Nvidia Buys Hugging Face | AI Field Notes #99

AnalysisThe federal government picked a side in the fight over whether AI companies can train on your work without paying. In a statement of interest filed September 2 in the New York Times case against OpenAI and Microsoft, a brief where the government tells a court where it stands, the Justice Department called the copying extraordinarily transformative and warned that ruling for the Times would leave only trillion-dollar firms able to afford licensed data. Left unmentioned: the administration is separately negotiating a government stake in OpenAI. A regulator arguing the defendant's case while angling to own a piece of it is a conflict you would flag in a first-year ethics class.

AnalysisThe company that already sells the shovels just bought the general store. Nvidia agreed on September 3 to pay $12.93 billion for Hugging Face, the platform where more than 18 million developers download over three million AI models and half a million datasets. Jensen Huang promised the hub stays open to AMD and other chips, the kind of thing you say when regulators are already circling. Nvidia calls the deal a deconcentration platform. Owning both the compute and the place people find models is the opposite of that word, and the antitrust reviewers in the US and EU know it.

AnalysisGoogle built a model good enough at finding software holes that it will not let most people use it. Alongside the general Gemini 3.8 Flash on September 2, it released a Cyber variant tuned to spot vulnerabilities and write patches, then locked it behind a vetting program called Fairwind, open only to government agencies, critical-infrastructure operators, and software maintainers who apply. More than 650 organizations are already in. The logic is that a tool which automates bug-hunting helps attackers as much as defenders, so access itself becomes the safety control. Capability is no longer the hard part. Deciding who gets the keys is.

AnalysisThe quiet winner of the AI buildout does not make a chip with its own logo on your desk. Broadcom reported on September 2 that its AI semiconductor revenue tripled to $16.7 billion in a single quarter, up 221% from a year earlier and now 56% of the whole company. It builds custom accelerators, chips designed for one customer's exact workload, for a short list of giants, and it shipped its Ironwood processor in volume to both Google and Anthropic. Nvidia still rules the merchant market, but the biggest buyers are quietly designing their way around it, one bespoke chip at a time.

AnalysisMeta is betting that most agent work does not need a genius, just a cheap model that finishes the job. Muse Spark 1.3 landed September 2 in Muse Code and the Meta Model API, tuned for long-running coding and multi-step agent tasks. Meta says it makes about 20% fewer tool calls and burns 25% fewer tokens than version 1.2 while holding a one-million-token context window, the amount of text it can weigh at once. The pitch is running cost. When an agent works for hours, the bill is the product, and trimming tokens is how you win the job over a smarter, pricier rival.

AnalysisA Chinese AI lab that barely existed three years ago is now heading for the public market. Moonshot AI, maker of the Kimi family of models, confidentially filed for a Hong Kong listing on September 3, carrying a valuation around $50 billion, more than eleven times its worth at the end of 2025. About 70% of its revenue comes from API sales, and its annual run rate has passed $300 million. The listing tests whether Western investors will fund a Chinese model builder directly, and whether Beijing will let one of its AI champions raise money on an exchange the US does not control.

AnalysisThe world's highest-grossing law firm has decided that buying AI off the shelf will not cut it. Kirkland & Ellis signed a multi-year deal with Palantir to build its own platform, and the first tool automates the grind of private-equity fund formation: the documents, side letters, and compliance checks that used to eat associate hours. The build leans on 250 Kirkland lawyers and more than 180 technology staff, part of a $500 million commitment the firm floated earlier this year. When the firm at the top of the profession starts automating its own junior work, the pyramid that trains young lawyers gets a lot narrower.

AnalysisCisco just handed every one of its 90,000 employees a personal AI agent, and the interesting part is the plumbing underneath. MyAgent runs supervised tasks across Outlook, Webex, Jira, and SharePoint, backed by more than 800 specialized subagents. To keep the bill sane, Cisco routes 50% to 60% of requests to cheap open-weight models, freely downloadable models it runs itself, sends 20% to 30% to plain software automation, and passes only the leftover sliver to an expensive frontier model. The rollout also landed the same month as layoffs, which is the part the internal launch deck tends to skip.

AnalysisYour doctor may soon be reading your chart through ChatGPT. OpenAI connected its healthcare version to Epic, the electronic records system that holds files on more than 325 million patients, so a clinician can pull your appointment notes, labs, and medications into a conversation or run ChatGPT inside the Epic chart itself. Access is read-only, so the model cannot write back into your record, and UCSF Health is among the first testing it. The convenience is obvious and so is the exposure: a summary that misreads a lab value or invents a drug interaction now sits one glance away from a treatment decision.

AnalysisCoding assistants forget everything the moment you close the session, and a wave of small open-source tools is trying to fix that without sending your code to anyone's cloud. Dejavu, shared on September 2 under an MIT license, free to use and modify, gives agents like Claude Code and Cursor persistent memory that runs entirely on your own machine, with no signup and no backend. It is one of several racing at the same problem, alongside projects that simply index the session history the agents already write to disk. The bet is that memory becomes a layer you own, sitting under whichever assistant you happen to use this month.

AnalysisAdobe spent an undisclosed sum to buy a six-person Indian startup that will then shut down, which tells you it wanted the people and the code more than the customer list. Rilo, founded in 2025 and valued at just $10 million after raising $1 million, built tools that let marketing teams spin up agents for competitor research and sales-call analysis. Adobe folds the technology and the team into its own agent push and closes Rilo to its existing users. This is the acquihire pattern at speed: a startup goes from seed round to absorbed in under a year, and its handful of paying customers are collateral.

AnalysisThe hard part of AI is no longer training the model. It is running the thing cheaply and close to users, inside whatever country's borders the law demands. Equinix said on September 2 it will launch an Inference Exchange with Nvidia and Together AI, letting a company run more than 200 open models across Equinix data centers worldwide while controlling exactly where each request is processed. The selling point is data residency: a bank in Frankfurt can keep inference on German soil to satisfy local law. It does not open until early 2027, so treat this as a direction, a flag planted well ahead of the product.