NVIDIA didn’t buy a model. It bought the town square.

$12.93 billion for Hugging Face. The open pledge is real. So is the landlord.

Town square at dusk with a large iron key hanging in the foreground
Use the square. Own a spare key.

$12,930,300,000.

That’s the number Jensen Huang put on the blog: NVIDIA agreed to acquire Hugging Face. Not a chip. Not a closed frontier lab. The place 18 million builders already go to find, share, and ship open models — 3M+ models, 500K+ datasets, 1M+ apps, 200K+ companies on the hub (NVIDIA blog).

The feed will split into two tribes by Monday: “open is dead” and “Jensen saved open.” Skip both. Adults read the pledge and the calendar.

The deal, without the vibes

Piece Number / fact Source
Headline price $12.93B total NVIDIA blog
Cash to stockholders ~$11.9B SEC 8-K
Employee retention up to ~$1B equity Same
Signed Sep 2, 2026 8-K
Expected close H1 2027 8-K
Review HSR — a real acquisition Public filings / coverage

Nothing flips for builders today. Hub stays under current ops until close. The planning window is the months between now and the deed.

The open pledge (quote it, don’t paraphrase into hopium)

Huang, in public:

Hugging Face will remain an open platform for the entire AI ecosystem… NVIDIA compute will not be required to build on or deploy through Hugging Face.

Also: multi-cloud, multi-accelerator, models from every builder. The 8-K echoes a commitment to keep upload/download open and support other silicon vendors.

Clem’s frame: resources and scale — 18M → 100M builders — and that “the planets aligned” after years of turning offers down (The Register).

Believe them if you want. Measure them either way.

The mistake (same shape as the Astra asterisk)

Last week’s post was about a 37-point score gap depending on the harness. This week’s gap is softer but bigger:

“Open” on a keynote ≠ “open” as a control you can enforce.

A pledge is a statement of intent. A structural guarantee is something that survives a product roadmap meeting in 2028. Until close — and after — watch the boring stuff:

  1. Search / ranking — what surfaces first for “best 7B instruct”?
  2. Inference endpoints — price, defaults, which accelerators feel first-class.
  3. Non-NVIDIA paths — do AMD / Google / custom silicon workflows stay first-class or become “also supported”?
  4. Policy choke points — export controls and “open weights from region X” risk is already named in deal risk language. That’s not conspiracy; it’s in the filing narrative.

You don’t need to rage-quit the hub. You need a mirror.

What to do Monday (before H1 2027)

  1. Inventory your critical HF dependencies — models, datasets, Spaces, CI that pip installs from the hub.
  2. Mirror what you can’t lose — weights + configs + licenses on storage you control (object store, private git-lfs, second registry). If it’s load-bearing for a customer, it isn’t “someone else’s S3.”
  3. Write the exit drill once — one page: how you redeploy if hub search, rate limits, or ToS shift. Don’t wait for a fire drill tweet.
  4. Keep using HF — the pledge is the best-case path. Using the platform and owning mirrors is how adults hedge.
  5. Track close conditions — HSR / EU review can slow or reshape this. Calendar the H1 2027 window; don’t build strategy on “it already closed.”

Same week as Astra. Not a coincidence in your content calendar.

Frontier labs are gating Critical cyber capabilities. The chip landlord just agreed to buy the open hub. Closed gets more powerful. Open gets a richer owner. Your job isn’t to pick a fandom. It’s to keep option value: a harness you understand (post 1) and weights you can still load if the town square changes landlords (this post).

If you only remember one line: NVIDIA didn’t buy a model. It bought the town square. Use the square. Own a spare key.

Sources

Want the one-page “HF exit drill” checklist? Reply MIRROR / join the waitlist.