Clem Delangue, the CEO of Hugging Face, spent a July 10 2026 TechCrunch interview making one argument out loud: companies are done renting their AI. His worry is concrete — that a handful of large providers could end up controlling everything if every studio pipes its data through the same three APIs. For an architecture office watching its BIM models, tender documents and client geometry drift into someone else’s datacentre, that reads less like a philosophy debate and more like a procurement decision — one you make this year.

←TODAY: The Hugging Face Hub lists 45,000+ models reachable through one API, and 50,000+ organisations already build on the open stack. →3012: The offices still standing are the ones that kept their weights, their data, and their inference on machines they controlled. Fulcrum: Renting intelligence is cheap until the day the lock changes — owning the runtime is the only clause that survives a pricing round.

The signal: the open stack grew up

Hugging Face is best known as a website, but the thing worth an afternoon is the software underneath it. The transformers library, the diffusers image stack, the safetensors weight format, PEFT for parameter-efficient fine-tuning — these are Apache-licensed Python packages you can install and run with no account, no key, and no outbound request. The scale behind them is real: the Hub routes 45,000+ models through one unified API, and Hugging Face counts more than 50,000 organisations on the platform — Meta, Amazon, Intel and Microsoft among them, per the company’s own front page. That is the shift Delangue is pointing at. The rented API is one door; the same models walk through a second door onto your own hardware.

The Tool: huggingface/transformers

The one project this tutorial is built on is transformers (github.com/huggingface/transformers), the library Hugging Face maintains as the common interface to language, vision and audio models. It matters to a computational designer because it collapses “find a model” and “run a model” into three lines of Python — and because everything it downloads lands in a local cache you can point at a folder you chose. As Brian Andrus puts it bluntly in DreamHost’s 9 July 2026 local-hosting guide, sending your data to someone else’s cloud can feel like handing a stranger your house keys.

Setup:

python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install "transformers[torch]" huggingface_hub
python - <<'PY'
from transformers import pipeline
clf = pipeline("sentiment-analysis")
print(clf("Local inference keeps the tender data on my own machine."))
PY

When that prints a label and a score, the whole stack works: PyTorch, a downloaded model, and a first inference — all on your laptop, nothing sent out.

First steps:

1. Swap the task. Change "sentiment-analysis" to "summarization" and feed it a paragraph from a real Ausschreibung. You will see the model pull the gist — useful for triaging a stack of tender PDFs.
2. Pin the model. Add model="BAAI/bge-small-en-v1.5" to a pipeline("feature-extraction", ...) call to turn text into a 384-dimension vector. That vector is the atom of every semantic search you will ever build over your own archive.
3. Point the cache. Set HF_HOME=./models before you run, and every weight downloads into your project folder — a directory you can audit, back up, and keep inside your own jurisdiction.

Because the models cache locally, the second run is offline. That is the property that turns a demo into an office policy.

Atelier:

Atelier: A Swiss studio adopting AI this year faces the same fork the municipalities do — rent the intelligence and accept the data-residency terms, or run a small model in-house and keep the geometry on the LAN. Offices that have been through a cloud-tool price change report the same lesson: the migration cost is never the licence, it is the data you can no longer extract. Your Monday move: stand up one transformers summarisation pipeline on a single studio workstation and route this week’s incoming tender PDFs through it — measure whether the local run is good enough before you sign anything that ships those documents off-site.

Hack:

Turn a clause into a vector that never leaves your desk — the first move of any private semantic search over your own drawings and specs.

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("BAAI/bge-small-en-v1.5")
vec = model.encode("Datenresidenz clause in a cantonal tender")
print(vec.shape)  # (384,) — embedding computed and stored locally

Those four lines are the primitive PAZ uses everywhere: text in, a fixed-length vector out, cosine distance for “what else in my archive says this.” It is the same mathematics that lets a NeRF store a whole building as the weights of one small network — the emission-absorption integral Kajiya and Von Herzen wrote down in 1984, now fitted by gradient descent — or lets 3D Gaussian Splatting hold a facade survey as ~10⁶ anisotropic Gaussians. Representation is the discipline; owning the representation is the sovereignty.

The move: read the clause yourself

So the real tool here was never the library at all; the durable move is asking, per pipeline, “where does this run and where does the data rest?” Switzerland kept its digital sovereignty less by grand design than by a few Kantone who insisted on local data residency when nobody else was reading the fine print — federalism as a slowness budget, spent well. You get to make the small version of that decision. When your office or your Gemeinderat procures an AI tool this year, read the data-residency clause yourself, and if it is missing, write it in.

Learn-it:

• Project source — the transformers library and its README: huggingface.co
• Why own-your-AI now — Clem Delangue on renting vs. owning: TechCrunch, 10 July 2026
• Hands-on local hosting walk-through: Local AI Hosting (2026 Guide)
• Is the Hub up? Operational status you can check before a run: StatusGator — Hugging Face Hub
• PAZ note: the same embedding primitive underwrites our Grasshopper↔Archicad Library and PAZ-GPT — keep the vector store where you keep the drawings.