Open weights AI lowers costs and boosts compute

Morgan Stanley Research argues that cheaper open-weights AI models could increase, rather than reduce, total demand for computing infrastructure as companies expand the number of tasks they can automate. The firm’s central case is a Jevons-paradox effect: lower inference costs make more AI workloads economical, pushing call volumes higher even when each individual task requires fewer resources.

That dynamic would preserve demand for GPUs, electricity, networking, security, and deployment software under several possible outcomes for the AI market, Morgan Stanley said. The revenue opportunity may shift between model developers, cloud providers, enterprise data centers, and infrastructure vendors, but the underlying need to run AI workloads would remain.

The report cautioned against treating open-weights models as a simple substitute for expensive proprietary systems. Companies adopting them still face costs for chips, cloud capacity, private facilities, fine-tuning, engineers, security controls, and ongoing operations. Whether self-hosting delivers savings depends heavily on the size of the model, the regularity of the workload, and how fully the infrastructure is used.

Lower model prices widen the set of viable AI tasks

Morgan Stanley cited an estimate that the average enterprise AI task can generate roughly $55 in value while carrying direct costs of about $2 to $5. A fall in inference pricing could therefore make AI worthwhile for many tasks that previously failed internal payback tests.

The effect is likely to be most visible in high-volume, repeatable work rather than a handful of headline-grabbing AI applications. Morgan Stanley identified customer-service record tagging, contract review, code testing, product-description generation, enterprise search, marketing content, data cleaning, and internal approval processes as examples of workflows that could move into production as costs fall.

A business might use a relatively costly frontier model for a complex legal analysis or a long sequence of agent-driven planning steps, while deploying a smaller open model to classify millions of support tickets. Multiplying the number of lower-cost calls can raise aggregate token use and electricity consumption, even if the computing load attached to each call is modest.

Morgan Stanley’s framing places the commercial question less on whether enterprises will spend money on AI infrastructure and more on where that spending will occur. Closed-model providers may capture API revenue, while companies running open models could direct more of their budgets toward cloud capacity, privately owned servers, model-management tools, power systems, and cybersecurity.

Enterprises are already mixing open and closed models

A McKinsey survey referenced by Morgan Stanley found that 63% of respondents were already using open models in their AI stacks. Many of those companies also retain closed models, indicating that the market has not settled into a single model architecture.

Open-weights systems, whose downloadable parameters allow companies to run and customize models in their own environments, are often used for coding, document parsing, frequent inference calls, and specialized domain tasks, according to Morgan Stanley. Closed models remain common where businesses require complex reasoning, higher reliability, or capabilities that are difficult to standardize across internal systems.

Usage data referenced in the report also points to growing developer interest in Chinese open models. Between February and July 2026, the weekly share of tokens routed by U.S. companies through OpenRouter to China-based open models exceeded 30% at one point, Morgan Stanley said. The firm added that the measure is likely weighted toward developers and startups using the routing platform, rather than representing spending patterns across the largest enterprises.

That distinction matters when assessing how quickly corporate infrastructure budgets could move. A developer can switch models through an API router rapidly, while a large company moving sensitive workloads to self-hosted infrastructure must address data governance, model testing, procurement, capacity planning, and security.

Self-hosting economics vary sharply by workload

Morgan Stanley cited an MIT study estimating that a shift from closed to open models could reduce average prices by about 70% and save consumers approximately $25 billion annually. The report noted that the research was conducted earlier in the market’s development, that model capabilities are not always directly comparable, and that engineering and operational costs may be understated.

A Carnegie Mellon study cited by Morgan Stanley produced a similarly conditional picture. It estimated that self-hosting open models can reach payback in as little as three months or take as long as six years.

Smaller models handling fixed, high-utilization tasks tend to produce the quicker payback periods, according to the report. Large models used for complex enterprise work can become far more expensive to operate when usage is inconsistent or when teams must repeatedly update models, fine-tune them for internal data, and maintain specialized infrastructure.

Those costs help explain why a hybrid model market remains a plausible outcome. Companies do not necessarily need to choose between one proprietary provider and a fully self-managed model stack. They can route workloads based on accuracy, speed, price, and data sensitivity.

Distributed deployments create a larger software and security layer

Morgan Stanley outlined three broad market scenarios. In the first, closed models retain a meaningful performance, reliability, and safety advantage, keeping workloads concentrated in hyperscale cloud data centers. The report estimated that Google, in an illustrative case where Gemini holds a leading model position and runs on Google infrastructure, could generate return on invested capital of about 45%; the figure could remain near 30% when the company acts mainly as an infrastructure provider.

The second scenario envisions a durable hybrid market. Frontier closed models would handle difficult reasoning and long-horizon agent tasks, while open-weights and smaller models would process high-frequency or cost-sensitive work. That arrangement spreads AI workloads across public clouds, private clouds, corporate data centers, and edge devices.

Morgan Stanley expects such a structure to increase demand for AI gateways that manage authentication and logging, routing software that selects models for each job, orchestration systems for multi-step agent workflows, and observability tools that track quality, latency, token consumption, and failures.

More distributed deployments also add governance burdens. Companies operating models across several environments need controls over data access and tool permissions, along with defenses against prompt injection, model extraction, weight tampering, and privilege misuse. Security software therefore remains relevant whether AI workloads are centralized in a hyperscaler’s facilities or spread across private and sovereign infrastructure.

In Morgan Stanley’s third scenario, open-weights models narrow the capability gap with leading closed systems and drive API prices lower. Enterprise budgets would move more heavily toward inference optimization, agents, industry-specific tools, and deployment in sovereign clouds or on-premises environments where latency and data-control requirements justify local operation.

Cloud platforms would retain substantial roles in that outcome. They can host open models while selling GPU compute, storage, networking, data services, and enterprise AI platforms, though the mix of revenue would differ from a market dominated by proprietary API access.

Across all three cases, Morgan Stanley identified NVIDIA, on-site power exposure, and cybersecurity software as recurring infrastructure beneficiaries. It also cited Bloom Energy, Williams, and Liberty Energy among companies with exposure to on-site generation, natural gas, or energy infrastructure associated with rising AI electricity requirements.

The report does not establish that every lower-cost model deployment will translate into more total compute. Some companies may simply replace existing AI calls with cheaper alternatives. Its broader conclusion rests on a more practical enterprise pattern: when the cost of processing documents, serving customers, testing code, or searching internal data falls far enough, organizations tend to apply the technology to many more routine tasks.