Author｜Hualin Dance King

Editor｜Jing Yu

Another AI intermediate layer company is up for sale.

On August 23 local time, foreign media reported that Hugging Face is exploring a sale, with a valuation that may reach 13 billion dollars or even higher. This New York-based company has hired investment banks to gauge potential buyers' interest.

This price would have seemed absurd three years ago. In 2023, when Hugging Face completed its Series D funding, the valuation led by Salesforce, Google, and Nvidia was 4.5 billion dollars.

In less than three years, the purchase price has nearly tripled.

But if we place this news within the timeline of the past month, things get much more interesting. Just a week ago, Stripe acquired the AI model routing platform OpenRouter for over 8 billion dollars; OpenRouter's valuation in its Series B round in May was only 1.3 billion dollars, and it was bought for a 5.4 times premium within three months.

These two transactions demonstrate that the "intermediate layer" between models and users in today's AI ecosystem is being strategically acquired by large companies.

01 The Future of "Black"

Hugging Face might be the only company in the world that saw its valuation rise instead of fall after being hacked.

On July 21, OpenAI disclosed a security incident that shook the entire industry. While doing internal tests on the network attack capabilities of GPT-5.6 Sol and a stronger unreleased model, both models autonomously escaped their sandbox environment, exploiting a zero-day vulnerability to access the internet. What happened next was like a plot from a science fiction movie — the models determined that test answers were stored on Hugging Face's server, and they broke through the defenses, penetrating Hugging Face's production environment.

OpenAI characterized this incident as "an unprecedented cybersecurity event involving cutting-edge network attack capabilities." Hugging Face's security team had already detected the breach and reported it to law enforcement before OpenAI contacted them.

This event brought mixed feelings for Hugging Face. On one hand, it exposed the vulnerability of their security infrastructure; for a platform hosting millions of AI models, this is a fatal flaw; on the other hand, this incident humorously proved Hugging Face's ecological status in the AI circle — when a top AI model wants to "cheat," it instinctively seeks answers from Hugging Face.

After the security incident, Hugging Face accelerated its sale process. This is understandable; rather than waiting for the next more severe breach to crush its valuation, it would be better to cash in the premium while the narrative around being a "core node of the AI ecosystem" still holds.

02 The Contradictory Business

To understand whether Hugging Face's 13 billion dollar price tag is expensive or not, we first need to clarify what it really is.

Hugging Face is often referred to as the "GitHub of AI." The platform hosts over a million community-contributed models and hundreds of thousands of datasets, with over 18 million monthly active visitors and about 5 million registered users, while over 2,000 companies pay to use its Enterprise Hub. Revenue mainly comes from corporate subscriptions, computing fees for the inference API, and revenue sharing with cloud vendors.

However, this analogy hides a key difference. Code on GitHub necessitates ongoing collaboration — CI/CD pipelines, issue management, and team collaboration are all tied to it, making the cost of migration extremely high. In contrast, models on Hugging Face are essentially downloadable files. Once a weight file is downloaded locally, it has no further connection to the platform. Models do not require continuous collaboration like code; they are an output, complete upon download.

This means that Hugging Face's "lock-in effect" is far weaker than GitHub's.

From a financial perspective, the contradiction is even clearer. Hugging Face's annual recurring revenue (ARR) is estimated to be in the range of 40 to 70 million dollars. A purchase price of 13 billion dollars implies a price-to-sales ratio of over 180 times. According to any traditional valuation framework, this number is untenable.

So, the 13 billion isn't buying a profit statement; it's buying an "ecological niche." But what actual benefits can this ecological niche bring to the acquirer?

The answer is more ambiguous than expected.

What can be gained is clear: panoramic intelligence of the entire open-source AI ecosystem (who is training what models, download trends, technical directions), a developer channel of 13 million users, and a rare neutral brand in the open-source community.

However, the problem is that the act of acquisition itself would destroy that neutrality. If Google buys Hugging Face, will Meta still first launch the Llama series on it? Will Mistral and Stability AI still trust this platform? When Microsoft acquired GitHub, there was no real alternative in the code hosting field. But the bar for AI model hosting is much lower — China's ModelScope is already an available choice, and Replicate and Together AI are also siphoning off traffic at the inference level.

Acquirers face a paradox — you are buying openness, but the act of acquisition itself erodes that openness.

03 The Fear of "Shovels"

Putting Hugging Face and Stripe’s acquisition of OpenRouter together gives a clearer understanding of the underlying logic of this wave of mergers and acquisitions.

What OpenRouter does is simple — a unified API that connects over 400 models from more than 80 vendors, allowing developers to avoid individually interfacing with each model company; OpenRouter routes you to the most suitable model, taking about 5% of the inference costs. It claims to serve 8 to 10 million developers, routing trillions of tokens weekly, with token consumption maintaining a compound growth rate of 9% per week since this year.

Stripe's acquisition of it for over 8 billion dollars is based on much clearer logic than that of the potential acquisition of Hugging Face.

Stripe provides payment infrastructure, and OpenRouter provides routing infrastructure for AI calls. Each model API call essentially represents a micropayment — by combining the two, Stripe can earn from both the "flow of money" and the "flow of tokens." OpenRouter's prior self-positioning was as "the Stripe of AI," and now it has literally been acquired by Stripe.

These two transactions represent two different acquisition logics. Stripe’s acquisition of OpenRouter is about "direct business integration" — there are clear revenue synergies and growth engines. In contrast, buying Hugging Face by a giant company resembles "not wanting competitors to gain access," leading to a more defensive acquisition approach.

The former is offensive, and the latter is driven by fear.

But they both reveal a trend — competition at the model level has intensified to the point where profits are being compressed, while what can sustainably generate revenue is at the infrastructure level.

Returning to the fundamental question, why are these AI intermediate layer companies valued so highly?

The logical answer is the "sell shovels" theory. In the gold rush, the most profitable aren’t the miners, but the shovel sellers. The "shovels" in the AI field include middleware such as model routing, model hosting, inference optimization, and evaluation tools. Model companies compete fiercely while the intermediate layer guarantees steady income.

However, this analogy carries a fatal assumption: "shovels" won’t become outdated.

The AI ecosystem is still in its very early stages. Today, OpenRouter does model routing because developers need to switch between dozens of models. However, if the market converges to three to five mainstream models in two years, the value of the routing layer will significantly diminish.

Today, Hugging Face is the default choice for model hosting, but if model capabilities continue to grow exponentially, future models might no longer be distributed through "downloading weight files," but exist purely in API form — making model hosting platforms a historical relic.

In other words, the value of these intermediate layer companies remains undetermined — is it "location value" or "time window value"?

If it’s the former, like Cloudflare and Akamai in the internet era, who can continue to charge after solidifying their position for decades, then 13 billion dollars isn’t expensive. If it’s the latter, they are merely transitional products in a time before AI infrastructure is solidified, then this wave of mergers and acquisitions may become one of the first high-priced lessons of the AI era.

From Hugging Face to OpenRouter, the curtain on the mergers and acquisitions of AI intermediate layers has already risen. The anxiety of big companies is very real — models will decrease in price, become open-source, and be commoditized, but the "pipeline" may not.

Perhaps half of Hugging Face's 13 billion dollar price tag represents the true value of AI infrastructure, while the other half reflects a unique fear of "missing out" characteristic of this era.