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

Author: Li Jia, Wall Street Journal

AI is entering a new phase from "being able to answer" to "being able to execute."

According to the Wind Trading Desk, Goldman Sachs' latest report shows that the commercialization of AI is shifting from "subscription by seat" to charging based on consumption, transaction volume, and outcomes; at the same time, Agents are transitioning from auxiliary tools to workflow executors, and the industry value is shifting from the model itself to proprietary data, business context, and domain expertise.

This means that competition in the AI industry is shifting from "whose model is stronger" to "who can truly master the workflow." Model capability remains important, but the ability to enter enterprise production environments, understand business context, and reliably complete tasks will become a more critical competitive barrier.

This judgment comes from Goldman Sachs' recent field investigation of 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 cutting-edge models, open-source models, world models, enterprise software, and proprietary data will change.

Agent Implementation: What Enterprises Truly Lack is Not Capability, but "Controllability"

If in the past AI solved the problem of "helping people complete tasks," then Agents are attempting to solve the problem of "completing tasks on their own." However, during large-scale deployment in enterprises, the biggest obstacle may no longer be model capability, but how responsibilities are divided and whether the entire execution process can be controlled.

The report cites Stanford researchers who point out that most enterprises are still in a mode of human supervision. Especially in fields such as law, risk control, insurance, and auditing, if a model makes an error, who bears the responsibility, how to trace the process, and whether it can be corrected in time may be as important as the model's capability itself.

Therefore, the workflows that are most likely to achieve automation first typically have three characteristics: clear decision boundaries, verifiable results, and reversible errors. Invoice processing is a typical case. AI is responsible for extracting fields and performing checks, with low-confidence cases reviewed by humans, and then completed through a reversible ERP process.

This also means that information service providers with trustworthy content, validated domain models, and mature regulatory relationships are more likely to enter enterprise production environments first.

Model Competition: Division of Labor Between Cutting-Edge Models and Open-Source Models

Regarding the debate of "open-source or closed-source," the signals released by Goldman Sachs during this investigation indicate that it is not a binary choice, but rather different models may correspond to different levels of workflows.

The cutting-edge model camp believes that enterprise benchmarking often underestimates model capability. In real production environments, the business losses caused by a decline in model accuracy may far exceed the savings in inference costs. Therefore, although many AI-native companies claim to adopt a multi-model strategy, they still heavily rely on cutting-edge models in their core production environments.

Another viewpoint suggests that the vast majority of enterprise workflows do not require cutting-edge intelligence. As the performance of open-source models continues to improve, customers are increasingly willing to trade off limited performance loss for lower inference costs. A venture capital firm predicts that in the next 12 to 18 months, about 90% of inference tokens will flow to open-source models.

This means that the future AI model market may form a clearer division of labor: cutting-edge models will be responsible for high-value, high-reliability complex tasks, while open-source models will handle larger-scale standardized tasks and the majority of token consumption.

World Models: AI Computing Power May Welcome a Second Growth Curve

In 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, with data coming more from physical systems, specific industries, and actual operational scenarios. This means 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 robotics is much larger than pure text generation, and these workflows often require higher computing power investment. As AI further enters the physical world from the digital world, the demand for computing power in model training, simulation, and inference may also see a new growth curve.

Goldman Sachs predicts that in the next five years, the demand for computing power may grow by about 24 times, and the supply-demand tension is expected to last longer, benefiting cloud computing and computing infrastructure companies like Microsoft, Oracle, and CoreWeave directly.