Delangue warns tech cartels will choke AI growth without open weights

A massive transition is reshaping the technology sector. For years, the narrative around artificial intelligence centered on massive, proprietary frontier systems. Startups and enterprise giants alike rushed to rent access to these digital brains, celebrating the convenience of a simple API call. But that honeymoon period is ending. As companies move beyond initial proof-of-concept experiments and begin running real workloads at scale, they encounter a harsh reality: proprietary APIs are financially unsustainable.

Clem Delangue, the visionary co-founder and CEO of Hugging Face, observes this pattern daily. Teams begin their building journey by prototyping with closed models. However, once traffic spikes, the monthly API bills force a sharp pivot. Companies are migrating toward open weights. They realize they cannot outsource their core technological capabilities to a third-party black box over which they have zero visibility, zero control, and zero ownership.

The Decentralized Dev Loop

The sheer volume of developer activity proves this movement is not a passing trend. On Hugging Face, developers register a new repository every seven seconds. The platform hosts nearly three million public models and one million datasets, powering workloads across half of the Fortune 500. This is the industrialization of software 2.0. Rather than relying on a single, all-powerful model to dictate every application, builders deploy specialized, highly customized systems designed for specific business actions.

This shift mimics the historical trajectory of traditional software development. The internet was not built on proprietary, closed-source black boxes; it was constructed on open foundations like Linux and Apache. By running open models on their own hardware or cloud instances, software engineers regain complete control over latency, compute costs, and proprietary data privacy. This structural freedom serves as the bedrock of real enterprise differentiation.

The Geopolitical Shift in AI Sovereignty

While American companies argue over theoretical safety protocols and push for defensive regulations, other global actors are moving ahead. According to Hugging Face metrics, Chinese models have emerged as dominant players in global downloads, capturing a staggering 41% of the platform's volume.

The Rise of Chinese Open Weights

High-performing Chinese systems like GLM 5.2 are attracting intense developer attention for their advanced agentic capabilities. Leading American technology platforms and prominent scale-ups are already running their systems on top of open weights developed in China. Prominent tech leaders, such as Brian Chesky of Airbnb, have publically highlighted the critical importance of these open frameworks.

The Research Bottleneck

This trend is deeply felt in academic institutions. Researchers at Harvard University and Stanford University rely heavily on open-weights systems to study AI mechanics. Because proprietary APIs do not expose model weights, they are impossible to audit, study, or systematically improve. If Western companies and institutions continue to depend on foreign open weights due to domestic restrictions, the center of gravity for AI research will inevitably shift away from Silicon Valley.

The Asymmetry of Closed-Door Safety

The dominant argument against open-source AI is safety. Proponents of proprietary networks argue that open weights allow bad actors to bypass guardrails and deploy systems for malicious purposes. Delangue rejects this premise. He asserts that keeping technology behind closed doors does not make it safe; it simply creates an asymmetry of capabilities.

Guardrails on commercial APIs are superficial and easily broken. Real security is achieved through transparency, which enables defenders to patch vulnerabilities, study structural flaws, and build robust defenses. Furthermore, the true danger facing the sector is not the open distribution of code, but the intense concentration of power. Allowing a tiny cartel of tech giants to control the foundational technology of our era represents a far greater threat to economic stability and human agency than open distribution.

Building for the Long Horizon

Hugging Face maintains a unique position in this fast-moving market. While other AI startups burn through billions of dollars in venture capital to rent massive compute clusters, Hugging Face has raised $400 million and prioritized extreme capital efficiency. The company remains close to profitability and only recently began utilizing funds raised three years ago. By avoiding the typical Silicon Valley pressure to maximize short-term revenue, the platform has cultivated a deep, long-term relationship with its 17 million builders.

This long-term focus allows Hugging Face to look beyond current market hype. While investors crowd around text-based LLMs, massive opportunities remain underfunded in specialized domains:

• Local AI: Running powerful, optimized systems directly on smartphones and local hardware rather than relying on expensive cloud systems.
• Physical Sciences: Applying transformer architectures to deep biology, chemistry, and drug discovery.
• Robotics: Merging open-source models with physical machines, where open code is essential to building user trust and protecting consumer privacy.

As AI continues to mature, the companies that thrive will not be those renting expensive, closed APIs. The future belongs to the builders who own their systems, optimize their weights, and control their digital destiny.