Will Open Source Models Kill 'Computing Demand'? Morgan Stanley Explores Three Futures of AI

Original Report: Morgan Stanley Research "Weighing In: Open-Weights Models & 3 States of the World", August 3, 2026

Compiled by: DaiDai, Frank, MSX Maitong Research Institute

Key Insights

• Open-weight models reduce the cost per call and deployment barriers, but do not necessarily lower total computing demand. As AI enters more enterprises, workflows, and devices, the increase in call volume may outpace efficiency gains, triggering the classic "Jevons Paradox";
• Enterprises have already entered a multi-model era, with open-weight models primarily handling programming, document processing, high-frequency calls, and specific domain tasks, while complex reasoning and agent workloads still rely more on cutting-edge closed-source models;
• Open weights do not equate to free, as the savings for enterprises may come from model licensing fees or some API costs, but they still need to bear costs for GPUs, cloud services, local data centers, fine-tuning, talent, security, and operations;
• Regardless of whether closed, hybrid, or open-weight models ultimately dominate, NVIDIA, on-site power, and security software are clearly defined beneficiaries across scenarios;
• The more prevalent open-weight models become, the more likely the AI value chain will spread from foundational model layers to inference, routing, orchestration, observability, local infrastructure, edge devices, and vertical applications;

Every so often, the AI market seems to experience a wave of "efficiency panic".

When model parameters shrink, inference costs drop, or a model company achieves near-state-of-the-art performance with fewer chips, the market quickly develops an intuition that, since the computational resources needed to accomplish the same tasks are decreasing, will the demand for GPUs, data centers, and electricity also peak?

The emergence of next-generation open-weight models like Kimi K3 and DeepSeek V4 Flash brings this question back to the forefront. They not only attempt to narrow the capability gap with cutting-edge closed-source models but also allow enterprises to download, modify, and deploy models independently, spreading capabilities that were once highly concentrated in a few U.S. model labs to a broader range of developers and enterprise tech stacks.

However, Morgan Stanley's latest report provides an answer that is precisely the opposite of this intuition:

Increased model efficiency may not kill computing demand; rather, it is more likely to make AI cheap enough to enter scenarios where it was previously not worth using, ultimately driving up overall call volume.

Based on this, the real discussion that needs to be revisited is where computing will occur, who will provide it, and which companies can profit from this diffusion.

1. Cheaper Models: Why Might They Consume More Computing Power?

One common misunderstanding about open-weight models is equating "lower model costs" directly with "lower infrastructure demand".

For most enterprises, the decision to adopt AI does not solely depend on how much computing power is needed to complete a task, but rather on whether the revenue it generates can cover the costs of the model, engineering, and infrastructure.

Morgan Stanley's calculations indicate that an enterprise AI task can average about $55 in revenue, while direct costs are around $2-5. Even considering data, engineering, security, and management costs, this revenue-to-cost ratio still suggests that many enterprise workflows have yet to be fully AI-enabled.

In this context, the result of model price reductions may not be a decrease in AI spending by enterprises, but rather more tasks crossing the threshold of economic feasibility.

In the past, enterprises might have assigned only the most critical, high-value tasks to AI; as inference prices continue to drop, tasks such as customer record classification, contract review, code testing, product descriptions, enterprise search, marketing materials, data cleaning, and even internal approvals could all be included in model workflows.

While the computing resources required for each task may decrease, the number of tasks, frequency of execution, and user scale may expand simultaneously.

This is precisely the "Jevons Paradox" that the report repeatedly mentions: when the efficiency of a resource increases and unit costs decrease, its total consumption may actually rise due to the rapid expansion of its application scope.

For example, more fuel-efficient cars have not led to a disappearance of global oil demand; cheaper internet bandwidth has not resulted in decreased data traffic. Similarly, more efficient models may not reduce GPU usage but could instead transform AI from a few high-value tasks into a ubiquitous foundational capability within enterprises.

Currently, this diffusion has already begun to occur.

