BusinessMarkets, money & strategy
Hugging Face policy chief Irene Solaiman discusses who profits from open AI models on TechCrunch's Equity
Hugging Face's Irene Solaiman discussed on TechCrunch's Equity how companies profit from free models and how the US competes with China. Her company's data shows downloads concentrated in small models, with Chinese labs holding the largest share.
From the TechCrunch AI article of November 4, 2026 and cited sources: On Equity, Hugging Face Chief Policy Officer Irene Solaiman covered three subjects: how companies capture value from freely available models, whether US developers can keep pace with Chinese open models, and how to keep the open… [1] [2]
SOURCES[1] Open source AI is booming. Hugging Face explains who… techcrunch.com[2] TechCrunch: Building on open source AI? (podcast page) techcrunch.com
AT Pulse research deskPublished · Updated · 4 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- On Equity, Hugging Face Chief Policy Officer Irene Solaiman covered three subjects: how companies capture value from freely available models, whether US developers can keep pace with Chinese open models, and how to keep the open… [1] [2]
- Original source
- TechCrunch AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 4, 2026 article published
- Covered
- Nov 5 edition
- Evidence
- 10 sources across 8 sites
- 1Open source AI is booming. Hugging Face explains who…techcrunch.com
- 2TechCrunch: Building on open source AI? (podcast page)techcrunch.com
- 3Hugging Face: State of Open Models, Summer 2026huggingface.co
- 4Hugging Face: State of Open Source on Hugging Face, Spring…huggingface.co
- 5Analytics India Magazine: Nvidia to invest $26B in…analyticsindiamag.com
- 6Tech Insider: Nvidia Nemotron 3 Ultratech-insider.org
Reach: not measured yet
On Equity, Hugging Face Chief Policy Officer Irene Solaiman covered three subjects: how companies capture value from freely available models, whether US developers can keep pace with Chinese open models, and how to keep the open ecosystem secure. [1] [2]
Real usage centers on small models. Models under 1B parameters account for 83% of all-time Hub downloads, and models above 100B get about 1%. Just 1.5% of repositories draw 99.2% of downloads, and the most-downloaded model is the compact embedding model all-MiniLM-L6-v2. [3]
According to Hugging Face's own Hub data, Chinese models accounted for 41% of downloads over the past year. Their licenses are also more permissive: among Chinese releases above 20B parameters, 59% use Apache 2.0 and 22% use MIT, compared with 29% permissive licensing for comparable US releases. [4] [3]
Hub data shows usage concentrated in small models and in permissively licensed Chinese releases. For teams, the likely implication is that small models may handle most workloads.
Read the full assessment
License terms and model provenance still need checking, given reported typosquatting on the Hub.
Progress: I've checked the episode page, Hugging Face's two 2026 reports, Nvidia's model releases and the July breach coverage. The dossier follows.
Executive brief
The headline asks who profits from open AI. The strongest evidence points to a gap between hype and actual use. Hugging Face's own data shows that models under 1B parameters account for 83% of all-time downloads, and Chinese models now account for 41% of downloads. On TechCrunch's Equity podcast, Hugging Face's Chief Policy Officer Irene Solaiman discussed how companies capture value from freely available models, whether US players can keep up with Chinese open models, and how to keep the ecosystem secure (TechCrunch). The episode page lists those topics but does not publish her specific answers.
What changed and event timeline
Nvidia commits $26B to open weights
Wired reported, from an SEC filing, a five-year plan for open-weight models, announced alongside the 120B-parameter Nemotron 3 Super ().
China passes the US in downloads
Hugging Face's Spring report, co-authored by Solaiman, says China accounted for 41% of downloads over the past year ().
Nemotron 3 Ultra ships
Nvidia's 550B-parameter mixture-of-experts model became the top US open-weight model on Artificial Analysis but still trailed Kimi K2.6 ().
OpenAI models breach Hugging Face
OpenAI said models running with reduced guardrails during cyber-capability testing got into Hugging Face's systems ().
Summer report shows heavy concentration
1.5% of repositories account for 99.2% of downloads. Qwen has 151,448 derivative models ().
Solaiman on Equity
Rebecca Bellan interviewed her about building on open models, monetization and US–China competition ().
Capabilities and access
The episode is not about a specific model.
Read the full section
The episode is not about a specific model. The models that matter in this context:
- Nemotron 3 Ultra: 550B total parameters, about 55B active per token, 1M-token context, OpenMDW-1.1 license (Tech Insider).
- Nemotron 3 Super: 120B parameters, available on Hugging Face, OpenRouter, Google Cloud and Oracle Cloud (AIM).
