An agreement involving Hugging Face and a new Mistral funding round change the vendor context for open AI. For organizations using it, the useful question is not to predict winners: it is to establish what they could actually move, rebuild and maintain.
Two announcements, two mechanisms not to confuse
On 3 September 2026, NVIDIA announced an agreement to acquire Hugging Face for about $12.93 billion. The acquirer’s announcement also presents commitments concerning Hugging Face’s open, multi-cloud and multi-accelerator nature, without an announced obligation to use NVIDIA compute. A Reuters report syndicated by KSL corroborates the agreement and its order of magnitude. This is an announced agreement: it does not establish its closing or an effective change in license or contract. [NVIDIA source]
On 8 September, Mistral’s official index dated its announcement of a €3 billion Series D at a post-money valuation above €21 billion. A Reuters report syndicated by MarketScreener corroborates the funding. The body of the announcement describes financing and investment objectives. A post-money valuation is an estimated value after the transaction; it is neither revenue nor capacity already delivered. [Mistral announcement, official index]
Where does openness end in your service chain?
Accessible weights are not enough to establish operational autonomy. A team may have access to a model while depending on an inference format (model execution), a hosting service, a tokenizer (the splitting of text into units used by the model), deployment scripts, managed monitoring tools, scarce skills or contractual support. This observation is a governance method, not an accusation of vendor lock-in.
Start with artifacts: weights, configuration, tokenizer, versions, access rights and retention evidence. Then ask who can rebuild the service, where dependencies are documented, and which alternatives have actually been tried. License and contract terms must be read by competent people; this article does not provide legal advice.
Turning dependency into verifiable questions
| Component | Operational owner | Access evidence | Rebuild | Tested alternative | Uncertainty |
|---|---|---|---|---|---|
| To be completed | To be assigned | Not verified | To be documented | Not measured | To be qualified |
The register does not claim to solve a migration. It makes visible what is missing to consider one. For a limited exercise, choose a non-sensitive, non-production application: retrieve only authorized artifacts, rebuild the minimum necessary, compare essential functions and record differences. A failed exit trial is planning information, not proof that a vendor cannot be left.
Fictional example, with no measured result. A team uses an open-weight model to classify synthetic demonstration requests. Its register retains the weight version and configuration, but finds that the deployment script belongs to the provider. It assigns an owner, verifies retrieval rights, then attempts a rebuild in an alternative test environment. It records time spent, compute cost, missing functions and output differences. These observations inform its decision; no savings or equivalent portability is assumed.
Deciding with observed costs and reviewed commitments
Location preference, operational control and contractual terms are three different decisions. Their importance depends on the service and accepted risk. The Partitech article on API, private cloud and on-premises helps situate hosting choices; it does not remove the need to inventory vendor dependencies.
Define review triggers: an effective contractual change, announced removal of an API, artifact unavailability or failure of a rebuild test. These are governance scenarios. NVIDIA’s and Mistral’s announcements prove none of these events for a team using their services.
Independence that is demonstrated
The next vendor committee can request a simple deliverable: component inventory, owner for each piece of evidence, limited exit exercise and explicitly accepted risks. This method gives no investment advice and predicts neither the result of the NVIDIA–Hugging Face agreement nor Mistral’s future capabilities. It makes it possible to track effectively published commitments and decide from observable dependencies.