Morgan Stanley Predicts April-June AI Capability Surge as Industry Splits on LLM Limits

Morgan Stanley has issued a client advisory predicting a significant acceleration in artificial intelligence capabilities within the next three months, warning that markets remain unprepared for what the bank characterizes as a "non-linear increase" in large language model performance expected between April and June.

The forecast arrives amid deepening disagreement within the AI research community over whether continued scaling of language models can achieve human-level intelligence. The bank's projection stands in tension with skepticism from prominent researchers who argue that current architectures face fundamental limitations in reasoning and physical world understanding.

"The market is not prepared for the non-linear increase in LLM capabilities, which, in our view, will become evident in April-June," Morgan Stanley told clients following its recent technology, media, and telecommunications conference, where multiple AI executives discussed upcoming model improvements.

OpenAI CEO Sam Altman signaled in February that capability jumps would arrive faster than previously anticipated. "The world is not prepared," Altman said. "We are going to have extremely capable models soon. It's going to be a faster takeoff than I originally thought."

The bank estimates nearly $3 trillion in global spending on AI-related infrastructure through 2028, projecting $2.9 trillion in data center construction costs alone to address compute demand that "vastly exceeds supply." OpenAI's recently released GPT-5.4 model reportedly achieved an 83 percent score on the GDPVal benchmark, designed to evaluate AI performance on economically valuable tasks, according to Fortune.

(Morgan Stanley's March advisory reflects growing Wall Street attention to AI capability timelines, with investment banks increasingly positioning artificial intelligence breakthroughs as material market events requiring client preparation and portfolio adjustment.)

Yet the optimism around language model scaling faces pointed criticism from researchers who pioneered the current AI wave. Critics argue that linguistic training alone cannot replicate human cognition, which they contend is fundamentally grounded in physical world interaction rather than text prediction. This technical disagreement carries strategic implications for capital allocation across competing AI architectures.

The capability debate has also intensified concerns about AI-generated content quality. NewsGuard launched an AI content farm detection tool in collaboration with Pangram Labs, designed to identify when news sites host significant portions of content created by large language models. "There's just going to be so much spam and bots and slop online that it's going to be pretty unusable without technology to help you wade through the slop," said a Pangram executive, reflecting industry anxiety that rapid capability gains may flood information ecosystems with low-quality synthetic content before quality controls mature.

Meanwhile, enterprise AI adoption continues through workforce development initiatives. LTIMindtree announced a partnership with IIT Kharagpur on March 16 to design training programs focused on AI skills, reflecting corporate efforts to build internal capabilities as the technology evolves. Separately, HAIL AI unveiled a multi-system architecture combining three coordinated AI systems for public websites, attempting to reduce hallucination risks through orchestrated outputs rather than relying on single language models.

The divergence between Wall Street's near-term capability forecasts and researcher skepticism about long-term architectural limits highlights uncertainty over whether current AI approaches face imminent breakthroughs or fundamental ceilings. Morgan Stanley's April-June timeline will test whether language model scaling continues delivering measurable capability gains or whether alternative architectures gain traction as limitations become apparent.