the speculative overbuild in model-capex names may be overstated, yet the broader enterprise software adoption curve could still be early and underowned.

Key Decisions for Investors

• Rotate out of high-multiple frontier-model narratives and into AI infrastructure beneficiaries over the next 1-3 months: prefer large cloud and networking names with visible inference demand over pure-play model developers; use any post-earnings rallies in the latter to trim exposure.
• Long a basket of AI application software with proven seat expansion and short a basket of capital-intensive model builders / inference-heavy private-market proxies where accessible; target a 6-12 month horizon as margin skepticism rises.
• Pair trade: long hyperscaler capex beneficiaries with strong pricing power and short semiconductor names most levered to training-cycle exuberance, but only on strength after the next AI conference cycle; stop if capex guidance re-accelerates for two consecutive quarters.
• Buy 3-6 month downside protection on the most crowded AI hardware winners if implied volatility is subdued; the catalyst is not immediate earnings collapse but multiple compression as investors differentiate between hype and monetizable usage.
• If open-source model adoption accelerates, increase exposure to on-prem security, data governance, and model-orchestration vendors; this is the cleaner second-order winner if enterprises prioritize cost and control over frontier performance.