A McKinsey survey cited in the report shows that 63% of surveyed enterprises are already using open models in their tech stacks, but most have not completely abandoned closed-source models; rather, they are using a combination of both, with open-weight models primarily utilized for programming, document parsing, high-frequency calls, and specific domain tasks, while cutting-edge closed-source models continue to handle complex reasoning, high reliability requirements, and tasks that are harder to standardize.

From February to July 2026, the proportion of tokens routed to Chinese open models via OpenRouter by U.S. companies exceeded 30% weekly at one point. This data may skew towards developers and startups and may not directly represent large enterprise spending, but it at least indicates that open-weight models have moved from laboratory concepts into real calling and deployment environments.

Of course, open-weight models do not equate to "free models".

When self-hosting, enterprises can avoid paying API fees calculated per token to model providers, but they still need to purchase or lease GPUs and bear costs for data centers, cloud services, fine-tuning, engineering teams, security, and daily operations. This means that using open-weight models via hosted APIs provided by model vendors or cloud platforms may still incur costs based on tokens or computing volume.

Therefore, what open weights change is not whether computing costs exist, but how enterprises pay for those computing costs, and whether value is derived from model providers, cloud platforms, or the enterprises' own infrastructure.

Morgan Stanley cites an MIT study estimating that transitioning from closed-source models to open models could reduce average prices by about 70%, saving consumers approximately $25 billion annually. However, the report also specifically notes that this study was completed some time ago, and the capabilities of different models are not entirely comparable, nor does it necessarily fully account for hidden costs such as engineering, fine-tuning, and operations.

Another study from Carnegie Mellon indicates that the payback period for self-deployed open models may extend from about 3 months to 6 years, where smaller models with fixed tasks and high call frequencies can more quickly amortize hardware costs; large models and complex enterprise applications may take a long time to prove that self-hosting is cheaper than using APIs due to insufficient utilization, frequent updates, and high fine-tuning costs.

Thus, the economics of open-weight models do not have a unified answer.

It depends on model scale, usage frequency, infrastructure utilization, enterprise engineering capabilities, and whether data must remain local. The more open the model, the more choices enterprises have, but the more technical responsibilities they also bear.

2. In the Next Phase, Where Will Computing Occur?

If the future of AI is dominated by a few closed-source models, then training and inference will continue to concentrate in large cloud platforms and massive data centers.

However, if open-weight models gain broader adoption, AI computing will not disappear but will instead diffuse from a few central nodes to private clouds, enterprise data centers, sovereign data centers, edge servers, and personal devices.

This means that the variables the market needs to focus on will shift from "how many GPUs are needed" to "where GPUs are deployed, who manages them, and how they are called".

In the closed-source model era, enterprises could outsource most complexities to model companies: connect to an API, pay per token, with providers responsible for model updates, infrastructure, security alignment, and some legal liabilities.

As we enter the multi-model era, this simple structure begins to break down.

Enterprises may assign the most complex tasks to cutting-edge closed-source models while delegating high-frequency and cost-sensitive tasks to open-weight models; sensitive data remains local, while general workloads run on public clouds; some requests are processed in data centers, while others are completed directly on computers, phones, or other edge devices.

The more models there are and the more dispersed their deployment, the more complex the enterprise AI tech stack will become:

• First, the gateway layer: enterprises need a unified entry point to manage authentication, calling permissions, rate limits, logging, and vendor failover for different models;
• Second, the routing layer: systems need to allocate each request to the most suitable model based on accuracy, latency, cost, data sensitivity, and task difficulty;
• Next, the orchestration layer: complex agent workflows often require multiple models, databases, and external tools to collaborate, with a single user request potentially broken down into dozens of steps and calls;
• Finally, the observability and evaluation layer: enterprises must continuously track model output quality, response latency, token usage, operational costs, failure reasons, and security risks.

This is also why the diffusion of open-weight models benefits not only model vendors.

