Two years ago, self-hosting your AI was a bad idea. In 2026, quietly, it has flipped. The objections everyone still repeats — it is too expensive, too complicated, and leaves you with weaker models anyway — were all true two years ago. None of them really survives 2026.
Objection one: the models are weaker. Not any more.
Start with the models, because it is the objection people cling to hardest. The open weights anyone can download today — Qwen, DeepSeek, Llama, Nemotron — now match frontier performance on the work most B2B applications actually need: reading documents, answering technical questions, and pulling clean structured data out of mess.
I benchmark this on every project, and on domain-specific tasks the gap is consistently 5% or less. Sometimes the local model simply wins. The frontier still leads on open-ended reasoning at the very edge; that is largely irrelevant to the jobs a manufacturer actually needs done.
Objection two: the hardware is out of reach. It costs about £5,000.
A single machine built around a 32GB GPU, roughly £5,000 to £6,000 all in, will run a 30-billion-parameter model, an embedding model and a reranker at the same time. After that, your cost per query is zero. No data centre on the other side of the world, no cloud bill that climbs with every prompt, and a fraction of the energy of the hyperscale equivalent. For a B2B tool serving 10 to 50 users, that one box is your entire inference infrastructure. We size the hardware to the application — not everyone needs 32GB. (The VRAM budget is here if you want the engineering detail.)
The real argument isn’t cost at all
For the engineering and manufacturing businesses I work with, the deciding factor is data sovereignty. When your documents hold proprietary processes, client specifications or defence-related work, sending those queries off to a cloud API is simply a non-starter. Self-hosted means the data never leaves the building. Full stop.
And the capital cost is no longer the blocker it was: one box pays for itself within two to four months of your first AI workflow going live. I have written separately about why owning beats renting over time and about the subsidised token pricing that makes cloud look cheaper than it really is. Cortex runs on exactly this basis: on your hardware, with no external AI calls.
The capability is sitting there, accessible, waiting on a decision rather than a breakthrough. So if it isn’t cost, and it isn’t capability, what is actually holding most SMEs back?
Weighing up deployment models? Read our full comparison: self-hosted AI vs cloud AI for engineering and manufacturing businesses.