AI Came for Math. Now It’s Chasing the Rest of Science

A study by Neil Thompson and colleagues found that leading AI models’ success rate on selected theoretical problems across six scientific fields rose from about 2% in mid-2024 to about 35% in mid-2026. Researchers say the progress could reshape theory-heavy research, while emphasizing that experts remain necessary to interpret and verify solutions; OpenAI has withdrawn some solutions after publication.

Analysis

The investable signal is an expansion in potential AI workloads, not yet a demonstrated acceleration in commercial scientific output. If researchers iterate more on theory, demand could shift toward sustained inference and expert verification—not just training—supporting AI compute and cloud providers over a 6–18 month horizon. But model efficiency, price competition, and the fact that this benchmark excludes experimental work could limit incremental spending. The nearer-term bottleneck is trust: withdrawn or incorrect solutions increase demand for validation and human review, slowing any near-term substitution of research labor.

The contrarian read is that task-solving scores may overstate economic value. Solving a well-specified problem is not the same as producing a reproducible result, connecting it to useful experiments, or generating a product. The study is a capability indicator, not evidence of near-term revenue, productivity gains, or reduced scientific headcount. Avoid treating it as a direct biotech catalyst.

Near term, expect limited fundamental repricing absent disclosed customer adoption or spending. Over 1–3 months, watch AI-lab partnerships, cloud workload disclosures, and evidence that scientific users are paying for inference and verification. Over 6–18 months, the upside case is greater compute utilization and faster R&D cycles; the downside is falling inference prices or weak translation from theory to experimental outcomes. Falsify the workload thesis if cloud/AI providers report slowing usage or capex, or if independent validation continues to show material error rates that prevent deployment.

Key Decisions for Investors

• No immediate directional trade on the study alone: its benchmark does not establish commercial demand, and the likely public-equity beneficiaries are indirect.
• Watch AI infrastructure and cloud exposures for entry on weakness only if upcoming disclosures show rising scientific or other sustained inference workloads; do not infer incremental revenue from model capability announcements alone.
• Treat scientific software, data, and verification tools as a watchlist theme rather than a current long: seek evidence of paid deployments, repeat usage, and validated results before underwriting revenue impact.
• Reassess in 1–3 months against cloud usage/capex commentary and independently verified model accuracy; reduce the AI-workload thesis if adoption remains experimental or error rates impede use in research workflows.