As foundational models become more accessible, enterprises are more willing to pay for "how to stably deploy models into production." Model gateways, task routing, agent orchestration, data governance, observability, and security software may become the segments of the value chain that are harder to compress.

The importance of security is particularly pronounced. When closed-source models run centrally, some security responsibilities are borne by model labs and cloud platforms; however, as open models enter enterprise local environments, private clouds, and edge devices, identity, data, endpoints, model weights, and runtime environments all need independent protection.

Enterprises not only need to prevent employees from sending sensitive data to the wrong models but also need to manage which databases different models can access, which tools they can call, and whether they may be subject to prompt injection, model distillation, weight tampering, or permission abuse.

The more dispersed the model deployment, the larger the attack surface; the more models there are, the higher the governance costs.

Therefore, what open-weight models truly bring is not the disappearance of AI infrastructure value, but rather the diffusion of value from a single foundational model and API layer to the entire system stack.

Future AI spending may no longer only reflect a few tech giants building super training clusters but will also manifest as enterprises purchasing servers, upgrading networks and storage, deploying security systems, building private AI platforms, and configuring stronger edge computing power on mobile phones and PCs.

The shift of computing demand from centralized to distributed does not mean a decrease in total volume; it simply means that beneficiaries are no longer concentrated solely among model labs and large cloud vendors.

3. Three Worlds of AI: Regardless of Who Wins, Computing, Power, and Security Are Inescapable

Morgan Stanley did not assign explicit probabilities to the three future scenarios but instead outlined the potential value distribution in the AI industry chain under the conditions of a closed-source model prevailing, a mixed architecture coexisting, and an open-weight model prevailing.

Scenario One: Closed-Source Models Maintain Cutting-Edge Capabilities

In this scenario, the performance, reliability, and security capabilities of cutting-edge models remain difficult to replicate, with a few capital-rich model labs continuing to lead.

Companies prioritize accuracy, ease of deployment, intellectual property compensation, and brand credibility over complete control of model weights, thus willing to continue paying for closed-source APIs and enterprise subscriptions.

Training and inference workloads will further concentrate on hyperscaler platforms like AWS and Google Cloud, with super training clusters continuing to drive demand for GPUs, high-speed networks, optical communications, and custom ASICs.

Platforms like Google and Amazon, which possess cloud infrastructure and model capabilities, are in a more advantageous position, and chip and network manufacturers such as Broadcom, Arista Networks, Lumentum, and Coherent may also benefit.

The report estimates that if Google can run the Gemini API on its own infrastructure with its leading model, the implied capital return could reach about 45%; even if the model is not in an absolute leading position and serves merely as an infrastructure provider, the relevant return could still approach 30%.

This indicates that the most important asset in a closed-source world is not just the model itself, but also the ownership of the computing power on which the model runs. Whoever owns the chips, data centers, and customer access has a greater ability to retain the profits generated from model calls.

Scenario Two: Long-Term Coexistence of Open and Closed-Source Models

This is the scenario that most closely resembles the actual usage state of enterprises today.

Closed-source cutting-edge models handle complex reasoning, long-process agents, and high-reliability tasks, while open-weight models and smaller models deal with high-frequency, cost-sensitive, low-latency, or highly specialized work.

Enterprises will not choose just one model but will switch and route based on tasks, which also means that AI deployments will exist simultaneously in public clouds, private clouds, on-premises infrastructure, and edge devices.

In this scenario, it is difficult for a single model vendor to gain complete monopolistic pricing power, but the entire AI software and infrastructure market will see the broadest demand.

• Microsoft, Amazon, and Google will still benefit from cloud workloads;
• Infrastructure and workflow software vendors like Datadog, Palantir, and Appian may benefit from model orchestration, data connectivity, and observability demands;
• Security vendors such as Palo Alto Networks, CrowdStrike, Fortinet, Zscaler, Netskope, and Okta will benefit from the continuous expansion of enterprise attack surfaces and identity boundaries;

At the same time, network device manufacturers like Cisco and F5, as well as enterprise infrastructure companies like Dell, HPE, and NetApp, may also gain new demand from on-premises and hybrid deployments.

