Quick Summary
Covering the latest research from top Wall Street investment banks

Open-weights models will drive AI diffusion, with different beneficiaries across three scenarios

Institution
Morgan Stanley
Date
2026-08-03
Authors
Stephen C Byrd, Sabrina E Hao, Brian Nowak, CFA, Adam Wood, Meta A Marshall, Sanjit K Singh, Erik W Woodring, Joseph Moore, Gary Yu, David Arcaro, CFA, Joe Laetsch, CFA, Robert S Kad, Ryan Lountzis
Company
-
Ticker
-
Industry
AI
Rating
-
NeutralLow confidenceThe report believes open-weights models will promote competition, AI diffusion, and enterprise adoption, but the eventual shares of open and closed models remain uncertain; accordingly, it evaluates beneficiary assets under three scenarios.
AuthorsStephen C Byrd, Sabrina E Hao, Brian Nowak, CFA, Adam Wood, Meta A Marshall, Sanjit K Singh, Erik W Woodring, Joseph Moore, Gary Yu, David Arcaro, CFA, Joe Laetsch, CFA, Robert S Kad, Ryan Lountzis
CoverageOther
Business segmentsopen-weights models、closed-weights models、cloud providers、model providers、security software、semiconductors、networking、on-prem infrastructure、edge devices、power infrastructure
Research firm divisions/subsidiariesMorgan Stanley(Other)

AI summary card

Open-weights models will drive AI diffusion, with different beneficiaries across three scenarios

Morgan Stanley believes open-weights models can lower costs, improve customization and deployment control, and expand AI adoption through the Jevons Paradox, but it remains to be seen whether enterprises ultimately favor closed, hybrid, or open models.

This report is thematic research/meeting notes and provides no individual company ratings, price targets, or expected upside.
AIopen-weights modelsclosed-weights modelsenterprise AI adoptionJevons Paradoxcloud computingsecurity softwaresemiconductors
  • 63% of surveyed enterprises already use open-weights models in their technology stacks, typically alongside closed models.
  • The main advantages of open-weights models include no per-token charges, customizability, and control over deployment location; the main drawbacks include fine-tuning costs, security risks, unclear accountability, and geopolitical risks.
  • The report proposes three states of the world: closed models win, hybrid architectures win, and open-weights models win, mapping each to beneficiary areas across cloud, models, security, semiconductors, networking, power, edge, and on-prem infrastructure.
  • NVDA and power-related demand are relatively resilient across multiple scenarios, while the degree of benefit for hyperscalers depends on closed-model capabilities, infrastructure roles, and open-model penetration.

Report interpretation

Overview

The report focuses on competition between open-weights and closed-weights models, discussing the release of Kimi K3, potential US restrictions on Chinese models, and market debate triggered by related policy letters. Morgan Stanley believes investors are concerned that smaller, more efficient open models could weaken demand for compute and AI infrastructure, but the report's core view is that open-weights models will accelerate AI diffusion through competition and lower costs, potentially broadening enterprise AI adoption.

Core views

The report believes enterprises currently use both open and closed models, with open-weights models primarily used for smaller use cases requiring speed, security, frequent calls, or domain-specific adaptation, such as coding and document parsing. Over the long term, if open models continue to catch up with frontier models, they will lower model API prices and expand demand; if closed models retain their lead, a small number of well-capitalized frontier labs, cloud providers, and training infrastructure providers will continue to benefit; if hybrid architectures become mainstream, model routing, orchestration, observability, security, and governance will become more important.

Analysis framework

The report uses a scenario-analysis framework that divides future enterprise AI model adoption into three states: closed models win, hybrid architectures win, and open-weights models win, then derives the beneficiaries across the industry chain for each. The analysis focuses on enterprise adoption rates, cost structures, performance gaps, deployment locations, data sovereignty, regulatory restrictions, software-stack complexity, and compute and power demand.

Methodology notes

  • thematic researchthree-state scenario analysis

    Projecting different industry beneficiary chains around competition between open and closed models

    Rather than forecasting a single path, the report divides the future into three scenarios—closed models win, hybrid models win, and open models win—and compares the relative beneficiaries across cloud, models, security, semiconductors, networking, on-prem infrastructure, and edge devices under each scenario.

  • economic mechanismJevons Paradox

    Efficiency improvements may expand total demand

    The report believes lower model costs and improved efficiency may not reduce demand for AI infrastructure; instead, total usage could expand as more enterprises and use cases adopt AI.

  • enterprise technology architecturemodel orchestration and observability framework

    The value of the software layer increases in multi-model environments

    When enterprises run both open and closed models, gateways, routing, orchestration, evaluation, governance, security, and observability layers become more important because enterprises need to dynamically select models based on cost, capability, latency, and compliance.

