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

Open source AI is booming, according to Hugging Face CEO Clem Delangue. The company has grown into something like a GitHub for AI in recent years, where AI builders can share and download open models and datasets, now used by roughly half the Fortune 500. Delangue has seen the same story play out again and again: companies start out on frontier APIs, but as they scale, the costs push them towards open source models.

On this episode of TechCrunch’s Equity podcast, Rebecca Bellan talked to Delangue about why the open vs closed source fight matters in the wake of Anthropic’s halted Fable release, and why he’s worried about the possibility that a handful of big companies could end up controlling everything.

TL;DR

A ~36:57 interview on TechCrunch’s Equity podcast (host Rebecca Bellan, published 10 July 2026) with Clément “Clem” Delangue, co-founder and CEO of Hugging Face — the “GitHub for AI” model/dataset-sharing platform, now used by roughly half the Fortune 500. Load-bearing claims, roughly following the episode’s nine chapters:

1. Open-source AI is booming, measured by platform volume. A new repository is created every 7 seconds on Hugging Face — almost 3 million public models and 1 million public datasets. Delangue reads this as evidence against the “one model to rule them all” narrative: the reality is companies using many specialized, customized models for specific use cases. Half the Fortune 500 now uses Hugging Face (private and/or open-source models).

2. The own-vs-rent flow: frontier APIs for experimentation, open models for production at scale. The “typical flow” is companies starting on frontier APIs to experiment and launch features, then — “when they really hit production and they hit scale” — switching to open-source models because frontier-model cost becomes too big. Delangue’s forward projection: in a few years, frontier models will be for experimenting and high-value tasks, while most production workloads run on private or open-source models. He is explicit this need not be “devastating” for OpenAI/Anthropic — they can still be “the most valuable companies in the world” on frontier reasoning even if most workloads run elsewhere.

3. “Companies need to own AI and own models instead of renting them.” The resonant driver (crediting Palantir’s Alex Karp) is control and transparency: an AI/tech company “doesn’t want to outsource your core capabilities… to another company through a black-box API that you don’t control, don’t have any visibility on, don’t have any ownership.” He frames AI as “the next generation of software” / software 2.0 — and software was always built in-house, not outsourced. Owning also removes government-shutdown risk (a model taken down “for safety reasons” leaves API renters “in the lurch”).

4. Agents lower the barrier to owning models. Hugging Face went from a few hundred thousand users three-to-four years ago to 16–17 million AI builders. With coding agents, “it’s becoming easier and easier for software engineers to run their own models, optimize their own models, train their own models” — across startups and enterprises. The platform is described as “part GitHub for AI, but also becoming a bit AWS” in services: teams start from an off-the-shelf open model (named: GLM 5.2, OpenAI’s open GPT, Nvidia Nemotron), deploy on their own infrastructure, then progressively optimize and post-train for their use case — “which creates your differentiation from other companies.”

5. China has overtaken the US in open-model downloads. Per Hugging Face’s own Spring 2026 report, Chinese models accounted for ~41% of downloads — most downloads, surpassing the US. Delangue calls this “a very big challenge”: in an ideal world more of the open source used in the US would be shared by American companies (he credits Nvidia as “the king of American open source AI” via Nemotron and datasets, plus startups like Reflection). US scale-ups (he names Cruise, Brian Chesky/Airbnb as vocal) and all academia (Stanford, Harvard) already build on Chinese open weights, “because you can’t really study an API — it’s a complete black box.”

6. Why open-source leadership matters, and the distillation rebuttal. Delangue’s causal claim: US AI leadership 2016–2023 came from open research and open source (“the T in ChatGPT came out from Google sharing an open source transformer”) — open source “creates the conditions for your AI leadership almost automatically.” If China keeps leading in open source, “I wouldn’t be surprised if… China starts to lead AI in general, next year or the year after.” He calls the “China’s only good because of distillation attacks” argument “very reductive and simplistic” — distillation is a small factor everyone (including US firms) uses; China simply has “really, really good research teams” taking a more open, collaborative approach.

7. Safety through transparency, not closed doors. Historically open source has been less dangerous than closed initiatives because it’s more transparent — capabilities are legible and defenders can patch. Guardrails on APIs are “very shallow and quite ineffective” (“very easy to jailbreak,” “possible to steal the weights”). You don’t make AI safe by keeping it “behind closed doors for just a few players” — you make it more dangerous by creating asymmetry of power between actors who can access/steal/misuse the weights and those who can’t defend themselves. Safer = “leveling up the playing field,” transparency, and making the attacks (not the tools) illegal.

