After the research in Silicon Valley, Goldman Sachs summarizes: Agents have entered the execution era, AI competition has shifted to workflows, and the rise of world models.

AI is entering a new phase, transitioning from "being able to answer" to "being able to execute." According to the Tracking Trade platform, Goldman Sachs' latest report shows that AI commercialization is shifting from "subscription-based access" to charging based on consumption, transaction volume, and outcomes. Meanwhile, Agents are evolving from being auxiliary tools to executing workflows, and industry value is transitioning from the models themselves to proprietary data, business context, and domain expertise. This indicates that competition in the AI industry is shifting from "whose model is stronger" to "who can truly master workflows." While model capabilities remain important, the ability to integrate into enterprise production environments, understand business contexts, and consistently complete tasks will become the more critical competitive barrier. This judgment comes from Goldman Sachs' recent field research on the AI industry chain in Silicon Valley. From August 18 to 19, Goldman Sachs visited AI startups, leading venture capital firms, and researchers from Stanford University, the University of California, Berkeley, and the University of San Francisco for the third consecutive year. Goldman Sachs believes that as Agents accelerate their implementation, the value distribution among frontier models, open-source models, world models, enterprise software, and proprietary data will all change. Agent Implementation: What Enterprises Truly Lack Is Not Capability, but "Control" If in the past AI was about "helping people complete tasks," Agents are now attempting to solve "completing tasks by themselves." However, during large-scale deployment in enterprises, the greatest obstacle may no longer be model capability, but rather how responsibility is defined and whether the entire execution process can be controlled. The report cites Stanford researchers pointing out that most enterprises are still in manual supervision mode. Particularly in fields such as law, risk control, insurance, and auditing, should a model make a mistake, determining who bears responsibility, how to trace the process, and whether it can be corrected in time may be as important as the model's capabilities. Thus, the workflows that are most likely to achieve automation first typically possess three characteristics: clear decision boundaries, verifiable outcomes, and the ability to roll back errors. Invoice processing is a typical case. AI is responsible for extracting fields and performing checks, with low-confidence cases handed over to human reviewers, then completed through a reversible ERP process. This also means that information service providers with trusted content, verified domain models, and mature regulatory relationships are more likely to enter enterprise production environments ahead of others. Model Competition: Division of Labor between Frontier Models and Open-Source Models Around the debate of "open-source versus closed-source," Goldman Sachs' investigation indicates that it is not a binary choice, but rather that different models may correspond to different levels of workflows. Proponents of frontier models believe that enterprise benchmarks often underestimate model capabilities. In real production environments, business losses due to decreased model accuracy may far exceed the savings in inference costs. Therefore, although several AI-native companies claim to employ a multi-model strategy, they still heavily rely on frontier models in core production environments. On the other hand, another viewpoint suggests that the vast majority of enterprise workflows do not require frontier-level intelligence. As the performance of open-source models continues to improve, clients are increasingly willing to exchange limited performance losses for lower inference costs. A venture capital firm estimates that in the next 12 to 18 months, about 90% of inference tokens will flow to open-source models. This indicates that the future AI model market may establish a clearer division of labor: frontier models will be responsible for complex tasks that require high value and high reliability, while open-source models will handle a larger scale of standardized tasks and most of the token consumption. World Models: AI Computing Power May See a Second Growth Curve Over the past 18 months, researchers have increasingly shifted their focus from LLMs to "world models." Unlike LLMs that primarily rely on internet data for training, world models need to understand environments, causal relationships, physical laws, and dynamic interactions in the real world. Their data comes more from physical systems, specific industries, and real operational scenarios. This implies that the importance of proprietary data may further increase. Goldman Sachs believes that the problem space corresponding to fields such as physics, industry, science, and companies like Siasun Robot & Automation is far greater than that of pure text generation, and these workflows often require a higher computational power investment. As AI further penetrates the physical world from the digital world, the demand for computing power for model training, simulation, and inference may witness a new growth curve. Goldman Sachs predicts that in the next five years, the demand for computing power may grow approximately 24 times, and tight supply and demand is expected to last for a longer period, benefiting companies like Microsoft, Oracle, and CoreWeave in the cloud computing and infrastructure sector. This article is adapted from "Wall Street Watch," author: Li Jia, GMTEight editor: Li Cheng.