{"slug": "put-the-model-in-the-basement", "title": "Put the Model in the Basement", "summary": "A Swiss provider operating a single 64-GPU cluster in Zurich could generate about CHF 7.4 million in annual revenue and CHF 3 million in EBITDA by selling dedicated Kimi K3-class inference to banks, pharmaceutical companies, and government bodies, with a three-year payback period on $7 million in initial costs and $370,000 in monthly operating costs, according to an analysis by Philipp Dubach. The model, Kimi K3, is a 2.8-trillion-parameter open-weight model, and the provider would offer Swiss data residency, audit access, and exit terms to meet sovereignty demands, though the business faces concentration risk and hardware depreciation.", "body_md": "This article follows [I Tried Kimi K3 Inside Claude Code](https://philippdubach.com/posts/kimi-k3-inside-claude-code/). That test asked if Kimi K3 could work inside a familiar coding setup. Here, I ask what changes when a Swiss provider runs a K3-class model for customers that need local control.\n\nBanks already treat data sovereignty as an operating issue. In [How DORA Made Sovereignty a Bank Problem](https://philippdubach.com/posts/dora-critical-cloud-providers-sovereignty/), I argued that banks must plan for concentration, exit, audit rights, and rule changes by critical foreign providers.\n\nAI gets the same treatment. Recent open-weight releases make local use more realistic. [Kimi K3](https://www.kimi.com/blog/kimi-k3) is huge. Alibaba is moving the Qwen family in the same direction. [Inkling](https://thinkingmachines.ai/news/introducing-inkling/) may matter more. It is a US-trained model with full weights, a one-million-token context, and a focus on customisation.\n\nAssume a Swiss provider operates one 64-GPU cluster in Zurich. It sells dedicated or managed K3-class inference to banks, pharmaceutical companies, government bodies, and other customers that need Swiss data residency.\n\nThis case assumes $7 million in initial costs, monthly costs of $370,000, and average use of 70%. Subscriptions cost CHF 15,000 to CHF 50,000 each month. With these values, one cluster makes about CHF 7.4 million a year. Its EBITDA is about CHF 3 million. The payback period is three years.\n\nThe provider sells control as well as tokens. It offers Swiss residency and a defined security boundary. It does not train on customer data. It gives customers audit access, continuity terms, and an exit route if it fails.\n\nMost companies will not put a model in a basement. Some will pay local providers to get the same control.\n\nA 2.8-trillion-parameter model costs a lot to serve. One cluster creates a concentration risk. Hardware ages fast. If use falls, the case loses money. A smaller model may work almost as well next year. Then the hardware loses value before the operator recovers its cost. I still think this market will exist.", "url": "https://wpnews.pro/news/put-the-model-in-the-basement", "canonical_source": "https://philippdubach.com/posts/put-the-model-in-the-basement/", "published_at": "2026-08-15 00:00:00+00:00", "updated_at": "2026-08-16 14:10:35.566737+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-infrastructure", "ai-products", "ai-policy"], "entities": ["Kimi K3", "Alibaba", "Inkling", "Thinking Machines", "Philipp Dubach"], "alternates": {"html": "https://wpnews.pro/news/put-the-model-in-the-basement", "markdown": "https://wpnews.pro/news/put-the-model-in-the-basement.md", "text": "https://wpnews.pro/news/put-the-model-in-the-basement.txt", "jsonld": "https://wpnews.pro/news/put-the-model-in-the-basement.jsonld"}}