The biggest investment implication of a mixed architecture is that AI spending will not be concentrated solely on training clusters but will spread layer by layer along the enterprise technology stack, and the more intense the competition at the model layer, the more enterprises will need neutral software and infrastructure to help them switch between different models and deployment environments.

Scenario Three: Open-Weight Models Become Mainstream

In this scenario, open-weight models gradually approach the capabilities of closed-source cutting-edge models, with foundational model intelligence becoming widely available and API prices significantly decreasing.

Enterprises are no longer willing to pay excessive premiums for general model capabilities but will fine-tune using private data, allocating their main budgets to inference optimization, agents, industry tools, and application deployment.

The focus of innovation shifts from pre-training to post-training and application layers, and AI infrastructure becomes more decentralized. Governments and large enterprises may deploy more models in sovereign clouds, private data centers, and on-premises environments for reasons of data sovereignty, privacy, latency, and avoiding vendor lock-in.

• Microsoft may gain a stronger position through Azure, GitHub, enterprise software, and the open model ecosystem;
• The strategic value of open model providers like MiniMax, ZhiPu/Z.ai, Alibaba, and Tencent will rise;
• Local infrastructure vendors like Dell, HPE, and NetApp, as well as endpoint device companies like HP and Apple, and system integrators and IT distribution channels will also see more significant demand;

However, an open world does not mean that cloud platforms lose value.

Most enterprises will still not fully build their own infrastructure, and open models may still run on Azure, AWS, or Google Cloud, with hyperscalers transitioning from being the sole model entry point to providers of open model hosting, computing resources, data services, and enterprise AI platforms.

Thus, cloud vendors face not a simple win or lose situation but a change in profit structure; for example, foundational model rents may decrease, but revenues from computing, storage, databases, security, and enterprise services may still grow.

Common Winners Among the Three Outcomes

In the asset matrix listed by Morgan Stanley, the most noteworthy aspect is not the companies unique to each scenario but those names that repeatedly appear across all three columns.

First is NVIDIA.

The success of closed-source models means more computing power is concentrated in large-scale training and inference clusters; the success of open-weight models means that inference nodes will spread to enterprises, on-premises, and edge environments.

The chip forms, customer structures, and individual cluster sizes corresponding to the two paths may differ, but both require continuous increases in computing power. Ultimately, what open models lower is the barrier to model usage, not the elimination of the computing process.

Second is electricity.

The closed-source world requires stable GW-level power to support super data centers, while the open world will increase the power demand for enterprise data centers, sovereign clouds, and on-site inference facilities. Improved model efficiency may reduce the power consumption per task but could also allow more tasks and devices to continuously run AI.

Companies like Bloom Energy, Williams, and Liberty Energy, which provide on-site power, natural gas, and energy infrastructure, have thus been repeatedly included in the list of beneficiaries in the report.

Third is security software.

Whether enterprises use closed-source or open models, they need to protect identities, data, and applications; as deployments spread from a few cloud platforms to on-premises and edge environments, security demands will only become more complex.

From this perspective, security software may be an underestimated secondary beneficiary in the spread of open-weight models.

Conclusion

Overall, what open-weight models may truly kill is perhaps never the demand for computing power, but the imagination that foundational model layers can maintain high rents over the long term.

As model capabilities gradually become widespread, enterprises will no longer pay just to "gain intelligence" but will allocate more budgets to how to deploy intelligence, connect data, manage workflows, protect systems, and ensure models continuously generate returns in real business.

Computing power will not simply shift from growth to decline—it may spread from a few super training clusters to more inference nodes, enterprise data rooms, sovereign clouds, and edge devices; from one-off model training competitions to long-term infrastructure investments covering every workflow in the enterprise.

Therefore, the biggest change brought by open-weight models is not that the AI CapEx cycle is about to end, but that this cycle is entering a new stage with more participants, more decentralized deployments, and a longer value chain.