Morgan Stanley: Open-Weight Models May Lower AI Costs but Could Increase Compute Demand via Jevons Paradox

Morgan Stanley (Morgan Stanley) in its latest report "Open-Weight Models and Three Future Scenarios" points out that open-weight (Open-Weights) models do not necessarily weaken AI compute demand.

On the contrary, lower usage costs may accelerate AI adoption, potentially triggering the classic "Jevons Paradox" (Jevons Paradox): After the cost per inference decreases, companies may apply AI to more work scenarios, ultimately increasing the overall demand for Tokens, compute power, electricity, and infrastructure.

According to a report by _Caixin Global_, the report also specifically explains the difference between "open source" (Open Source) and "open weights" (Open Weights). The market often confuses the two, but in reality, they are not the same.

Open-weight models publish the model parameters after training, allowing users to download, deploy, and fine-tune the model, but they do not necessarily publish the training data or complete program code, and the license terms are not necessarily fully open. Meta's Llama, Google's Gemma, and China's Qwen, DeepSeek, and Kimi AI models all belong to this category.

As for true open-source models, in addition to publishing model weights, they also publish program code, training processes, and data details. The representative case listed in the report is AI2's OLMo.

The main difference between open-weight and closed-weight (Closed Weights, also commonly referred to as closed models) is not in the model scale, speed, or efficiency, but in how users obtain, modify, and deploy the model.

Closed-weight models are usually hosted by model suppliers, and companies must pay subscription fees or license seat fees and pay based on Token usage. Open-weight models are mostly free to obtain, but companies still need to bear the costs of GPU, cloud services, training, fine-tuning, operations, and information security. Companies can build their own computing power or rent cloud GPUs by the hour, or use hosted APIs to charge by Token.

Therefore, open-weight only waives the model licensing fee, and does not mean that the total cost of ownership (TCO) is zero.

Current corporate adoption situation

According to a survey by McKinsey of 41 countries and 700 technology executives, 63% of respondents have already started adopting open-source models, and most of them use them in parallel with closed-weight models. From February to July 2026, more than 30% of OpenRouter Tokens used by American users were directed to Chinese open-weight models.

However, Morgan Stanley points out that this data may more reflect the usage situation of startups and technical teams. Currently, the share of open-weight models in corporate large language model (LLM) applications and overall model spending is still a minority.

The reason is that in addition to price, large enterprises also attach great importance to the reliability of the model, brand reputation, technical support, copyright protection, and continuous upgrade capabilities. Most enterprises currently still tend to adopt the standardized products of model suppliers or directly cooperate with cutting-edge AI laboratories to customize models, in order to reduce the engineering burden brought by self-training and maintenance.

Currently, open-weight models are mainly applied to software development, document parsing, customer service, and other professional fields. These applications usually have three common characteristics: high usage frequency, high requirements for response speed, or the need to combine with corporate internal data for customization.

The main advantages of open-weight

The biggest attraction of open-weight models is that they can effectively reduce the variable costs of high-frequency inference.

The report cites research by the Massachusetts Institute of Technology (MIT) that if companies switch from closed-weight models to open models, the average cost can be reduced by up to 70%, saving about $250 billion annually for global users.

For companies that need to execute tens of billions of multimodal inferences every day, if they continue to charge by Token, the cost may be difficult to bear, so open-weight models have considerable appeal.

However, the report also proposes three limitations.

First, this research was completed in December 2025; second, not all open models have lower costs per Token, for example, Kimi K3 is an exception; in addition, companies still need to bear the costs of computing power, hosting, training, and fine-tuning.

The deployment scale will also directly affect the investment recovery period.

According to research by Carnegie Mellon University, if the deployment scale is less than 300 billion parameters, mainly applied to scenarios such as customer service and document analysis, the investment can be recovered in as little as about 3 months; if 2000 billion to 1 trillion parameters are deployed for large models, used for advanced research or enterprise-wide generative AI applications, the investment recovery period may be as long as 6 years.

Therefore, the economic benefits of open-weight models are not simply measured by "free replacing paid," but depend on the model usage volume, GPU utilization, model scale, and the company's own technical capabilities.

Risks and limitations

In addition to cost, another important advantage of open-weight models is that companies have higher deployment autonomy.

Companies can deploy models on-premises or in private clouds to meet data sovereignty, local storage, and industry regulatory requirements. Surveys show that about 40% of corporate executives choose to adopt open-weight models precisely because they can decide where to deploy the model, especially in highly regulated industries such as finance, healthcare, and government.

However, having higher deployment autonomy also means that companies must bear more responsibilities.

The report points out that after the open model is released, developers have difficulty restricting how third parties modify the model weights.

For example, in April 2026, the British _Financial Times_ and the AI security organization Alice used the Heretic tool to remove the security protection mechanisms of Meta and Google's open models in just a few minutes. Malicious use, model distillation attacks, and corporate internal data leaks may all become risks that companies have to bear themselves.

In terms of intellectual property rights, commercial models such as OpenAI and Anthropic usually provide a certain degree of copyright protection under the premise that enterprises comply with security regulations and use the models correctly; in contrast, open-weight models generally do not provide similar mechanisms, and related legal responsibilities may need to be borne by the deploying enterprises themselves.

Geopolitics is also an important variable. If the United States restricts open-weight models from China, it may affect the continuity of corporate operations. Even if American companies recognize that Chinese models have higher cost-effectiveness, they may still turn to American open models or closed model suppliers due to concerns about future policy changes.

Other adoption barriers include conversion costs, brand trust, and some companies' concerns about foreign models.

Morgan Stanley presents three possible scenarios for industry development.

Closed-weight models win

If the capabilities of cutting-edge models are always difficult to replicate, the market will continue to be dominated by a few well-funded AI laboratories.

Companies will be willing to pay a premium for higher accuracy, reliability, intellectual property protection, and deployment convenience, and training and inference work will continue to be concentrated on super-large cloud platforms.

In this scenario, Google, Amazon, Meta, NVIDIA, Broadcom, Arista Networks, Lumentum, Coherent, and other optical communication and network equipment companies will be the main beneficiaries.

If Gemini 4 helps Google regain its leading position in AI models, Morgan Stanley estimates that its incremental capital return rate for operating model APIs on its own infrastructure will be approximately 45%; if it mainly provides infrastructure services, it will be approximately 30%.

Hybrid architecture becomes mainstream

Closed-weight models are responsible for complex inference and high-level tasks such as AI agents; open-weight models and small models undertake high-traffic, cost-sensitive, and vertical application scenarios.

The same set of corporate workflows may also dynamically switch between different models according to cost and model capabilities.

In this scenario, Amazon, Google, and Microsoft, the three major cloud platforms, are expected to benefit; the value of model orchestration, infrastructure software, security, SaaS, and network equipment suppliers will also be enhanced simultaneously.

SAP, ServiceNow, Cisco, and F5 are expected to become important beneficiaries.

Open-weight models win

If the performance of open-weight models gradually approaches the most advanced models, the basic model capabilities will become more popular, and the API price will also drop significantly.

The focus of future innovation will shift from large-scale pre-training to model fine-tuning, inference optimization, AI agents, tools, and various dedicated applications.

AI deployment will also become more decentralized, and some workloads will be transferred from public clouds to private data centers, sovereign clouds, local servers, and edge devices.

Microsoft, Alibaba, Tencent, MiniMax, and other model or cloud platform companies, as well as Dell, HPE, NetApp, HP, Apple, TD Synnex, and other companies are expected to benefit.