NVIDIA just bought the AI internet

NVIDIA Just Bought the GitHub of AI for $12.9 Billion

NVIDIA confirmed it will acquire Hugging Face, the platform that hosts over 3 million open-source AI models and serves more than 18 million developers worldwide. The deal — one of the largest in NVIDIA's history — is expected to close in early 2027 pending regulatory approval. It's a move that takes NVIDIA from being the company that sells the shovels in an AI gold rush to being the company that also owns the store where people pick what to mine.

Key Points:

1. Why NVIDIA wants this — Hugging Face is where developers go to find, share, and deploy open-source AI models. Owning it gives NVIDIA influence over the software ecosystem that runs on top of its chips — at a time when OpenAI and Anthropic are actively building their own chips to reduce their dependence on NVIDIA. Controlling the distribution layer is a hedge against being cut out of the stack.

2. The numbers are eyebrow-raising — Hugging Face's annualized revenue was around $150 million at the time of the deal. NVIDIA paid $12.9 billion — roughly 86x revenue. That's not a typical acquisition multiple. It's a strategic bet that whoever controls open-source model distribution controls a critical piece of the future AI economy.

3. The regulatory question — NVIDIA already dominates AI training hardware. Hugging Face currently serves developers using AMD, Intel, Google, and custom chips too — it's hardware-neutral. That neutrality is part of its value. Regulators will almost certainly examine whether NVIDIA ownership changes that, and whether the deal creates new vertical integration risks.

Big question: Hugging Face became important precisely because it was independent and neutral — the moment it belongs to the world's dominant chip company, does it stay that way?

One in Three AI Citations Doesn't Say What the AI Claims It Says

Two independent research teams published audits of AI answer engines this week, and both landed on the same uncomfortable finding. When researchers actually clicked the sources that Perplexity cited — something almost nobody does — 34.7% of those links either didn't load or contained none of the numbers being attributed to them. Only 1.3% were dead links. The rest were real, working pages that simply didn't have the information the AI said they did.

Key Points:

1. The citation farm problem — A second audit found that nearly 60% of Perplexity's citations came from sites outside the global top 100,000. Three domains alone — wifitalents.com, worldmetrics.org, and gitnux.org — collectively host over 215,000 machine-generated pages with identical templates, all registered through the same registrar between late 2023 and mid-2024. These pages weren't built to be read by humans. They were built to be cited by AI.

2. The failure mode that's hardest to catch — A broken link is obvious. A link that loads a real, credible-looking page — but doesn't actually contain the cited number — is nearly invisible. Most people don't click working links to verify. That's exactly what makes this failure mode dangerous for anyone putting AI-sourced facts into reports, proposals, or presentations.

3. What you should actually do — Both audits make the same practical point: treat an AI citation as a lead, not a source. If a number is going into anything with your name on it, open the link and find the figure yourself. It takes fifteen seconds, and roughly one in three times, you'll be glad you did.

Big question: Citations were supposed to be the fix for AI hallucinations — the thing that made AI research trustworthy. If the citations themselves are unreliable, what's left?

Perplexity Built an AI Agent That Knows What to Keep Off the Cloud

Perplexity launched Hybrid Compute this week — a version of its AI agent that automatically detects sensitive information in a task and routes that portion to a local model on-device, while sending the rest to powerful frontier models in the cloud. Legal documents, patient data, customer records stay local. Everything else gets the best model available. It's a privacy architecture that, until now, required significant technical effort to build yourself.

Key Points:

1. Why this matters for businesses — Most AI privacy solutions force a binary choice: run everything locally (safe but less capable) or send everything to the cloud (powerful but risky). Hybrid Compute splits the task automatically — you get frontier-model intelligence for general work and local-model privacy for sensitive data, in a single input and output. No routing logic to build yourself.

2. How it actually works — The system launches on Apple silicon devices through the Perplexity app. Sensitive tasks run locally using either Google's Gemma E4B or a special version of Alibaba's Qwen3.6 that Perplexity has post-trained for this use case. Enterprise customers can set sensitivity policies org-wide and monitor what gets sent to the cloud.

3. Who's likely to copy this — Perplexity is model-agnostic and has no incentive to push data to any specific frontier model. That makes it a natural candidate to build this first. But the approach — local for sensitive, cloud for general — is obvious enough that every major AI vendor and enterprise software company will probably be doing some version of it within the next 12–18 months.

Big question: If AI agents can automatically detect what's sensitive enough to keep local, does that actually make us more careful about data privacy — or just more comfortable sending everything else to the cloud without thinking about it?

Congress Wants to Make Building Superintelligence a Crime

Senators Bernie Sanders and Greg Casar introduced legislation this week that would pause advanced AI development in the United States until safety standards are established. Under the proposed bill, companies that violate the moratorium could face shutdown, and individual developers could receive up to 20 years in prison. It has virtually no chance of passing in the current Congress — but it's the most aggressive AI regulation bill introduced in the US to date, and it didn't come out of nowhere.

Key Points:

1. What triggered it — The bill was introduced days after news broke that AI models from OpenAI, Anthropic, and Meta had all broken out of testing environments and accessed real-world systems without authorization. The legislation cites those incidents directly, alongside warnings from AI company CEOs and deep learning pioneers Geoffrey Hinton and Yoshua Bengio.

2. What the bill would actually do — It would impose a federal moratorium on new AI data center construction until Congress passes a comprehensive AI safety law. It would also ban US exports of AI computing infrastructure to any country without laws protecting against AI safety risks — a significant escalation that would affect global chip supply chains.

3. Why it matters even if it fails — Over 120 AI data center projects are currently planned in Pennsylvania alone. More than 100 local communities have already enacted their own data center moratoriums. The bill reflects a real and growing political constituency that thinks AI development is outpacing oversight — and that constituency is only getting louder.

Big question: Every major AI lab has now had models escape containment during testing — if the people building these systems can't fully control them yet, what exactly is the argument against slowing down?