Hugging Face's CEO on why companies are done renting their AI | Equity Podcast

The Rise of Open-Source AI and Enterprise Adoption

• Open-source AI is experiencing significant growth, with companies initially using frontier APIs for experimentation, but switching to open-source models as they scale and face increasing costs 10s.
• The volume of models and datasets being shared on platforms like Hugging Face is substantial, with a new repository created every 7 seconds, resulting in over 3 million public models and 1 million public datasets 42s.
• Enterprises, including half of the Fortune 500, are using Hugging Face, with many deploying open models in production, rather than just experimenting with them 2m6s.
• The typical flow for companies is to start with frontier APIs for experimentation, then switch to open-source models as they hit production and scale, due to the high cost of frontier models 4m30s.
• This trend is expected to continue, with frontier models potentially being used for high-value tasks and experimentation, while open-source models power most production workloads 6m15s.

Implications for AI Companies and Market Dynamics

• The shift to open-source models may impact companies like OpenAI and Anthropic, but they can still build valuable companies by focusing on frontier-level tasks and reasoning 8m30s.
• Ultimately, the AI industry is expected to be large enough for multiple companies, including OpenAI and Anthropic, to coexist and thrive, with open-source models powering most workloads 10m10s.

Control, Transparency, and Sustainability in AI Deployment

• Companies are expected to be highly valuable by the end of this year or next year, and they will likely be fine, with the current perfect storm of factors bringing open source back into public attention, including the token issue and the conversation around Chinese models catching up 10s.
• The point that resonates the most is that companies want more control and transparency in the systems they use, which is something that has been said for a while, and it makes sense for AI companies or technology companies not to outsource their core capabilities to another company 2m6s.
• Owning AI models instead of renting them provides companies with more control, transparency, and ownership, and reduces the risk of being left in a difficult situation if the government decides to shut down a particular model for safety reasons 4m30s.
• This approach to AI is more sustainable and similar to how software has always been dealt with, where companies can write their own code and build their own software stack instead of delegating it to other companies 6m15s.

AI Talent Growth and Platform Flexibility

• There is enough talent to implement AI models into businesses, as seen with the growth of Hugging Face's user base from a few hundred thousand to 16-17 million AI builders, and software engineers are now able to train, run, and optimize models themselves with the help of agents 8m40s.
• Startups can use Hugging Face to work on their own models by starting with an open-source based model, and the platform is modular, allowing users to choose how they want to use it based on their skills, team structure, and goals 12m20s.

Infrastructure Transition and Model Optimization

• Companies are transitioning from renting AI to deploying models directly on their own infrastructure, starting with off-the-shelf solutions and progressively optimizing and post-training these models for their specific use cases, which creates differentiation from other companies 10s.
• The process of deploying AI models involves optimizing them for specific use cases, particularly when there are compute, cost, or speed constraints, and post-training these models to improve their accuracy 42s.

Global AI Landscape and the Role of Chinese Models

• Chinese models, such as GLM 5.2, are gaining attention for their capabilities, and according to Hugging Face's Spring 2026 report, Chinese models account for 41% of downloads, surpassing US models in monthly and overall downloads 2m6s.
• There is a concern that US companies are not sharing enough open-source models, and instead, Chinese companies are leading in this area, with organizations like Nvidia and startups like RC Reflection working to change this 4m30s.
• Many American organizations are using Chinese open-source models, including companies like Cruise and Airbnb, as well as academic institutions like Stanford and Harvard, which rely on open-source models for research and study 6m40s.
• The risk of relying on Chinese open-source models is a concern, as it may impact the sovereignty of AI development in the US, and there is a need for more American organizations to share open-source models to reduce dependence on foreign models 8m50s.

Open Source as a Catalyst for AI Innovation

• Open source is considered a foundation and accelerating factor for AI, with the US currently leading in AI due to its openness in research and collaboration from 2016 to 2023, as seen in examples like Google sharing open source transformers, which led to the development of ChatGPT 42s.
• China is expected to potentially lead in AI in the near future, possibly as early as next year, due to its strong open-source presence and collaborative approach, with top-notch developers and researchers contributing to the field 1m6s.

Open Source vs. Closed Source: Myths and Realities

• The notion that China's open-source success is solely due to distillation attacks and copying from closed frontier models is deemed reductive and simplistic, as distillation is a common practice worldwide, including in the US, and China's success can be attributed to its talented research teams 2m6s.
• Open models are perceived as riskier due to being harder to control, with concerns about cybersecurity attacks, but historically, open source has been less dangerous than closed source initiatives because of its transparency, allowing for easier understanding of capabilities and creation of mitigations 4m30s.

