Who Captures Value When Open-Source AI Models Go Mainstream?

The Economic Puzzle of Open AI

Enterprises across Asia and beyond are embracing open-source AI models at an accelerating pace. The appeal is straightforward: lower deployment costs, granular control over model behaviour, and the flexibility to customise without vendor lock-in. Yet as these freely available models proliferate, a harder question emerges for companies building in this space: if the core technology is open, where does sustainable economic value actually reside?

Irene Solaiman, Chief Policy Officer and SVP of Strategy at Hugging Face, recently addressed this tension in a discussion about the state of open AI. Her organisation sits at the centre of the open-model ecosystem, hosting tens of thousands of models and datasets. The platform has become infrastructure for a generation of startups, researchers, and product teams trying to move beyond proprietary APIs. But the rise of open models, particularly those from Chinese labs that match or exceed Western capabilities, complicates the competitive landscape.

At Opentechwire, we've tracked how open-source AI has evolved from a niche research preference into a credible enterprise option. The shift is no longer theoretical. Governments in Southeast Asia are piloting open models for public services. Indian AI startups are fine-tuning Llama and Mistral variants rather than paying per token to closed providers. Even large corporations are running open models on-premises to satisfy data residency and compliance requirements.

The economics, however, remain unsettled. If the model weights are free, what exactly are customers paying for, and which layer of the stack captures margin?

Where Value Pools in an Open Ecosystem

The traditional software business model hinges on controlling access to the core asset. Open-source AI inverts that logic. The model itself, often the result of hundreds of millions of dollars in compute and talent, is released without a licence fee. Value must therefore accrue elsewhere in the stack.

According to Solaiman, startups entering this space need to think carefully about where they sit. Infrastructure providers, those offering optimised inference engines, edge deployment, or fine-tuning pipelines, have carved out defensible positions. Companies that reduce latency, cut inference costs, or simplify the path from model to production are solving real pain points that enterprises will pay to avoid.

Another emerging pocket of value lies in domain-specific curation. Open models are general-purpose by design. Adapting them for healthcare, financial services, or industrial automation requires proprietary datasets, evaluation frameworks, and regulatory expertise. Startups that own these vertical layers, rather than the models themselves, can build sustainable businesses even as foundation models commoditise.

Platform plays also matter. Hugging Face itself exemplifies this: the models are open, but the infrastructure for hosting, versioning, collaboration, and deployment is a paid service. The platform becomes the economic layer, not the algorithm.

The China Factor and Competitive Dynamics

Open-source AI is not a Western monopoly. Chinese labs, including DeepSeek, Alibaba's Qwen team, and others, have released models that rival or surpass American counterparts on several benchmarks. These models are not only competitive on quality but are often optimised for efficiency, running on less expensive hardware and delivering strong performance per dollar of compute.

For American and European startups, this represents both a challenge and an opportunity. The challenge is straightforward: if a high-quality open model from China costs nothing and performs well, the competitive moat around any given model narrows further. Differentiation must come from integration, trust, compliance, or ecosystem lock-in, not from model superiority alone.

The opportunity is more subtle. Open models from any geography can be adapted, fine-tuned, and deployed locally. A Singaporean fintech can take a Chinese open model, retrain it on local transaction data, and run it in a jurisdiction-compliant data centre. The geopolitical origin of the base model matters less than the infrastructure, governance, and application layer built on top of it.

Solaiman's comments reflect a broader industry recognition: the race in open AI is not winner-takes-all. It is a race to define standards, build ecosystems, and capture the adjacent layers where sustainable economics live.

Navigating Trade-Offs for Builders

For startups deciding whether to build on open models, the decision matrix has become more complex. Closed APIs from providers like OpenAI or Anthropic offer simplicity and continuous improvement but come with per-token costs, rate limits, and dependency risk. Open models offer control and cost predictability but require infrastructure investment, model operations expertise, and ongoing maintenance.

Solaiman emphasises that the choice depends on the use case. For prototyping or low-volume applications, closed APIs remain the fastest path. For high-volume production workloads, or for applications where data cannot leave a private environment, open models become economically and operationally superior.

The trade-off extends to model selection. Larger open models, those with hundreds of billions of parameters, deliver stronger performance but require expensive inference infrastructure. Smaller, distilled models can run on edge devices or modest cloud instances but sacrifice capability. Startups must match model size to their latency, cost, and accuracy requirements, a task that remains more art than science.

Another consideration is the pace of model releases. The open ecosystem moves quickly. A model that is state-of-the-art today may be surpassed in weeks. Startups that tightly couple their product to a specific model version face technical debt. Those that abstract the model layer and design for model swapping gain resilience but add engineering complexity.

Policy and the Future of Open AI

The regulatory environment surrounding open AI remains in flux. Export controls, particularly those targeting advanced chips and AI capabilities, have already reshaped the supply chain. US restrictions on GPU exports to China have not prevented Chinese labs from training competitive models, but they have influenced architecture choices and spurred domestic chip development.

Solaiman's role at Hugging Face includes navigating these policy crosscurrents. Open models raise questions that closed systems sidestep: who is responsible when a freely available model is misused? How do you enforce safety standards without controlling distribution? Can export controls even function in a world where model weights can be downloaded by anyone, anywhere?

The answers are not yet settled. Some governments favour mandatory licensing or registration for frontier models, even open ones. Others argue that transparency and community scrutiny, the core strengths of open source, provide better safety guarantees than opacity. The debate will shape not only what models are released but where they are hosted, who can access them, and under what terms.

For startups, this uncertainty is both risk and opportunity. Those that build compliance, auditability, and governance tooling into their products may find demand from enterprises navigating regulatory complexity. Those that ignore policy risk building on foundations that could shift beneath them.

What Startups Need to Know

The open-source AI wave is not a temporary phenomenon. It reflects a structural shift in how models are developed, distributed, and monetised. But openness alone does not guarantee success. Startups must choose their layer carefully, whether infrastructure, application, data, or vertical integration.

They must also remain adaptable. The models will keep improving, the competitive set will keep expanding, and the policy landscape will keep evolving. The companies that thrive in this environment will be those that treat the model as a commodity input, not the core product, and build defensible value in the layers that matter to customers: speed, reliability, compliance, and outcomes.

The question is no longer whether open AI will succeed. It already has. The question is who will capture the value it creates, and how.