Open-weight models will accelerate AI diffusion, with industry winners depending on three competitive landscapes

Overview

The report discusses the competition and coexistence of open-weight models and closed-source models in enterprise AI technology stacks. Morgan Stanley believes the market may be underestimating the role of open-weight models: their core distinction is not model size or quality, but whether users can access and adjust weights, and whether they can deploy flexibly in public cloud, private cloud, on-premises or edge environments. Open-weight models can reduce the variable cost of high-frequency inference, support specialized customization and meet data sovereignty requirements, but they also bring fine-tuning costs, talent bottlenecks, security responsibilities and geopolitical risks. The distribution of industry value over the coming years depends on which of the three scenarios—closed-source, hybrid or open-weight—dominates.

Core views

First, open-weight models are conducive to competition and technology diffusion, and price declines may expand total AI demand through Jevons paradox rather than simply weakening computing power demand. Second, enterprises are more likely to maintain hybrid architectures in the near term: closed-source frontier models handle complex reasoning and agentic tasks, while open or small models take on high-frequency, cost-sensitive and domain-specific work. Third, the more dispersed the number of models and deployment locations, the greater the importance of gateways, routing, orchestration, evaluation, observability, security and governance. Fourth, if closed-source models win, leading labs, hyperscale clouds, training networks, semiconductors and on-site power supply benefit more; if open-weight models win, value will shift more toward fine-tuning, inference optimization, applications, private cloud, on-premises infrastructure, edge devices and security software.

Analysis framework

The report combines enterprise adoption surveys, token usage data, total cost of ownership and return-on-investment studies, and uses scenario analysis to compare model performance, pricing power, deployment architectures, regulatory constraints and enterprise procurement behavior, then maps each scenario to relevant stocks and industry segments.

Methodology notes

• Scenario analysisThree states of the world

  Assumes respectively that closed-source models win, hybrid architectures coexist and open-weight models win.

  By changing model performance gaps, price competition, enterprise deployment methods and regulatory conditions, the report identifies the relative beneficiaries across cloud, models, software, hardware, networking and energy segments under different scenarios.

• Economics frameworkJevons paradox

  A decline in unit inference costs may stimulate usage growth and increase total AI demand.

  After open-weight models improve efficiency and push down prices, more enterprises and workloads become economically viable for AI adoption, so demand for computing power and related infrastructure may not necessarily decline as unit costs fall.

• Cost analysisTotal cost of ownership

  Compares licensing fees, token fees, computing power, hosting, fine-tuning, talent and operations and maintenance costs.

  Although open-weight models can avoid licensing fees or per-token charges, they are not free; deployment scale and model parameter count significantly affect their relative economics and return-on-investment period.

• Industry chain mappingBeneficiary analysis of AI deployment architectures

  Allocates industry beneficiary directions according to the extent to which workloads migrate to public cloud, private cloud, on-premises and edge.

  Centralized closed-source architectures are more favorable to leading cloud and frontier model providers, while distributed open architectures increase demand for on-premises hardware, edge devices, security, orchestration, routing and observability.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• GOOGL, AMZN, MSFT

  Cloud platforms and model distribution channels

  Strengths

  They have scaled computing power, enterprise customer relationships and hosting capabilities; under closed-source or hybrid architectures, they can absorb large amounts of centralized workloads.

  Weaknesses

  Price competition from open models may compress excess returns from model APIs and push some workloads to on-premises or edge environments.

  Comparison

  Key beneficiaries in the closed-source scenario include GOOGL and AMZN; the hybrid scenario includes AMZN, GOOGL and MSFT; in the open-weight scenario, the report highlights MSFT.

  Risks

  Proprietary models lagging frontier models, declining returns on capital expenditure, customer multi-cloud adoption and regulatory restrictions.

• NVDA, AVGO

  AI semiconductors and custom computing infrastructure

  Strengths

  Growth in training and inference demand directly supports demand for accelerators, networking and custom chips; NVDA is viewed as a beneficiary with relatively strong resilience across scenarios.

  Weaknesses

  Smaller and more efficient models may reduce the computing power required per task, and open architectures may also change chip and system procurement structures.