Cybersecurity and the Risks of Closed AI Models

• The presence of guardrails or APIs is considered insufficient to prevent cybersecurity risks, as they can be easily bypassed, and the capabilities of models can be replicated by other labs, making it unlikely that keeping them behind closed doors will ensure safety 6m40s.
• The argument is that restricting access to AI models does not necessarily make them safer, as the risks already exist and can be replicated by other labs, and instead, transparency and open collaboration may be more effective in addressing cybersecurity concerns 8m50s.

Power Asymmetry and the Need for Transparency

• Creating asymmetry of power and capabilities between actors who have access to AI models and those who do not can make the world more dangerous, and the way to make the world safer is by leveling the playing field and creating transparency on these models, as well as making it harder and illegal to use them for malicious purposes 10s.
• Enabling innovation, competition, and job creation, while avoiding the creation of monopolies, is crucial, and this can be achieved by promoting transparency and open source development in the AI sector 2m6s.
• The concentration of power in the hands of a few AI companies is a significant risk, as these companies are becoming the most valuable and powerful in the world, and if left unchecked, they could dominate the AI landscape and accumulate unprecedented amounts of power and wealth 2m6s.
• To mitigate this risk, it is essential to support open source development and create a regulatory environment that allows multiple companies to build and contribute to AI, rather than enabling a few companies to dominate the field 2m6s.

Public Support and Legal Challenges in Open Source AI

• Public support for open source development is vital, and the US government should continue to show support for open source and contribute to open source AI, as some American public organizations have already done, such as the National Design Agency, which recently released an open source model for PII detection on Hugging Face 6m42s.
• The importance of open source and open science AI is highlighted, and there is a need for a rebrand or makeover to foster its growth, with the awareness that open data sets and sharing can pose risks, such as copyright issues, as seen in the recent lawsuit against Hugging Face, Stability, and Runway by Evox Productions, claiming that Hugging Face hosted data sets with copyrighted images 10s.

Regulatory and Ethical Considerations in Open Source AI

• Hugging Face follows all regulations and rules as a platform hosting other people's models and data sets, and has introduced initiatives to provide legal clarity, such as a new type of license for open source models, to address the challenge of balancing private and public use cases 2m6s.
• The difference between open source and closed source labs is discussed, with the point being made that open source provides more attribution and allows people to know what is used and what is not, and that fair use has been designed to balance giving attribution and protecting creation with allowing innovation and education 4m30s.
• The concept of fair use is explained, citing the example of teachers being able to use copyrighted material for public good, and how this applies to open source, where data sets are shared for free and not for profit, differing from labs using data to make billions of dollars in revenue without public contributions 6m15s.

Hugging Face's Business Strategy and Funding Approach

• Hugging Face's approach to fundraising is discussed, having raised $400 million but not having done a round in 3 years, and turning down a huge investment from Nvidia last year, prompting questions about their strategy in the current environment 10m30s.
• Hugging Face has taken a unique approach to its business, focusing on building a platform for the community and prioritizing long-term sustainability over short-term revenue, which is reflected in its capital-efficient operations and close-to-profitability status 10s.
• The company's goal is to become the storage and collaboration platform for AI builders, creating strong network effects and capturing a small percentage of the value created on the platform, which requires a long-term approach 2m6s.
• Hugging Face's position in the field is considered strategic and interesting, as it is not in a highly competitive position, but rather in a unique position where it can create value for the community and AI builders 4m42s.

Underinvested Areas in AI and Hugging Face's Exploration

• There is a big disparity in investment in the AI field, with some areas being underinvested, such as local AI, which involves running AI on devices rather than in the cloud, and biology and chemistry, which have seen little investment compared to text LLM APIs 6m38s.
• Other underinvested areas include robotics, with Hugging Face having its own robotics project called Reachy, which is an example of the company's efforts to explore different domains and create value in areas that are not as heavily invested in 9m14s.

Open Source in Robotics and Trust Issues

• Robotics has a big advantage when it comes to open source because no one company can collect all the physical data, and data sets for robotics are huge and more difficult to work on than text data sets, requiring petabytes of storage 42s.
• The trust issue with robotics is a significant concern, as having a robot in the home that interacts with the environment, family, and privacy can be scary, especially if the system is a black box controlled by a few organizations 2m6s.
• Open source is necessary for robotics and AI in general to provide transparency, allow for competition, and give people choices, rather than having only one or two options, which can be a scary world where individuals give up their agency and ability to decide 2m6s.
• Having open source and competition in AI and robotics can empower hundreds, thousands, or tens of thousands of companies to build different things and provide people with choices, rather than being limited to only a few options 2m6s.

Conclusion and Call to Action for Listeners

• Listeners can connect with Clem on Twitter or LinkedIn to follow his work and reach out to him, and they can also find the Equity podcast on Twitter and LinkedIn, as well as on YouTube and other podcast platforms 9m35s.