Nvidia Reportedly Acquiring Hugging Face for USD 12.9 Billion

Nvidia is reportedly paying USD 12.9 billion for Hugging Face. This acquisition aims to secure the layer where developers select artificial intelligence models, influencing the open-weights ecosystem's growth and future compute demand. Hugging Face hosts 2.96 million public model repositories. However, 85.6 per cent are downloaded fewer than 200 times, with just 1.5 per cent accounting for 99.2 per cent of all downloads, according to its State of Open Models report.

Credit: UNSPLASH

The Information first reported this agreement, citing an informed source. CNBC, Fortune, and Forbes have since published the figure. Neither company has confirmed the transaction. A Business Insider account detailed a valuation above USD 13 billion without a signed agreement. The reported price stands as published, not as a formally filed sum.

And this USD 12.9 billion valuation represents nearly 86 times Hugging Face's approximate USD 150 million annualised revenue, also close to three times its USD 4.5 billion valuation after its 2023 Series D funding round. Nvidia, Google, Amazon, Salesforce, IBM, Intel, AMD, and Qualcomm all participated. This multiple indicates payment for market position, specifically where devs choose operational models.

The platform saw substantial expansion last year. Public model repositories increased to 2.96 million, datasets to 1 million, and hosted demonstration applications to 1.44 million. All categories recorded growth exceeding 20 per cent. Traffic predominantly flows through limited model families. Derivatives of Alibaba's Qwen models comprise 151,448 repositories, approximately 2.6 times Meta's footprint.

These Qwen derivatives are added at 180 to 210 new repositories daily. Google's derivatives total 82,506. Most activity focused on smaller models, with only 3 per cent of 2026 downloads exceeding 70 billion parameters. The largest families originate from China.

So, the acquisition targets the address developers frequent for models and the interface for family choices. This strategy mirrors Nvidia's historical practices. CUDA was free for almost two decades, leading to widespread engineer adoption, valuing its installed user base over direct monetisation. Nvidia's USD 96.2 billion revenue for the three months ending Jul. offers context.

The reported USD 12.9 billion acquisition price equals roughly six weeks of Nvidia's sales, securing a key position at the top of the developer funnel. But an open-weight model runs on a downloader's existing hardware, which for most developers supports CUDA. Each fine-tune, quantization, or local experiment consumes compute cycles on the machine's accelerator.

So, this contrasts with closed "frontier" models, which concentrate demand among few data-centre operators building proprietary silicon and negotiating aggressively. Open models, however, spread demand across hundreds of thousands of individual devs purchasing hardware at retail, without negotiation.

A broader open ecosystem shifts the market towards a more favourable structure for accelerator suppliers. Washington's export restrictions also influenced this, showing buyers outside the United States that permissions could be revoked. Their practical response involved obtaining open weights they could control. Qwen derivatives on Hugging Face reflect this eighteen months later.

And the acquisition's success depends on Hugging Face retaining universal appeal and perceived neutrality. Developers choose models because the hub appears unbiased. A repository adopting vendor channel characteristics risks losing its justifying quality. Concerns surfaced quickly after the report.

Nvidia's USD 20 billion Groq deal in Dec. provides a precedent for preserving independent operation as a nonexclusive technology licence with a leadership hire. This left Groq independent with a separate cloud business. Nvidia paid substantially to keep an existing operation running unchanged.

So, the critical assessment will monitor whether AMD, Intel, and hyperscaler runtimes perform equally well on the hub in 12 months compared to the CUDA path. Continued parity suggests Nvidia bought a position on a public road. Performance degradation for non-CUDA systems could prompt developers to move elsewhere, making the USD 12.9 billion an archive acquisition.

Distribution remains a durable asset. Models are publicly released by labs in three countries. Companies capturing economic benefit are those at layers through which all activity must pass. Nvidia controls the compute layer, and has now reportedly paid for the model choice layer. Broadcom and TSMC hold comparable positions for similar structural reasons. Free models raise the value of surrounding layers, a worth this transaction quantifies.

• Nvidia is reportedly acquiring Hugging Face for USD 12.9 billion.

• The acquisition aims to control the distribution of AI models to developers.

• Hugging Face hosts 2.96 million repositories, but activity is concentrated in a small fraction.

• The valuation of 86 times annualised revenue reflects a payment for market position, not direct earnings.

• The continued neutrality of Hugging Face is considered crucial for the acquisition's long-term value.

Source: Forbes