Asset mapping & comparison

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

  • GOOGL, AMZN, MSFT
    cloud service providers
    Strengths
    They can absorb AI workloads in closed or hybrid scenarios; if Gemini returns to the frontier, GOOGL could also benefit from the economics of model APIs.
    Weaknesses
    Falling open-model prices could limit excess returns from closed models; the degree of benefit for different cloud providers depends on model capabilities and customer deployment preferences.
    Comparison
    GOOGL and AMZN stand out when closed models win; AMZN, GOOGL, and MSFT all benefit in the hybrid scenario; the report emphasizes MSFT in the open-model victory scenario.
    Risks
    Changes in the competitive model landscape, price declines, and enterprise shifts toward on-prem or edge deployment.
  • NVDA, AVGO
    semiconductors and AI compute
    Strengths
    The report believes NVDA benefits across multiple scenarios; when closed models win, demand for training clusters and high-intensity compute remains strong, also benefiting AVGO.
    Weaknesses
    If smaller open models significantly improve efficiency, compute demand per task could decline.
    Comparison
    The closed-model scenario lists NVDA and AVGO; the hybrid scenario lists NVDA.
    Risks
    Improved model efficiency, lower inference prices, and a shift in compute demand from training toward inference or the edge.
  • PANW, CRWD, ZS, NTSK, OKTA, FTNT, SAIL, VRNS
    security software
    Strengths
    Distributed, on-prem, and multi-model deployments increase demand for security, governance, identity, and access controls.
    Weaknesses
    When closed models win, some security burdens may be assumed by frontier model labs.
    Comparison
    Security software is an important beneficiary across all three scenarios, with distributed workloads making it more important in open and hybrid scenarios.
    Risks
    Enterprise AI deployment progresses more slowly than expected, or security budgets fail to grow in line with the number of models.
  • DDOG, PLTR, APPN
    infrastructure software, orchestration, and observability
    Strengths
    Multi-model environments require routing, orchestration, monitoring, evaluation, and cost and quality tracking.
    Weaknesses
    If enterprises continue to adopt a single closed-model API, increases in software-stack complexity may be limited.
    Comparison
    The hybrid scenario lists DDOG, PLTR, and APPN; the open-model victory scenario lists PLTR.
    Risks
    Platform consolidation, replacement by cloud-provider-native tools, and enterprise self-built capabilities.
  • DELL, HPE, NTAP, P, HPQ, AAPL
    on-prem infrastructure and edge devices
    Strengths
    When open models win, more workloads migrate to on-prem, private cloud, sovereign cloud, and edge devices.
    Weaknesses
    If closed models retain their lead, enterprises may continue consuming models through cloud and APIs, reducing demand for local hardware.
    Comparison
    These assets primarily benefit when open-weights models win.
    Risks
    High costs of local deployment for large models, a shortage of enterprise engineering talent, and uncertain payback periods.
  • BE, INIO, SEI, WMB, LBRT
    on-site power and energy infrastructure
    Strengths
    When closed models win, training superclusters drive gigawatt-scale demand for reliable power.
    Weaknesses
    If model efficiency improves substantially and workloads become more distributed, incremental demand from centralized, hyperscale training clusters could weaken.
    Comparison
    The report specifically emphasizes on-site power suppliers in the closed-model victory scenario.
    Risks
    Slower data-center construction, grid interconnection constraints, and regulatory and capital-expenditure cycles.

Key data

  • Enterprise open-model usage rate63%A McKinsey survey of 700 technology leaders across 41 countries shows that 63% of respondents use open models, typically alongside closed models.
  • Share of US enterprise OpenRouter tokens routed to Chinese open models30%+The report states that from February through July 2026, more than 30% of OpenRouter tokens used weekly by US companies were routed to Chinese open models, although the data may be more representative of startups than large enterprises.
  • Average AI task benefit and cost$55 benefit vs. $2-5 costThe report cites its earlier estimates to illustrate that enterprise AI adoption still has substantial runway.
  • Potential price reduction from open modelsapproximately 70%The report cites an MIT study indicating that shifting from closed to open models could reduce average prices by 70% and save consumers approximately $25B annually, while noting limitations related to the study period, model differences, and total cost of ownership.
  • Payback period for open deploymentup to 3 months to up to 6 yearsA Carnegie Mellon study shows that payback periods for smaller deployments can be as short as 3 months, while large-scale deployments with 200B-1T parameters may take up to 6 years.
  • GOOGL model API ROIC scenariosapproximately 45% or approximately 30%If Gemini returns to the frontier, the report estimates that GOOGL's ROIC from running model APIs on its own infrastructure would be approximately 45%; as an infrastructure provider, it would be approximately 30%.

Impact & implications

The investment implication is that open models do not necessarily weaken the AI investment thesis; instead, lower costs and expanding use cases could generate broader enterprise AI spending. A closed-model victory would favor frontier model labs, cloud, training compute, networking, and power; a hybrid victory would favor cloud, infrastructure software, security software, and model orchestration; an open-model victory would favor on-prem infrastructure, edge devices, security, observability, routing, and application-layer deployment.

Risks

  • Potential US restrictions on Chinese open-weights models could create business continuity risks.
  • Enterprises using open models must bear fine-tuning, hosting, talent, and compute costs; open does not mean free in terms of total cost of ownership.
  • Open-weights models face risks including removal of safety alignment, malicious modification, and unclear accountability.
  • Enterprise trust, brand, security, and compliance concerns regarding foreign models could limit adoption.
  • If closed frontier models continue to lead by a wide margin, penetration of open models into core enterprise workloads could fall short of expectations.

What to watch

  • Changes in the performance gap between open-weights models such as Kimi K3, Qwen, DeepSeek, Llama, and Gemma and closed frontier models.
  • Regulatory restrictions by the US on Chinese open models and by China on overseas frontier models.
  • Changes in the actual share of open, closed, and hybrid architectures in enterprise AI spending.
  • Model API prices, per-token costs, inference efficiency, and the pace of enterprise use-case expansion.
  • The pace at which enterprises deploy AI workloads on-prem, in private clouds, sovereign clouds, and at the edge.
  • Budget growth for model orchestration, routing, observability, security, and governance software.
Zhejiang ICP No. 2022035445-5
Disclaimer: Market data, charts, indicators, research views, and other information provided on this website are intended solely for information display, research communication, and educational reference. They should not be regarded as personalized investment advice, securities recommendations, trading instructions, solicitations, or guarantees of return. While we strive to improve the reliability of our data and content, such information may still be subject to delays, errors, incompleteness, or untimely updates due to source differences, methodological limitations, system processing, or market volatility. Users should exercise independent judgment based on their own circumstances and bear all risks and responsibilities arising from the use of this website.

Settings

Sign in to view recent logins