Here’s what’s on our plate today:

• 🧪 The infrastructure behind the open AI explosion.

• 📰 Google Search's quiet death, Gen Z boos the AI bots, and a new global AI fault line.

• 🛠️ Weekend To-Do: run a model locally, browse Hugging Face, try a Chinese frontier model.

• 🗳️ Poll: What does the rise of open-source AI mean for the next phase?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• Hugging Face runs the other internet: What started as a failed chatbot is now the central hub of open AI, hosting 2M+ models, 500k+ datasets, and 11M users, with 30% of Fortune 500 companies maintaining accounts there.

• The llama.cpp acquisition closes the loop: By absorbing the team behind the tool that lets LLMs run on laptops, Hugging Face now controls model hosting, model definition, and local inference. No other company owns all three.

• China quietly won the download war: Chinese-developed models account for 41% of Hugging Face downloads, ahead of the U.S., with Alibaba’s Qwen family alone spawning 100k+ derivatives, more than Google and Meta combined.

• The stakes: A movement built to decentralize AI now depends on one French-American startup, and the field’s center of gravity has shifted to Chinese labs and independent developers rather than Silicon Valley.

The infrastructure behind the open AI explosion

For most users, current artificial intelligence tools are cloud-based services accessed through a browser tab or a smartphone app. You type a prompt into ChatGPT, Claude, or Gemini, a remote server processes the request somewhere inside a massive data center, and a response appears within seconds.

The system feels immediate and personal, but the underlying infrastructure remains far removed from the user: owned, operated, and controlled by a small group of technology companies that manage the models, the computing power, and the cost of access. This centralized, platform-driven version of AI has come to define the industry itself, shaping everything from investor enthusiasm and enterprise adoption to regulatory debates over the technology’s future.

However, beneath the surface, a second version of AI has been quietly accumulating mass. This version of AI lives on hard drives rather than behind APIs. The models can be downloaded, copied, modified, and run on a laptop with Wi-Fi switched off, eliminating reliance on hyperscalers’ massive data centers. Instead, this ecosystem relies on open model repositories, independent developers, university researchers, small labs, and a sprawling online community of hobbyists and quantizers who compress models to run on smaller consumer machines.

Many of these systems are built in the open, shared freely across the internet, and used without subscriptions, enterprise contracts, or corporate approval. They rarely appear in earnings calls or government briefings, yet in 2026, they account for a growing share of experimentation, customization, and real-world development across the AI landscape. What began as a niche movement among researchers and hobbyists is now reshaping who controls AI development, which countries dominate the open ecosystem, and how dependent the industry remains on Silicon Valley platforms.

The most visible institution in that second economy is a company called Hugging Face, founded in 2016 by three French entrepreneurs who originally tried to build a chatbot for teenagers, failed at it in the consumer sense, and stumbled into something closer to a public utility for AI distribution. Hugging Face is, in effect, the index page of the open-source AI world. Watching what happens there is one of the better ways to see what the technology actually is, separate from what its largest companies say it is.

What the platform really sells

Hugging Face describes itself, with deliberate modesty, as a place where AI developers share models, datasets, and applications. According to its own Spring 2026 State of Open Source report, the site now hosts around 11M users, more than 2M public models, and over 500k public datasets. Roughly 30% of the Fortune 500 maintain verified accounts on it. Its last public valuation, set in an August 2023 funding round reported by Fortune, was $4.5B, with money from Google, Amazon, NVIDIA, Salesforce, IBM, AMD, Intel, and Qualcomm. That investor list is unusual in AI because no single hyperscaler dominates it, which is part of why the platform reads as neutral ground rather than someone’s territory.

The neutrality is commercially relevant because the thing Hugging Face actually sells is not the models, since they are mostly free to download. What gets billed is hosting, enterprise features, inference services, and convenience. The arrangement it has made with the open-source world looks something like this: it will keep the doors open at low cost, and the world will, over time, route a meaningful share of its AI workloads across its infrastructure. The value of that position grows in proportion to how much of the stack it covers, which is why what happened on February 20, 2026, is interesting.

The acquisition that closed the loop

That day, Hugging Face announced it was absorbing a small team called ggml.ai. Relatively unknown, the team’s main product is a piece of software called llama.cpp, which lets a large language model (the kind that normally needs a data center) run on a regular computer by translating its weights into a format light enough for consumer hardware. That format is called GGUF, and it has become the default way open models are packaged for local use. If you have ever run an AI model on a MacBook or a phone, the code almost certainly passed through this project.