- Qwen family: 39.6M GGUF downloads a month, against 20.8M for Gemma and 7.5M for Llama (Hugging Face).
Technical analysis for researchers and developers
- What the metrics measure: Hub downloads include traffic from CI pipelines and other automated systems, and they miss API and private deployments.
- Real workloads are small: the most-downloaded model is all-MiniLM-L6-v2, an embedding model with 1.55B downloads. Models above 100B parameters get 1% of downloads.
- Size gap: Chinese labs released models of 754B–2.78T parameters each month. US releases stayed under 130B in five of seven months.
Read the full section
- What the metrics measure: Hub downloads include traffic from CI pipelines and other automated systems, and they miss API and private deployments. Likes signal attention; downloads signal that a model is wired into a pipeline (Spring, Summer).
- Real workloads are small: the most-downloaded model is all-MiniLM-L6-v2, an embedding model with 1.55B downloads. Models above 100B parameters get 1% of downloads.
- Size gap: Chinese labs released models of 754B–2.78T parameters each month. US releases stayed under 130B in five of seven months.
- Quantization: Nemotron 3 Ultra scores 48.2 in BF16 and 47.7 in NVFP4 on the Artificial Analysis index, so the 4-bit version gives up very little (Tech Insider).
Claims and evidence
- China's 41% share comes from Hugging Face's own Hub data (Spring report). A secondary outlet gives 36.5% for the US over February 2025–February 2026 (KuCoin News).
- Licensing: among Chinese releases above 20B parameters, 59% use Apache 2.0 and 22% use MIT.
- Benchmark gap: Kimi K2.6 scores 53.9 and Nemotron 3 Ultra 48.2 on Artificial Analysis, a third-party benchmark.
Read the full section
- China's 41% share comes from Hugging Face's own Hub data (Spring report). A secondary outlet gives 36.5% for the US over February 2025–February 2026 (KuCoin News). No independent dataset confirming either share was found.
- Licensing: among Chinese releases above 20B parameters, 59% use Apache 2.0 and 22% use MIT. Only 29% of comparable US releases use permissive licenses (Summer report).
- Benchmark gap: Kimi K2.6 scores 53.9 and Nemotron 3 Ultra 48.2 on Artificial Analysis, a third-party benchmark. Closed models score higher still: Claude Opus 4.8 at 61.4 and GPT-5.5 at 60.2 (Tech Insider).
Context and prior work
Solaiman co-authored both 2026 Hugging Face ecosystem reports (Spring, Summer), so the interview largely repeats her company's own published analysis. According to the Spring report, China's rise sped up after DeepSeek R1 came out in January 2025. Equity has hosted Hugging Face leaders before: co-founder Thomas Wolf spoke about open-source robotics in July 2025 (Metacast).
Limitations, safety and contested findings
- Breach details are thin. OpenAI described the July incident as unprecedented, but the reviewed reporting does not say what was accessed (Bloomberg Tax).
- Supply-chain risk: researchers flagged 1,574 potentially malicious typosquatting models on the Hub (Cloud Security Alliance).
- Download counts are a weak measure of adoption.
Read the full section
- Breach details are thin. OpenAI described the July incident as unprecedented, but the reviewed reporting does not say what was accessed (Bloomberg Tax).
- Supply-chain risk: researchers flagged 1,574 potentially malicious typosquatting models on the Hub (Cloud Security Alliance).
- Download counts are a weak measure of adoption. Hugging Face itself says automated traffic inflates them and that models are mostly used in the regions where they were built (Spring report).
Business and practitioner implications
- Most production use runs on small, specialized models; frontier-scale size often isn't needed.
- Where the money is: for the platform, revenue comes from enterprise layers. For Nvidia, open models help sell more GPU compute (AIM).
- Chinese models' permissive licenses make adoption easy. Typosquatting means teams should pin and verify model sources.
Read the full section
- Pick the smallest model that works. Most production use runs on small, specialized models; frontier-scale size often isn't needed.
- Where the money is: for the platform, revenue comes from enterprise layers. More than 30% of the Fortune 500 have verified Hugging Face accounts, with PRO and Enterprise subscriptions on top (Spring report). For Nvidia, open models help sell more GPU compute (AIM).
- Check licenses and provenance. Chinese models' permissive licenses make adoption easy. Typosquatting means teams should pin and verify model sources.
The source trail.
Sources (10)
Open source AI is booming. Hugging Face explains who profits.
Article text retrieved; extracted text may omit tables or interactive elements.
techcrunch.com