Nvidia in talks to buy Hugging Face: the investor refused in January is the buyer in August

Four facts, stated with the qualifiers they came with. Nvidia has been in talks to acquire Hugging Face at a valuation of more than $13 billion, per one source familiar with the matter. No deal has been reached. The talks could still fall apart. Microsoft also met with Hugging Face, and those talks are not ongoing. All four are from Business Insider’s report of 27 August, by Katie Roof, Geoff Weiss and Ashley Stewart — Roof being the reporter who broke the “exploring a sale” story four days earlier.

One correction before anything else. The Hacker News thread that had reached 537 points by our collection on 27 August (and was still climbing) ran under the title “Nvidia agrees to acquire Hugging Face for $13B”. That is not what the article says. Nothing has been agreed; “in talks” and “agrees” are different claims, and the gap between them is exactly where acquisition stories go to die. If you read the discussion yesterday, you read hundreds of comments reacting to a stronger headline than the reporting supports. It happens; it is worth knowing it happened here.

The timeline reads better backwards

Last week we wrote that two distribution layers went up for sale in the same month, and built the piece around one January fact the coverage kept missing. That fact now has a punchline:

Read the valuation column as a curve: $4.5B in 2023, $7B refused in January, more than $13B on the table in August. The number nearly doubled in seven months — and what is being repriced is the position, not any product announcement. Seven months from “we do not want a single dominant investor” to talks about selling the whole company — to the company that sentence was about. A commenter in the thread, transitorykris, compressed it fairly: well, if they didn’t want a dominant investor, they didn’t get one — they got acquired. I would resist the urge to narrate that as hypocrisy. Boards change their minds when the number roughly doubles; that is what numbers are for. The more useful reading is what it says about the asset: the price of a neutral distribution layer is rising faster than the cost of staying neutral.

Business Insider also supplied the context for why Nvidia can even have this conversation: the company said this week it has $18 billion committed to equity investments for the rest of its fiscal year, on top of $47.9 billion it already holds in private companies. At that scale, a $13B acquisition is not a bet-the-company move. It is portfolio allocation.

What the thread actually argued about

The thread (a few hundred comments at collection time, still growing) sorted into four questions. The price got its share of incredulity too — but the incredulity was itself an argument about what is being bought, which is question three below.

The GitHub question. bensyverson reached for the obvious precedent — Microsoft buying GitHub in 2018 for $7.5B — and hoped for the same hands-off stewardship. rvz supplied the financial rhyme: GitHub went for roughly thirty times its reported revenue then — and if you run the same arithmetic on the unverified $150M revenue figure floated in this thread, the multiple here lands near ninety. More defiant, not less. The comparison is genuinely load-bearing, but note where it bends: Microsoft did not sell the machines developers ran their code on. Nvidia sells the machines these models run on. Vespasian put the optimistic version in one line — handing out digging spots, when you are the shovel manufacturer, sounds reasonable. To extend his image one step further than he did: the pessimistic version is the same business, with the digging spots drawn to route past everyone else’s mines.

There is also a sentiment timestamp worth preserving. kpw94 dug up the HN thread from six months ago in which ggml.ai — the company behind llama.cpp — announced it was joining Hugging Face, and whose top comment called the platform “more ‘Open AI’ than OpenAI”. The question attached: does that sentiment survive with Nvidia as the boss. Six months is a short shelf life for a compliment like that, and nobody in yesterday’s thread repeated it.

The open-source-record question. A long thread relitigated Nvidia’s history with open source, Torvalds quote included. make3 made the counterpoint that matters: Nvidia has a structural incentive to keep the hub open, because open models sell GPUs. noosphr supplied the asterisk — llama.cpp, the project from the announcement above, is precisely the tool for running models on hardware that is not Nvidia’s. An open hub and an owner whose margins depend on one kind of silicon can coexist; the question is what happens at the margins, and the margins are where quantized weights for competing accelerators live. andy99 named that concern precisely: crowding out or downplaying “non Nvidia-relevant quants”.

The business-model question. Gigachad asked what Hugging Face’s business model actually is, and the thread’s honest answer was: mostly hosting an enormous amount of free traffic. One commenter cited revenue around $150M a year against a $13B price — I have not verified that figure and report it as a commenter’s number — and another commenter read the same figure as evidence of runway rather than weakness, so the thread was not one-sided. The shape of the question stands regardless. A hub that loses money on free distribution gets bought either to monetise the distribution or to own the position. Neither of those is “keep everything exactly as it was”.

The quiet hypothesis. RachelF offered the darkest reading: good free models on local hardware are a threat to Nvidia’s investments in the large closed labs, and owning the hub is a way to manage that threat. I file this as a hypothesis, not a finding — but it is the one reading under which the acquisition is strategically coherent rather than merely expensive, which is exactly why it deserves to be written down before we know the outcome.

Business Insider’s own report, to its credit, names the tension without any of the HN heat: “Nvidia ownership could also complicate one of Hugging Face’s strengths: its neutrality. The platform supports models and hardware from across the industry, including Nvidia competitors such as AMD and Intel.”

Porting last week’s three measurements from routers to hubs

Last week’s piece ended with three measurements for auditing a router’s neutrality after an acquisition — the fee measured against direct prices, how quickly competitors’ new models land in the catalogue, and whether routing stays explicable. A model hub is not a router, but each measurement has a direct translation, and all three can be started today, deal or no deal:

1. Terms of free distribution. The router version watched the fee gap. The hub version watches what free downloads cost: rate limits, authentication requirements, storage quotas for hosted weights. These are the hub’s price sheet, and like every price sheet, the interesting changes arrive quietly.

2. Treatment of competing hardware’s artifacts. The router version watched catalogue latency. The hub version watches whether quants and runtimes targeting non-Nvidia silicon — the llama.cpp ecosystem above all — keep first-class placement: search ranking, model-card presence, download mirrors. A pattern across months, not a single incident, is the unit of evidence.

3. Granularity of published numbers. Identical in both versions. Hugging Face publishes download counts per model and per organisation; we used exactly those numbers to measure where each country’s open-weights output actually flows. If that granularity coarsens after a change of ownership, the audit trail for the first two measurements disappears with it. The baseline costs one line a month:

   curl -s "https://huggingface.co/api/models?author=<org>" \
     | jq '[.[] | {id, downloads}] | {models: length, dl30d: (map(.downloads) | add)}'

   One calibration note from running this on 24 August: the endpoint truncates at 1,000 models per organisation, so the largest publishers read low. For watching your own organisation, or the policy toward any normal-sized one, it is exact enough to catch a slope change.

The reason to start now rather than after a deal closes is baseline. A measurement without a pre-acquisition baseline cannot distinguish drift from policy — and the pre-acquisition window is the only time you can still collect one.

Our disclosure, unchanged

We said this last week and it stays said: we are a router vendor, neutrality is part of what we sell, and every measurement above applies to us as much as to anyone Nvidia might buy. The talks may collapse — Business Insider says so in plain terms, and one of the two parties has walked away from the other once already this year. If they do collapse, you will have spent a cron job on nothing. If they don’t, you will be one of the few people holding a before picture of the largest neutral distribution layer in AI changing hands. The trade is available today, and the asymmetry is entirely in your favour.