The project’s author, the Bulgarian engineer Georgi Gerganov, first published llama.cpp in March 2023, just after Meta’s LLaMA weights leaked online. Within days, the code had been ported to Windows, then to a Pixel 6, then, slowly, to a Raspberry Pi. Most of the consumer-facing tools people use to run AI on their own machines today, Ollama, LM Studio, GPT4All, and several others, are wrappers around it.

What the acquisition gives Hugging Face is coverage rather than revenue, since llama.cpp remains free and open-source. In its own announcement, the company described the deal as uniting model hosting (the Hub), model definition (its transformers library), and local inference (llama.cpp and GGUF) under one organization. No other company controls all three. That gives Hugging Face an unusual level of influence over the open AI ecosystem. A movement built to avoid central control is increasingly depending on a single platform. Some users in the GitHub discussion under the announcement raised that concern, asking whether quotas or login requirements might eventually be implemented. For now, the open-source license is the main safeguard, since the code itself can be forked if governance drifts.

What the rails are carrying

That concentration only matters if the traffic flowing across the rails is large and interesting, and the same Spring 2026 report makes clear that it is. Reading the data is the part of the story most worth your attention, because it reveals an open-source AI that looks very different from the one Western policy debates assume.

Chinese-developed models now account for 41% of all downloads on Hugging Face, surpassing the United States. Independent developers, meaning people not affiliated with any company, have risen from 17% to 39% of downloads in roughly three years. Industry’s share has fallen from around 70% before 2022 to roughly 37% in 2025. Derivatives of Alibaba’s Qwen family alone, according to a US-China Economic and Security Review Commission report (a US government source) published in March 2026, exceed 100k models, more than Google’s and Meta’s open ecosystems combined.

The numbers should be read carefully rather than dramatically. Download counts favor smaller models, which are easier to run and are automatically pulled by software pipelines, and Chinese labs have released disproportionately many small, deployable ones. Frontier capability on the most demanding benchmarks still tends to lag behind closed American models like Claude, ChatGPT, and Gemini, as CNN reported in its coverage of DeepSeek’s V4 preview in late April. Western labs have also alleged that some Chinese groups have distilled capabilities from closed models without permission, suggesting that part of the surge reflects transfer rather than independent invention. The Stanford AI Index, however, as Al Jazeera summarized, now describes the capability gap between Chinese and American models as effectively closed.

What the success of the other AI might mean

Taken together, the acquisition and the report describe an open-source AI economy that has succeeded in roughly the way its early evangelists hoped, and in some ways they did not anticipate. Models are genuinely accessible, individuals can compete with corporate research divisions, and the cost of running serious AI on a personal machine has fallen by orders of magnitude. The composition of who is doing that building, though, has shifted in ways that complicate the older story about Western leadership in AI. The center of gravity now sits with Chinese labs and unaffiliated individuals rather than with the corporate research divisions that dominated the field five years ago. The institution that holds the whole thing together is a French-American startup whose valuation looks modest next to OpenAI’s or Anthropic’s, and whose reach across the open layer of the field is increasingly hard to substitute.

Five years ago, AI looked like a race between a handful of American companies building ever-larger models inside giant cloud platforms. In 2026, it increasingly resembles something messier: a global ecosystem of downloadable models, independent developers, Chinese research labs, and open infrastructure held together by companies that were barely known outside developer circles only a few years ago.

Friday Poll

Headlines You Actually Need

• Google Search's quiet death: Google Search as we knew it is officially over, with AI Overviews and chat-style results pushing the classic blue-link era out of mainstream use.

• Gen Z boos the AI bots: Young workers are pushing back hard on AI tools entering the workplace, a generational shift that's surprising executives expecting easy adoption.

• A new global AI fault line: A fresh BBC report frames the rising friction between US and Chinese AI ecosystems as one of the defining geopolitical fault lines of the decade.

Weekend To-Do

• Run a model locally: Install LM Studio or Ollama and pull a Qwen or Llama model, the easiest way to feel why open-source AI is exploding.

• Browse the Hugging Face Hub: Spend 20 minutes on Hugging Face filtering models by task and language, and you'll see how much of AI now lives outside the big labs.

• Try a Chinese frontier model: Test Qwen Chat or DeepSeek to compare against ChatGPT or Claude, the capability gap is smaller than the headlines suggest.

Meme Of The Day

The Toolkit

• Regie.ai: AI sales agent that researches prospects, writes personalized outbound at scale, and handles follow-ups so reps can focus on closing.

• Replit AI: A browser-based coding environment with an AI agent that builds, debugs, and deploys apps from prompts, with no local setup required.

• Sourcegraph: Code intelligence platform with Cody, an AI assistant that understands your entire codebase so it can answer questions and write code that actually fits.

Rate This Edition