Open-weight models drive enterprise cost reduction and efficiency gains, with AI competition shifting toward a multi-model landscape

Overview

The report focuses on three developments in open-weight models over the past week: first, Chinese models improving profitability through direct price increases and commercial licensing; second, US enterprises adopting open-weight models and achieving clear returns on investment; and third, Meta returning to the open-weight ecosystem. Morgan Stanley believes the AI market is rapidly moving toward a multi-model landscape, where open models help reduce unit costs, accelerate enterprise adoption, and sustain compute demand, while the US-China model price gap will determine the competitive boundaries between open and closed-source models.

Core views

Chinese models are shifting from low-price competition toward tiered pricing and commercialization, entering a phase of larger model scale, higher barriers to entry, and improved monetization. At the same time, prices for US closed-source models are expected to continue falling, and the rising share of open models is also lowering overall token prices. Enterprises will not simply choose between open and closed-source models, but will dynamically route based on task complexity, cost, customization needs, security, and trustworthiness. Open-weight models are better suited to high-frequency, cost-sensitive, and industry-specific tasks, while frontier closed-source models may continue to dominate complex reasoning and agentic workloads.

Analysis framework

The report combines model pricing, licensing terms, enterprise earnings case studies, and industry-chain scenario analysis to assess the impact of open, closed-source, and hybrid model paths on enterprise adoption speed, token demand, cloud computing capital expenditure, and technology industry-chain beneficiaries, and uses a “three states of the world” framework to map related stocks.

Methodology notes

• Scenario analysisThree states of the world

  Assumes respectively that closed-source models win, open and closed-source models coexist in a hybrid mix, and open models win.

  This framework is used to compare the relative benefits for cloud services, model providers, semiconductors, networking, data centers, cybersecurity, enterprise software, and edge hardware under different model landscapes.

• Demand elasticityJevons paradox

  A decline in unit compute or token costs may stimulate more usage, causing total demand to rise rather than fall.

  The report argues that price reductions for open and closed-source models will increase AI usage, so a decline in unit costs should not be directly equated with a decline in demand for compute infrastructure.

• Technical architectureMulti-model routing

  Allocating different workloads to the most suitable model based on task performance, cost, and risk.

  Enterprises can use a centralized routing layer to mix open and closed-source models, reducing spending while maintaining token usage and service quality.

Asset mapping & comparison

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

• Closed-source model winner basket: GOOGL, AMZN, META, PANW, CRWD, ANET, NVDA, AVGO

  Benefiting from frontier model capability barriers, centralized cloud workloads, and high-intensity training demand.

  Strengths

  Capital, compute, data, security, and deployment capabilities create relatively high barriers, while enterprises place greater emphasis on accuracy, reliability, and liability protection.

  Weaknesses

  A model oligopoly may keep compute costs relatively high and limit enterprises’ control and customization of models.

  Comparison

  Compared with the open-model scenario, this is more tilted toward hyperscale cloud platforms, frontier labs, and centralized data centers.

  Risks

  Rapid catch-up by open models, continued price cuts for closed-source APIs, and stronger enterprise in-house capabilities may weaken concentration.

• Hybrid model basket: AMZN, GOOGL, MSFT, DDOG, PLTR, PANW, SAP, NOW, CSCO, NVDA

  Benefiting from the coexistence of open and closed-source models, as well as enterprise demand for routing, orchestration, observability, and security capabilities.

  Strengths

  Can cover complex reasoning, high-frequency tasks, and industry-specific tasks, allowing enterprises to flexibly select models based on performance and cost.

  Weaknesses

  Architecture complexity, governance costs, and cross-model data security requirements are higher.

  Comparison

  More adaptable than a single path, and also the state the report believes is increasingly aligned with actual enterprise adoption.

  Risks

  Model standardization, rapid price convergence, or the formation of an overwhelming advantage by a single platform could reduce the value of independent orchestration tools.

• Open-model winner basket: MSFT, BABA, Tencent, PLTR, DELL, HPE, NTAP, AAPL, SAP, NOW, SHOP, NVDA

  Benefiting from the popularization of open models, declining token costs, and more workloads shifting toward local or edge deployment.

  Strengths

  Supports industry data customization, model control, private deployment, and lower inference costs, helping expand enterprise adoption.

  Weaknesses

  Enterprises need to take on more responsibility for deployment, security, updates, and model governance.

  Comparison

  Compared with the closed-source model scenario, this is more favorable for local infrastructure, edge devices, channel partners, and open-model providers.

  Risks

  Tightening commercial licensing, open-model price increases, price cuts for frontier closed-source models, and regulatory restrictions may weaken the cost advantage.

Key data

• Relative API pricing of Chinese large modelsAbout 15% to 20% of comparable US productsThe price gap remains significant, but price increases by Chinese models and price cuts by US models are expected to narrow the gap.
• DeepSeek V4-Flash base priceUS$0.14 per 1 million input tokens; US$0.28 per 1 million output tokensThe report estimates that if prices double, the profit margin of a 200MW data center running this model could approach that of Kimi K3 and other frontier models.
• Commercial licensing revenue shareMoonshot can charge up to 30%Mainly aimed at customers generating more than US$20 million in revenue using the relevant models, with the impact likely falling more on hyperscale cloud providers offering model services.
• Rippling token spendingDeclined from 40% of R&D personnel costs to 15%Achieved after adopting GLM 5.2 and customized model routing, while token usage remained stable at about 600 billion per month.
• Pinterest open-model transaction costLess than 8% of the cost of comparable closed-source proprietary modelsAchieved by using open models post-trained on proprietary data and a model routing layer.

Impact & implications

A multi-model landscape is favorable for model orchestration, observability, cybersecurity, and distributed infrastructure, and may accelerate enterprise AI penetration through lower costs. A closed-source model victory would benefit hyperscale cloud platforms, frontier model labs, and centralized training infrastructure; a hybrid landscape would benefit cross-model routing, infrastructure software, and security vendors; and an open-model victory would be more favorable for local deployment, edge devices, enterprise hardware, and channel distributors. Semiconductor and on-site power suppliers have a certain basis for benefiting across multiple scenarios.

Risks

• Improved profitability of Chinese models may create a reinvestment cycle in R&D, accelerating the narrowing of the capability gap with US frontier models and changing the competitive landscape.
• Restrictions on Chinese models by the United States or other regions may cause supply disruptions, compliance risks, and business continuity risks.
• If open models significantly replace high-priced closed-source calls, while incremental usage is insufficient to offset the decline in unit prices, token revenue and some infrastructure demand may come under pressure.
• Tighter commercial licensing and revenue-sharing requirements may weaken open models’ original advantages of low cost and permissive usage.
• Enterprises need to bear additional security, governance, maintenance, liability protection, and talent costs when deploying open models.
• The scenarios discussed in the report are still evolving rapidly, and the ranking of beneficiaries across the relevant industry chain is highly uncertain.

What to watch

• Whether the US-China large-model API price gap continues to narrow, and the pace of price cuts for low-end and mid-tier US models.
• The scale of price increases and implementation of commercial licensing by Chinese model providers such as DeepSeek, Moonshot, and Alibaba.
• Enterprise adoption rates of open-weight models, model routing penetration, and verifiable cost-saving cases.
• Whether the decline in overall token prices leads to higher usage and validates Jevons paradox.
• The capabilities, adoption rates, and ecosystem-building progress of Meta Muse Spark 1.2 and other US open-weight models.
• US-China AI policy, sovereign AI requirements, and potential model bans.
• Changes in cloud capital expenditure, data center power, and semiconductor demand under the three scenarios of open, closed-source, and hybrid models.