  Comparison

  The closed-source scenario lists NVDA and AVGO, while the hybrid scenario lists NVDA; the report's front page emphasizes that NVDA has a beneficiary logic across multiple states.

  Risks

  Model efficiency improving faster than workload growth, customers' in-house chips, supply expansion and high valuation.

• PANW, CRWD, FTNT, ZS, NTSK, OKTA, SAIL, VRNS

  Security and identity governance for distributed AI workloads

  Strengths

  After models and data are distributed across cloud, on-premises and edge environments, demand for access control, endpoint security, identity management and governance increases.

  Weaknesses

  Under centralized closed-source architectures, some security responsibilities may be handled internally by frontier model labs or cloud platforms.

  Comparison

  Security software benefits in all three scenarios, but incremental demand is broader under open-weight or hybrid architectures.

  Risks

  Security budget consolidation, substitution by platform-native features, malicious modification of open models and unclear responsibility attribution.

• DDOG, PLTR, APPN

  Orchestration, observability and infrastructure software

  Strengths

  Multi-model environments require unified gateways, routing, logging, evaluation, cost monitoring, workflow orchestration and failover.

  Weaknesses

  Functions may be built into cloud platforms, model providers or open-source tools, and independent vendors face platform competition.

  Comparison

  The hybrid scenario highlights DDOG, PLTR and APPN, while the open-weight scenario lists PLTR.

  Risks

  Enterprise deployment progress slower than expected, product commoditization, open-source alternatives and customers' in-house development.

• DELL, HPE, NTAP, P, HPQ, AAPL

  On-premises infrastructure and edge devices

  Strengths

  After open model performance improves, enterprises may move more inference to on-premises and edge environments due to cost, latency, privacy and data sovereignty requirements.

  Weaknesses

  On-premises deployment of large models is capital intensive, the payback period may be long, and enterprises need operations and maintenance talent.

  Comparison

  These companies mainly benefit from the decentralized scenario in which open-weight models win, with weaker elasticity under the centralized closed-source scenario.

  Risks

  Enterprises continuing to prioritize APIs, declining public cloud costs, longer hardware refresh cycles and excessive model deployment complexity.

• SAP, NOW, SHOP

  Enterprise applications and SaaS layer

  Strengths

  They can embed open or closed-source models into existing workflows and use enterprises' proprietary data to develop domain applications and agents.

  Weaknesses

  Commoditization of foundation model capabilities may reduce the differentiation and pricing power of some AI functions.

  Comparison

  SAP and NOW appear in both the hybrid and open-weight scenarios, while SHOP is mainly listed in the open-weight scenario.

  Risks

  Insufficient monetization of AI functions, inference cost pressure, data governance obstacles and customers building applications in-house.

• ANET, LITE, COHR, CSCO, FFIV

  Data center optical communications, networking and traffic management

  Strengths

  Closed-source training clusters require high-bandwidth networks, while hybrid and distributed architectures increase demand for connectivity and traffic management across clouds, data centers and the edge.

  Weaknesses

  Beneficiary names diverge as the shares of centralized training and distributed inference change.

  Comparison

  The closed-source scenario highlights ANET, LITE and COHR; the hybrid scenario highlights CSCO and FFIV.

  Risks

  Data center capital expenditure cycles, customer concentration, technology iteration and supply chain volatility.

• BE, INIO, SEI, WMB, LBRT

  On-site power supply for hyperscale AI training facilities

  Strengths

  When closed-source frontier models continue expanding training clusters, demand for gigawatt-scale reliable power and on-site energy solutions will be strengthened.

  Weaknesses

  Project cycles are long, and demand realization depends on data center construction progress, permitting and fuel infrastructure.

  Comparison

  The report mainly lists these names as beneficiaries in the scenario where closed-source models win, while also believing that power demand is supported across multiple states.

  Risks

  Data center project delays, energy price fluctuations, regulatory approvals, financing costs and computing efficiency improving more than expected.

• MiniMax, Knowledge Atlas, BABA, Tencent

  Open-weight model providers

  Strengths

  If open models approach frontier performance, they can expand adoption through broad distribution, ecosystem building and local-market customer bases.

  Weaknesses

  Foundation model prices are falling rapidly, direct monetization ability may be limited, and continuous investment in training and security governance is required.

