AI Giants Shift Billions to World Models as Large Language Model Progress Stalls in 2025

Major AI firms are increasingly investing in world models as progress in large language models (LLMs) shows signs of slowing in late 2025[2][3]. Venture capital and direct funding rounds have shifted focus, with billions now flowing to startups and initiatives developing multimodal, agentic, and simulation-based AI systems, marking a critical transition in the field.

Why World Models Are the New Frontier

World models are AI systems designed to emulate and predict environments, enabling agents to plan, reason, and interact more flexibly than traditional LLMs. While LLMs—like OpenAI’s GPT series—excel at generating text, their ability to understand complex, dynamic, real-world scenarios is limited. World models combine multiple modalities (text, vision, action) and can simulate entire environments, making them essential for robotics, advanced automation, and scientific discovery[3].

The slowing pace of LLM breakthroughs in 2025 stems from:
– Diminishing returns in scaling model size for improvements in text-based reasoning and generation.
– Increasing costs and energy requirements for training ever-larger models.
– Saturation in commercial applications, as most major use cases for LLMs (chatbots, summarization, code generation) have matured[2].

Funding Shifts: Billions for Multimodal and Agentic AI

The investment landscape reflects this transition:
– OpenAI is raising up to $40 billion in a round led by SoftBank, with much of the funding earmarked for next-generation AI infrastructure and world modeling capabilities[1].
– The Allen Institute for AI (Ai2) received over $150 million from Nvidia and NSF for the “Open Multimodal AI Infrastructure to Accelerate Science,” a five-year initiative focused on open, multimodal world models tailored for scientific research[1].
– Meta and other tech giants are backing startups like Databricks and Skild AI, aiming to build flexible models that understand and manipulate data, images, and actions across cloud and robotics platforms[1][3].

Projections suggest global spending on agentic AI—systems capable of autonomous decision-making and planning—could reach $155 billion by 2030, with world models at the core of this growth[3].

The Technical and Strategic Appeal of World Models

World models are attractive for several reasons:
– Multimodality: They process and integrate text, images, audio, and sensor data, moving beyond the “language-only” bottleneck[1][3].
– Simulation and Reasoning: World models simulate environments, predict consequences, and enable agents to plan steps toward complex goals—a leap from reactive language generation.
– Generalization: These models can learn from fewer examples and adapt to new settings, critical for scientific research, robotics, and real-world automation.

Leading AI labs now prioritize building open, reproducible world models, collaborating with universities and national research foundations. Nvidia’s investment in hardware and infrastructure (e.g., Blackwell Ultra architecture for Ai2) highlights the computational intensity and ambition of these projects[1].

Implications for Industry and Research

This pivot has broad effects:
– Robotics: World models enable robots to understand their environments, predict obstacles, and make strategic decisions, accelerating automation in manufacturing, logistics, and healthcare.
– Drug Discovery and Science: Multimodal models trained on scientific literature and experiment data can simulate molecular interactions, optimize experiments, and generate novel hypotheses at scale.
– Cloud and Enterprise AI: Companies like Databricks integrate world models to help organizations analyze complex data types, automate workflows, and develop intelligent agents for business optimization[1][3].

Startups with a focus on vertical AI—tailored to specific industries—are also benefiting, as seen with EliseAI’s $250M round for healthcare agentic models[1].

Challenges and the Road Ahead

Despite the excitement, building effective world models is technically challenging:
– Data requirements: Training models to understand and simulate real-world complexity needs vast, diverse datasets—often requiring new forms of data collection and annotation.
– Computational costs: World models demand advanced hardware and efficient training algorithms, driving up infrastructure investments.
– Evaluation and safety: Ensuring reliability, interpretability, and safe deployment in dynamic environments remains an open research question.

Nonetheless, the consensus among major investors and research labs is clear: the future of AI lies in multimodal, agentic, and world modeling systems, not just ever-larger language models[2][3]. As LLM progress plateaus, the race to build AI that can reason, plan, and act in the real world is defining the next era of innovation.

In summary: 2025 marks a pivotal year as big AI firms redirect funding from LLMs to world models, betting that these complex, multimodal systems will unlock new capabilities—and value—across industry and science[1][2][3]. For startups, researchers, and enterprises, the message is clear: mastering world modeling is now the key to staying at the forefront of artificial intelligence.

Original source: Ars Technica – Big AI firms pump money into world models as LLM advances slow