8. Concentration of power is “the biggest risk in AI.” In Delangue’s view the AI companies becoming the most valuable companies are also becoming the most powerful — he cites an AI company’s leverage vis-à-vis the (renamed) US Department of War as something he’d have called “crazy a few months ago.” A world where “a few companies completely dominate AI, getting an amount of power and wealth you’ve never seen before” — “basically similar to if there were just one or two companies able to do software” — is “the real dangerous scary scenario.” Regulatory environments that let only a few build frontier AI accelerate it.

9. What the US government should do; the legal-risk posture. Public support and a “rebrand” for open source: reverse the recent counter-incentive where open source is reflexively questioned as “unsafe” (10 years ago sharing research was celebrated). Public organizations can contribute open data/models (example: a US public agency released an open PII-detection model on Hugging Face). On legal risk — Hugging Face is named (with Stability and Runway) in an ongoing Evox Productions copyright suit over hosted datasets — Delangue can’t comment on the case but argues open sharing gives more attribution and legibility than closed labs “using the whole web without any copyright,” and invokes fair-use’s education/innovation balance (nonprofit/public-good use vs. a lab “making billions without public contribution”).

10. Hugging Face’s contrarian business model. No funding round in 3 years (raised ~$400M), and it turned down a large Nvidia investment last year. Rationale: a long-term responsibility to the community trusting the platform with their data/models; capital-efficient (doesn’t need “billions of dollars of compute”), close to profitability, only recently “started to touch” the money raised 3 years ago. As a platform it aims to “create 100× more value than we would if we weren’t a platform, and capture 1–2% of it” — strong network effects, a “storage and collaboration platform for AI builders,” likened to a social network’s long-term arc.

11. Underinvested opportunities: local AI, bio/chemistry, robotics. Asked about a bubble, Delangue: “probably in an LLM-API bubble, but definitely not in an AI bubble.” Underinvested: local AI (running AI on your phone/laptop/own data center rather than the cloud) — driven by investor “mimetic behavior” chasing hot topics; biology and chemistry; and robotics — Hugging Face’s own Reachy Mini robots. Robotics data is “huge… petabytes” (video/image vs. text), and the trust problem is even sharper than for the rest of AI: a home robot interacting with your family/privacy shouldn’t be “a black box controlled by a few” — “especially if these organizations’ CEO is not the most stable person in the world.” Open source gives competition and choice; “a world where you have one or two choices is a very scary world because you give up some of your agency.”

What was actually ingested

The full auto-generated (ASR) English caption track, ~36:57 runtime, nine chapters, all present and consistent with length_seconds: 2217. Fetch-pipeline note: the youtube-transcript-skill Playwright path hit the known long-format “transcript panel did not render” failure (GH #2) at both 30s and 60s timeouts, so captions were pulled via yt-dlp (--write-auto-subs, English asr track); the VTT’s rolling-window duplication was collapsed to unique lines and chapters interleaved by timestamp before landing the raw file. ASR cleanups applied: “Clem Delong”→“Clem Delangue”, “Taking Face”→“Hugging Face”, “NeMo tron”→“Nemotron”, HTML entities decoded. Model names spoken by the participants (GLM 5.2, Nemotron, GPT 5.6, Anthropic Mythos/Fable) are preserved verbatim. Uncertain proper nouns left as heard: “National Design Agency” (the PII-detection-model releaser), “Evox Productions” (the lawsuit), and the closing producer credits.

Dynamic-capabilities tagging

• digital-sensing/digital-scouting — Delangue is scanning technological and competitor trends: the open-vs-closed resurgence, the frontier-API-cost curve, China overtaking the US in open-model downloads (~41% per HF’s Spring 2026 report), and screening frontier labs as digital competitors. See digital-scouting.
• digital-seizing/balancing-digital-portfolios — the interview’s core operating decision is a portfolio balance between internal and external options: frontier APIs for experimentation/high-value tasks vs. owned/open models for production at scale, with an explicit “appropriate speed of execution” (start off-the-shelf, then optimize/post-train). See balancing-digital-portfolios.
• digital-transforming/navigating-innovation-ecosystems — Hugging Face is a digital ecosystem (“GitHub for AI,” 16–17M builders, half the Fortune 500) that companies join, and through which they interact with external partners and exploit new ecosystem capabilities; the platform strategy (“create 100× the value, capture 1–2%”) is ecosystem-navigation as a business. See navigating-innovation-ecosystems.
• strategic-renewal/business-model — two business-model claims: (a) the enterprise shift from renting to owning AI is a change in the value-creation/value-capture logic of AI adoption; (b) Hugging Face’s own contrarian, capital-efficient, community-first platform model (no round in 3 years, turned down Nvidia, network-effects “social network” arc). See business-model.
• contextual/external-triggers — disruptive digital competitors (Chinese open models), the regulatory environment (US administration limiting private model releases; open-source “rebrand” needed), and the concentration-of-power dynamic are the external triggers framing the whole conversation. See external-triggers.