  Comparison

  The report lists these companies as model-layer beneficiaries in the scenario where open-weight models win.

  Risks

  Potential U.S. restrictions on Chinese models, cross-border business continuity, licensing policies, security concerns and gaps from frontier performance.

Key data

• Enterprise adoption rate of open models63%Results from McKinsey's survey of about 700 technology leaders across 41 countries; most adopters also use closed-source models.
• Share of OpenRouter tokens from U.S. users flowing to Chinese open modelsover 30% per weekStatistics cover February to July 2026; the report notes the sample may be more skewed toward startups than large enterprises.
• Average economics of AI tasksapproximately $55 in benefits, corresponding to $2–5 in costsUsed to illustrate that enterprise AI adoption still has substantial room for growth.
• Potential price impact of open modelsAverage prices may fall by 70%, and consumers may save about $25B per yearCites a December 2025 MIT study; not all open-weight models have lower unit token costs.
• Share of enterprises valuing deployment control40%In McKinsey's survey, surveyed enterprise leaders preferred open-weight models because they could control the hosting location.
• Payback period for on-premises deploymentapproximately 3 months to 6 yearsThe payback period for small deployments can be as short as about 3 months, while large 200B–1T parameter deployments may take as long as about 6 years.
• Potential ROIC of GOOGL model APIsapproximately 45%Applies to a scenario in which Gemini is at the frontier and operates model APIs on its own infrastructure.
• Potential ROIC of GOOGL infrastructure provider modelapproximately 30%Applies to a scenario in which Gemini is not at the frontier and GOOGL plays more of an infrastructure provider role.
• Incremental ROIC of major AI inference business modelsabove 25%The report believes the three main business models in the AI inference era have relatively attractive incremental unit economics.

Impact & implications

For investors, the key is not simply to bet on open or closed-source models, but to identify the sensitivity of each industry segment to the three scenarios. Closed-source dominance favors a small number of platforms with advantages in capital, model performance and cloud infrastructure, as well as semiconductor, networking and power supply companies supporting hyperscale training; hybrid architectures are most favorable to software vendors that can provide routing, orchestration, evaluation, governance and security across models; open-weight dominance would shift value further from foundation model pre-training toward fine-tuning, inference optimization, agents, applications, on-premises infrastructure and edge devices. Improvements in computing efficiency may expand the total volume of AI workloads, so NVDA and power-related demand have relatively stronger resilience across scenarios.

Risks

• The United States may restrict Chinese open-weight models, and China may also restrict overseas frontier models, creating risks for multinational enterprises in business continuity and supplier selection.
• Open weights can have safety alignment removed or be maliciously fine-tuned, increasing cybersecurity, model misuse and responsibility attribution risks.
• Keeping up with frontier versions requires repeated fine-tuning, and enterprises may lack computing power, talent, time and engineering resources.
• The absence of per-token fees for open-weight models does not mean total cost of ownership is lower; the payback period for large on-premises deployments may be as long as several years.
• Brand, trust, switching costs, indemnification protections and concerns about foreign models may slow enterprise migration.
• Model performance, prices and efficiency change rapidly, and the probabilities of the three scenarios and the corresponding stock beneficiary relationships may continue to adjust.
• The beneficiary companies listed in the report span multiple industries, and individual stock performance is still affected by valuation, competition, capital expenditure cycles and company execution.

What to watch

• Whether open-weight models can continue narrowing the performance gap with frontier closed-source models in reasoning, agents and security.
• Changes in the shares of open, closed-source and hybrid architectures in enterprise AI model spending, rather than only observing trials or token traffic.
• Regulatory developments in the United States and China regarding model origin, weight release, data sovereignty and security responsibility.
• The full lifecycle total cost of ownership of open models, including hosting, computing power, fine-tuning, talent, security and version upgrades.
• The pace at which enterprise workloads migrate to private cloud, on-premises and edge environments, and the actual payback period of GPU clusters.
• Procurement and monetization trends for model gateways, routing, orchestration, evaluation, observability and governance software.
• The relative speed of model efficiency improvements versus new workloads, to assess whether Jevons paradox can continue supporting computing power and electricity demand.
• Whether GOOGL's Gemini can return to the frontier, and the ROIC differences between its model API and infrastructure businesses.