Ai what’s next

Here is a deep-dive list of the next ten major things this scale of AI compute (heading toward ~150 million H100-equivalents by late 2028) is positioned to find or unlock. These build directly on the chart’s existing milestones (superior weather models, Alzheimer’s insights, and the 2026 disproof of the planar unit-distance conjecture). Global AI compute has already been roughly tripling yearly and doubling every ~7 months; the projected jump multiplies training runs, inference fleets, search spaces, and agentic experimentation by another large factor. That enables deeper exploration of combinatorial, physical, and biological search spaces that were previously intractable. A cascade of additional open mathematical results and formalized proofs Beyond the unit-distance conjecture, expect systematic progress on other long-standing combinatorial, number-theoretic, and geometric problems (more Erdős-type questions, improved bounds, and counterexamples). Systems already generate candidate constructions, verify them formally, and iterate with human mathematicians. With 10×+ more effective search and verification compute, AI will routinely produce publishable advances, assist in formalizing proof sketches at research level, and push toward harder benchmarks (e.g., FrontierMath-style problems projected solvable around 2027 on current trends). This turns mathematics into a higher-throughput collaborative enterprise.47 Novel therapeutic candidates and disease mechanisms at far higher rate Building on 2025 Alzheimer’s-related gene/pathway insights, larger models plus massive molecular simulation and multi-omics search will identify causal mechanisms, repurposed drugs, and de-novo designs for complex diseases (cancer, fibrosis, neurodegeneration, aging-related pathways). End-to-end AI loops (target identification → molecule generation → property prediction → experimental prioritization) already shortened some timelines dramatically; the extra compute multiplies the number of candidates screened and the fidelity of predictions for protein–ligand and protein–protein interactions. Expect more Phase I/II candidates originating primarily from AI systems.69 New functional materials discovered via inverse design and billion-atom simulations AI can already simulate systems of billions of atoms and screen vast composition spaces for stability, magnetic, catalytic, or mechanical properties. The next wave will yield practical candidates for room-temperature (or higher-Tc) superconductors, better battery electrolytes/electrodes, carbon-negative concretes, efficient catalysts for green chemistry/fuels, and quantum materials. Autonomous labs close the loop by synthesizing and testing the top hits, turning materials discovery from years-long campaigns into months.61 Higher-fidelity whole-cell and multi-scale biological models Accurate prediction of arbitrary protein–protein interactions, full cellular digital twins, and systems-level simulations of disease states or developmental processes. Current protein-structure and interaction tools are strong but limited by data and compute; the coming capacity supports training on vastly larger simulated + experimental datasets and running longer, more detailed dynamics. This feeds directly into synthetic biology, personalized interventions, and understanding of complex traits. Practical advances in controlled fusion and plasma physics AI-optimized plasma control, magnet design, and real-time disruption prediction for magnetic-confinement devices. Digital twins of fusion systems, accelerated by orders-of-magnitude more simulation throughput, will help stabilize plasmas longer and identify better operating regimes—accelerating the path from experimental reactors toward net-energy systems. Transformative climate and Earth-system insights at regional and process scales Building on the 2024 weather-forecasting leap, next-generation models will resolve finer-scale processes (clouds, turbulence, biogeochemistry, ice-sheet dynamics) with higher accuracy and longer lead times. Expect new understanding of tipping elements, improved extreme-event attribution, and optimized interventions (e.g., for agriculture, water, or carbon management). The compute enables ensembles and hybrid physics-AI models previously too expensive. Fully agentic AI scientific researchers capable of multi-week autonomous loops Systems that generate hypotheses, design experiments or simulations, write and debug the necessary code, analyze results, critique their own work, and iterate—functioning as “AI research interns” scaling toward automatic researchers. Time-horizon benchmarks already show rapid lengthening of reliable autonomous work; the extra compute supports longer coherent reasoning, tool use, and parallel exploration. This multiplies human scientist productivity and opens problems that require exhaustive search.46 AI-designed next-generation AI hardware and software stacks Models that co-optimize chip architectures, interconnects, compilers, kernels, and even data-center layouts. Early examples already improve matrix-multiplication algorithms, recover global compute efficiency, and accelerate chip design cycles (from years toward months). With this compute base, AI will help design the successors to Blackwell/Rubin-class accelerators, photonic interconnects, and more efficient training/inference algorithms—creating a positive feedback loop. Progress on hard continuum and many-body physics problems Better approximations or numerical insights into turbulence, Navier–Stokes regularity questions, quantum many-body systems, and cosmological simulations (galaxy formation, black-hole wave behavior). AI already assists with PDE solving and symbolic discovery; massive compute enables exhaustive exploration of solution spaces, discovery of new invariants or reduced-order models, and hybrid symbolic–numeric advances. Personalized multi-omics medicine and real-time diagnostic/therapeutic systems Integrating genomics, proteomics, imaging, clinical records, and continuous monitoring at population scale to uncover individual disease trajectories, optimal interventions, and early predictive signatures far beyond current Alzheimer’s-style findings. The inference capacity supports always-on models that personalize treatment in real time, while training compute unlocks foundation models trained on vastly larger multimodal biomedical corpora. Key caveats These outcomes are not automatic. They require continued algorithmic progress, high-quality data (or high-fidelity simulation data), energy infrastructure, experimental validation loops (robotic labs, clinical trials), and human scientific judgment. Bottlenecks in power delivery, data quality, or regulation could slow deployment even if the raw chips arrive. Nonetheless, the historical pattern—each major compute increase unlocking qualitatively new scientific results—strongly suggests the list above is the direction of travel for 2026–2028 and immediately beyond. The chart’s trajectory is the enabling condition for turning today’s promising prototypes into routine, high-impact discoveries.