Linked entities and concepts

• Channel/publisher: TechCrunch — first appearance; dangling (single-source, deferred per author-entity promotion). The show is TechCrunch’s Equity podcast.
• Created Hugging Face entity — this is the second source citing Hugging Face (after the Agentic Evaluations Workshop, where “Hugging Face” is the author:), so the org clears the second-source promotion bar; central subject here (its CEO, platform metrics, business model, and Reachy robots).
• Dangling (single-source, deferred, per the person-entity second-source rule — central-subject status does not itself trigger promotion for individuals, precedent: Sal Khan / Jensen Huang on their first sources): Clément “Clem” Delangue (Hugging Face co-founder/CEO — central subject; promote on a second substantive Delangue source) and Rebecca Bellan (TechCrunch Equity host).
• NVIDIA — the strongest neighbour; the own-vs-rent thesis convergence (see this source’s relationships:), and both cite NVIDIA’s Nemotron open-weight family.
• Hugging Face Agentic Evaluations Workshop — same organization; open-evals/open-benchmarks ethos as the evaluation-side twin of this interview’s open-model thesis.
• Concepts: open-source-ai (created in this ingest — the own-vs-rent / open-weight / AI-sovereignty / concentration-of-power theme this source most substantively advances), enterprise-ai-adoption (the frontier-experiment → owned-production adoption flow), foundation-models (open vs. closed frontier), dynamic-capabilities (the five W&W cells above).
• Products/terms new to the wiki, held in body prose rather than promoted to entities: GLM 5.2 (open model, Zhipu-lineage as spoken), Reachy / Reachy Mini (Hugging Face open-source robots), Evox Productions (copyright plaintiff), and the passing mentions of Reflection, Cruise, Airbnb/Brian Chesky, Stability, Runway, Palantir/Alex Karp.

Relationships

See frontmatter. Two typed supports edges — to NVIDIA (own-vs-rent convergence, shared Nemotron citation) and to the Hugging Face workshop (same-org open-AI stance). Considered, not linked: AI Index Report 2026 (tracks open-weight release trends and US-vs-China model output, thematically adjacent to the Spring 2026 China-downloads figure, but no shared specific claim or data point — too thin for a typed edge); Hoffman (frontier-vs-specialized decision rule overlaps the frontier-experiment/open-production flow, but Delangue’s axis is open-vs-closed sourcing, not model specialization sequencing — the Huang edge already carries this convergence more directly).

Source quality note

Auto-generated (ASR) transcript, no manual/human-curated caption track fetched; the yt-dlp fallback and dedup are documented in What was actually ingested above. This is a vendor-CEO interview — Delangue leads the largest open-model platform, so claims favouring open source (cost-at-scale switching, safety-through-transparency, concentration-of-power risk) are self-interested and should be read as advocacy, not neutral analysis; the platform metrics (3M models, 16–17M builders, half the Fortune 500, ~41% China downloads) are first-party and uncorroborated here. Per Lifecycle rules, a single vendor-advocacy source does not lift concept confidence above 0.75 on its own.

Debates and supersession

• Open-vs-closed as a safety question. Delangue’s “open source is less dangerous because it’s transparent / concentration is the real risk” is a contestable position — it runs against the closed-labs framing (echoed in the interview’s own premise) that halted the Anthropic Fable and OpenAI GPT 5.6 releases “due to cybersecurity concerns.” The wiki holds both framings without resolving them; see responsible-ai and the new open-source-ai Debates section. No supersession.
• “Companies are done renting their AI” vs. the frontier labs’ trajectory. Delangue’s own hedge (“I’m not worried for them… probably the most valuable companies by next year”) makes this a claim about where workloads run, not about frontier-lab viability — compatible with, not contradicting, the wiki’s frontier-model material. Flagged so future ingests keep “most production workloads move to open/owned models” distinct from “frontier labs decline.”