In an artificial intelligence (AI)-driven economy, the key question is how firms build systems that are safe, compliant, scalable and, above all, cost-effective.
Recently, US-based AI software firm Palantir’s chief executive, Alex Karp, criticised the token-based pricing model used by OpenAI and Anthropic, arguing that enterprises are being pushed towards expensive AI usage. His comments reflect a growing concern that the cost of using frontier AI models can rise rapidly as usage scales, particularly for organisations deploying AI across large teams and high-volume workflows.
The divide is reshaping enterprise AI adoption, software development and future technology spending. It is becoming increasingly significant as the performance gap between leading AI models narrows.
As Stanford University’s AI Index Report 2026 notes: “At the technical frontier, leading models are now nearly indistinguishable from one another. Open-weight models are more competitive than ever.”
What separates proprietary AI from open-weight AI?
The debate between proprietary and open-weight AI is both philosophical and commercial. It centres on who controls the underlying technology and how widely it should be made available.
Proprietary, or closed, AI refers to models whose internal workings, model weights and training processes remain under the control of the developer. Users typically interact with these systems through APIs, chatbots or enterprise software, while the provider manages the infrastructure, updates and feature releases.
Companies such as Anthropic follow this model, while OpenAI and Google are primarily associated with proprietary systems, even though both have also released some open-weight models. In many cases, usage is billed on a token basis, meaning customers pay according to the amount of text processed or generated.
Open-weight AI takes a different approach. Developers publicly release the trained model weights — the numerical parameters that determine how an AI model generates responses — allowing organisations to download and run the models independently. Although these models are often described as “open source”, the term is frequently used loosely.
Many prominent models, including those from DeepSeek and Alibaba, provide access to model weights without releasing the complete training data, source code or datasets required to recreate the model from scratch. As a result, they are more accurately described as open-weight rather than fully open-source AI.
How do the two approaches work?
The differences between proprietary and open-weight AI become clearer when viewed through how organisations deploy them.
With proprietary AI, the model typically runs on infrastructure operated by the provider. Customers send requests through APIs or web-based interfaces, with computation performed in remote data centres. The provider controls software updates, performance improvements, security patches and access to new capabilities. Customers generally have limited ability to modify the underlying model beyond configuration options or prompt engineering.
Open-weight AI allows organisations to download the trained model weights and deploy them on their own servers, private cloud environments or third-party cloud infrastructure. Because the model weights are available, organisations can operate the model independently.
This also gives organisations greater flexibility. They can fine-tune models using their own data, optimise them for specific industries or workflows, decide when to install updates, and integrate them more deeply into existing software environments. Instead of paying continuously for every AI request, the primary cost shifts to computing infrastructure, including graphics processing units (GPUs), storage and system management.
The result is a different operating model. Proprietary AI functions primarily as a cloud-based service, while open-weight AI increasingly resembles enterprise software that organisations deploy and manage themselves.
How do the business models differ?
The technical differences underpin two contrasting approaches to monetisation.
For providers of proprietary AI, the model itself is the commercial product. Revenue is generated through API consumption, premium chatbot subscriptions, enterprise licensing agreements and access to advanced capabilities. As customer usage increases, recurring revenue generally rises in parallel because organisations continue paying for every interaction with the model.
The open-weight ecosystem follows a different commercial logic. While the model itself may be released free of charge or under an open licence, companies generate revenue from the surrounding services rather than the model alone.
These include managed cloud hosting, enterprise deployment services, technical support, model fine-tuning, security features and software tools that simplify implementation. Cloud providers also benefit from increased demand for computing resources, while hardware companies supplying GPUs and AI accelerators stand to gain as more organisations choose to operate models themselves.
In this ecosystem, value shifts away from ownership of the model towards the infrastructure, software and services that enable organisations to deploy AI at scale. The model becomes a foundation on which businesses build commercial offerings rather than the sole product being sold.
Why are enterprises paying attention to open-weight AI?
If proprietary models are perceived to have an advantage, the competitive landscape is narrowing, according to Stanford University’s AI Index Report 2026.
The report found that the US-China AI model performance gap has effectively closed, with US and Chinese models trading the lead multiple times since early 2025. In February 2025, DeepSeek-R1 briefly matched the top US model, and by March 2026, Anthropic’s leading model held a margin of just 2.7 per cent. The report also noted that while the US still produces more top-tier AI models and higher-impact patents, China leads in publication volume, citations, patent output and industrial robot installations.
The emergence of increasingly capable open-weight models is prompting organisations to reassess how they deploy AI.
One of the principal attractions is long-term cost efficiency. For organisations processing millions of AI requests, operating a model on dedicated infrastructure can become more economical than paying recurring API charges. The economics depend on workload size and infrastructure utilisation, but the potential savings become increasingly significant at scale.
Data governance is another consideration. Running AI models internally also gives organisations greater control over proprietary information, an important factor for sectors such as finance, healthcare and government, where regulatory requirements and confidentiality obligations are stringent.
Open-weight models also offer greater opportunities for customisation. Businesses can fine-tune models for industry-specific terminology, internal knowledge bases and specialised workflows that may not be adequately supported by general-purpose proprietary systems. Because organisations control deployment, these models can often be integrated more closely with existing enterprise software and operational processes.
At the same time, adopting open-weight AI introduces new responsibilities. Organisations must manage infrastructure, cybersecurity, software updates and model maintenance. They also require engineering expertise to deploy, optimise and monitor these systems effectively. For many businesses, the additional operational complexity may offset some of the financial benefits.
Even so, proprietary AI retains important advantages. Providers generally deliver the latest capabilities first, manage infrastructure and security, offer integrated developer tools and enterprise support, and continuously improve their models without requiring customers to maintain them. For organisations seeking a plug-and-play solution, these advantages can outweigh the higher operating costs.
Why does proprietary AI still lead?
Despite growing interest in open-weight models, proprietary AI continues to dominate many of the industry’s most demanding applications.
The report also cautions that measuring progress is becoming more difficult. Many established AI benchmarks are approaching saturation, frontier developers are publishing fewer technical details, and independent testing does not always corroborate performance claims made by model developers. As a result, competitive advantage is increasingly determined not only by benchmark scores, but also by deployment options, ecosystem maturity, developer tools and total cost of ownership.
What could this battle mean for AI’s future?
The competition between proprietary and open-weight AI is unlikely to produce a single winner. Instead, the industry may evolve much as previous computing platforms have, with multiple business models coexisting.
Comparable patterns have emerged before. Microsoft’s Windows and Linux developed distinct ecosystems for different user groups. Apple’s iOS and Google’s Android adopted contrasting approaches to software control and openness. Cloud computing also evolved into a mix of fully managed services and self-managed infrastructure, allowing organisations to choose the model that best suited their requirements.
Artificial intelligence appears to be following a similar trajectory. Consumers may continue to favour proprietary AI assistants that offer ease of use, integrated experiences and continual feature updates. Enterprises, meanwhile, are likely to adopt a broader mix of proprietary and open-weight models depending on their requirements for cost, customisation, security and regulatory compliance.
Ultimately, the industry’s defining competition may not be over which company builds the most capable AI model. It may instead be over which approach to ownership, deployment and monetisation becomes the dominant foundation for the next generation of computing. Rather than replacing one another, proprietary and open-weight AI are likely to coexist, serving different needs across an increasingly diverse AI ecosystem.
