Artificial Intelligence is rapidly evolving from a productivity tool into something much larger: a form of digital infrastructure.

Around the world, governments are exploring how AI can improve public services, support economic development, strengthen institutions, and accelerate progress toward national development goals. Here in Trinidad and Tobago, similar discussions are taking place as Government and UNDP collaborate on initiatives related to Digital Public Infrastructure (DPI), digital identity, interoperability, AI governance, Digital Credentials, Trust and Safety, and the country's emerging Open Source Programme Office (OSPO).

At the same time, a heated debate has emerged among Silicon Valley leaders, governments, universities, startups, and policymakers around the world.

That conversation is about Open Models and Open Source AI.

In July 2026, NVIDIA CEO Jensen Huang made his first-ever post on X (formerly Twitter) to share an open letter supporting open-weight AI models. The letter, co-signed by dozens of technology companies and organizations, argued:

"AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models."

The statement immediately sparked debate across the technology sector because it touched a fundamental question:

Should the future of AI be controlled primarily through proprietary services owned by a handful of companies, or should governments, universities, local innovators, and businesses also have the ability to inspect, modify, deploy, and operate advanced AI models themselves?

Supporters of open models see openness as a pathway to innovation, competition, and digital sovereignty. Advocates of proprietary systems argue that closed models provide stronger safety controls and more sustainable business models for the enormous investments required to develop frontier AI.

So who is right?

Before you form an opinion, it helps to understand what these terms actually mean.

An Open Source AI Crash Course

One challenge in discussing AI is that terms such as "open model," "open weight," and "open source" are frequently used interchangeably, even though they describe very different things.

Think of AI like a car.

A company could let you drive the car for a fee (proprietary). It could let you look under the hood and inspect the engine, without being able to alter any component (open-weight). Or it could give you the complete engineering plans and permission to modify the design to suit your needs (open-source).

Each represents a very different level of openness.

Most people are familiar with proprietary AI models, such as ChatGPT, Claude, Gemini, and Microsoft Copilot. These systems are owned and controlled by the companies that developed them. 

Using a proprietary model is a bit like using Netflix.

You can enjoy the service, but you do not own the platform, control its pricing, or decide how it operates.

You access the technology through an application or API (to be explained shortly), and the provider determines what happens behind the scenes.

Users benefit from powerful capabilities, but have limited visibility into how the models were built and little control over how they evolve.

At the other end of the spectrum is Open-Source AI. The Open Source Initiative (OSI), an organization that has long defined what constitutes "open source" software, recently published an Open-Source AI Definition  that emphasizes users' ability to study, modify, use, and share AI systems. To achieve this level of openness, users typically need access not only to the model itself but also to information about how it was created, trained, and maintained.

Between these two extremes sits a category that has become increasingly important: Open Weight Models.

These models publicly release the "weights" that contain the model's learned knowledge. In practice, this means that organizations can often download the model, run it themselves, customize it, or adapt it for specific use cases. Models such as Llama, DeepSeek, Qwen, Mistral, Gemma, and Kimi fall into this broader category.

A useful analogy is baking a cake.

A proprietary model lets you buy the finished cake.

An open-weight model lets you buy the cake and inspect many of the ingredients.

A truly open-source AI system provides the recipe, instructions, and permission to bake your own version.

It’s best to visualize the “openness” of AI models on a spectrum, with one end being fully closed systems and the other end being fully open. All AI models fall somewhere along this spectrum, according to the amount of transparency and downloadability.

Understanding Tokens: The Hidden Economics of AI

Before we go further, it is worth understanding one more important concept: tokens.

AI models think in tokens.

AI systems do not process information the way humans read words and sentences. Instead, they break text into smaller units called tokens.

Every question you ask an AI model consumes tokens. Every response generated by the model consumes tokens.

Most commercial AI providers charge based on this usage.

A useful analogy is a taxi meter.

You do not pay for owning the taxi. You pay for how far you travel.

Similarly, many AI services charge according to how much of the model you use.

This detail may seem technical, but it has enormous implications for governments, businesses, and public institutions seeking to deploy AI at scale.

 

How Organizations Typically Adopt AI
 

Most people begin their AI journey through a public platform such as ChatGPT, Gemini, Claude, DeepSeek Chat, Kimi Chat or another of the many public chatbot interfaces.

This is the simplest option. Users open a website or mobile application and immediately gain access to sophisticated AI capabilities. The provider manages the infrastructure, security, maintenance, and upgrades.

Think of this like renting a car. The provider owns the vehicle, maintains it, fuels it, and upgrades it. You simply use it when needed.

As organizations become more sophisticated, they often move beyond public chat interfaces and begin integrating AI directly into their own systems through APIs.

If you have ever interacted with a customer service chatbot, a document search tool, or a smart online application form, there is a good chance an API was involved behind the scenes.

At this stage, AI is no longer simply a tool for individual productivity. It becomes embedded within organizational workflows.

Over time, new questions begin to emerge.

What happens if token costs continue to grow?

How much control do we have over our data?

Can we customize the model for our specific needs?

These questions often lead organizations toward a third option: downloading and hosting an open-weight model themselves.

Instead of sending requests to an external provider and paying per interaction, the organization operates the model directly on infrastructure it controls.

This progression, from public platforms to APIs to self-hosted models, mirrors the journey many governments and institutions are beginning to explore today.

Why Governments Care About Open Models

For most individuals, the distinction between proprietary and open models may not matter very much.

If an AI tool helps write an email or summarize a document, the type of model running behind the scenes may feel irrelevant.

Governments, however, face a different reality.

Public institutions manage:

• Health information
• Tax records
• Social protection data
• Education systems
• Digital identity platforms
• National registers and administrative databases

As AI becomes integrated into these systems, questions about control, resilience, sovereignty, and long-term sustainability become increasingly important.

This is where open models begin to attract attention.

The Token Cost Question

One of the most overlooked aspects of AI adoption is the long-term cost associated with token-based pricing.

Initially, proprietary AI services can appear highly attractive. Organizations can begin using advanced AI capabilities without purchasing servers, hiring specialized staff, or operating technical infrastructure.

However, success can create its own challenge.

The more an organization uses AI, the more tokens it consumes.

Imagine a government eventually deploying AI to support thousands of public officers, citizen service portals, hundreds of schools, social programmes, and national contact centres.

Under a token-based model, costs generally grow alongside adoption. The more the tool is used, the more expensive the AI becomes to run.

Open models change this equation.

Instead of paying primarily for usage, organizations invest in infrastructure, expertise, and operations. While this often involves higher upfront investment, it can create more predictable costs over time, particularly for large-scale public sector deployments.

Interestingly, even when accessed through APIs with consumption of tokens, some open models often cost less than proprietary alternatives.

The reason is relatively simple: competition.

A proprietary model can generally only be hosted by its creator. An open-weight model can potentially be hosted by universities, startups, cloud providers, governments, and technology companies around the world. More providers means more competition, and competition often puts downward pressure on prices.

This does not make open models "free", but it does create additional choices.

Organizations are no longer limited to a single provider.

Why This Matters for Small Island Developing States

For countries such as Trinidad and Tobago and other Small Island Developing States (SIDS), the discussion becomes especially interesting.

When people hear "open-source AI," they sometimes assume governments must build massive data centres or train their own equivalent of ChatGPT.

That is rarely the objective.

Training a frontier AI model can cost hundreds of millions or even billions of dollars and requires infrastructure that few countries possess.

Most countries do not need the most advanced frontier models. The more practical question is:

How can countries leverage existing AI models in ways that strengthen national capability, protect sensitive information, develop local expertise, and avoid creating new forms of technology dependence?

Open-weight models create possibilities that did not previously exist.

Instead of sending sensitive information to external systems, organizations can bring the model to the data.

Instead of relying entirely on foreign providers, they can build local expertise around deployment, customization, governance, and evaluation.

For governments concerned about digital sovereignty, these distinctions can be significant.

Open Models Are Not a Silver Bullet

It is important to be realistic.

Open models solve some problems, but they introduce others.

Organizations that choose to operate their own models assume responsibility for infrastructure, cybersecurity, maintenance, upgrades, governance, and performance management.

In many cases, the bigger challenge is not technology but capacity.

Running AI systems requires engineers, data specialists, cybersecurity professionals, and governance frameworks. While open models can reduce certain forms of dependence, they also create new responsibilities.

There are also important questions about trust and safety.

Whether a model is open or proprietary, governments must consider issues such as bias, misinformation, privacy, accountability, and responsible use.

For this reason, discussions about AI adoption are increasingly occurring alongside discussions about AI governance and Trust and Safety frameworks.

The question is no longer simply whether to deploy AI.

The question is how to deploy AI responsibly.

Why Open-Source Programme Offices Matter

This is one reason governments around the world have begun establishing Open Source Programme Offices, or OSPOs.

An OSPO provides a focal point for policy, governance, coordination, and capacity building related to open technologies.

Rather than advocating for any particular technology, an OSPO helps organizations assess where open approaches create value and where proprietary approaches may remain preferable.

For Trinidad and Tobago, this work is closely connected to broader conversations around Digital Public Infrastructure, interoperability, digital identity, open standards, local innovation ecosystems, and long-term digital resilience.

The Future Is Probably Hybrid

The debate between proprietary AI and open AI is often presented as a choice between competing visions of the future.

In reality, the most practical path forward may involve both.

Proprietary models can provide tremendous value for general productivity, research assistance, document drafting, and other low-risk use cases.

Open models may be particularly attractive for citizen-facing services, government knowledge platforms, sensitive data environments, and situations where customization, sovereignty, or cost predictability are especially important.

The future is therefore unlikely to be fully open or fully proprietary.

It is likely to be hybrid.

Final Thoughts

For UNDP, this discussion is not ultimately about technology.

It is about development.

The real question is how countries can use AI to improve public services, expand opportunity, strengthen institutions, and accelerate progress toward national development goals.

Open models offer compelling opportunities to reduce vendor lock-in, build local capacity, support digital sovereignty, encourage innovation, and create more sustainable pathways for AI adoption. At the same time, they require thoughtful governance, investment in skills, strong Trust and Safety frameworks, and realistic assessments of capacity and risk.

Open models will not replace proprietary models, just as proprietary models will not eliminate the value of openness.

The future is likely to involve both.

Success will depend less on choosing sides and more on building the governance, partnerships, institutional capacity, and policy frameworks needed to use both responsibly.

As Trinidad and Tobago continues to advance its digital transformation agenda, conversations about AI governance, Digital Public Infrastructure, digital sovereignty, Trust and Safety, and Open Source AI will become increasingly important.

UNDP is proud to support these discussions and help ensure that the benefits of AI are harnessed in ways that are inclusive, practical, and aligned with sustainable development.