Own Your Intelligence

Mistral’s new training platform targets enterprises whose domain knowledge is too specialized for rented AI models.

Mistral used GTC to announce a new open-source model, a formal verification agent, a founding seat in NVIDIA’s Nemotron Coalition, and an enterprise training platform called Forge. Forge is the announcement that reveals the most about where Mistral believes enterprise AI is heading.

The platform allows organizations to train models on their own proprietary data across the full model lifecycle. Forge assumes that enterprises adopting AI agents for operational roles will need to own the underlying models.

The Operational Ceiling

The dominant architecture for enterprise AI in 2026 pairs a general-purpose frontier model with retrieval-augmented generation or a round of lightweight fine-tuning. The model provides broad capability, and the company’s data fills in the specifics at inference time. For customer service bots, internal knowledge search, and document summarization, this stack performs adequately. Most enterprise AI deployments run on rented intelligence, and most of them work.

The demands change as enterprises move from copilots and chatbots to agents that execute multi-step workflows autonomously. An agent selecting compliance-relevant tools, enforcing internal approval policies, and chaining decisions across a dozen steps needs domain logic available in its weights. Mid-workflow retrieval requires querying a vector store for the right policy document, trusting that embeddings captured the relevant terminology, and confirming that the retrieved context is current. At each of those steps, the wrong document, stale data, or missed jargon can send the entire chain off course.

Mistral’s head of product, Elisa Salamanca, described working with a hedge fund whose proprietary quantitative languages were too specialized for retrieval-based approaches to handle. The firm used reinforcement learning to develop custom benchmarks and train a model that outperformed on them. The hedge fund’s problem is a leading indicator. As enterprise AI shifts from answering questions to running operations, the gap between general-purpose model knowledge and institutional knowledge widens, and the organizations that feel it first are the ones whose institutional knowledge is most specialized.

Targeted Investment

Forge’s announced partners map the terrain where custom training makes sense. ASML builds the lithography systems on which advanced semiconductor manufacturing depends, and Ericsson operates telecom infrastructure at global scale. The European Space Agency, DSO National Laboratories Singapore, and HTX Singapore work in defense and aerospace. Each operates in a domain where institutional knowledge is proprietary, highly technical, and central to competitive advantage or national security.

Futurum’s analysis of Forge concluded that the platform targets a real but narrow segment. Customer service automation, the most popular enterprise AI use case, does not require custom model training, because general-purpose models with structured prompting handle it adequately. Full-lifecycle training also demands significant GPU compute, clean internal data pipelines, and organizational willingness to commit to AI as long-term infrastructure. Most companies lack at least one of these prerequisites.

The organizations in Forge’s early adopter profile are also among the highest-value AI customers in their respective industries. Semiconductor lithography, telecom infrastructure, aerospace, and defense are domains where a competitive edge in AI translates directly to revenue or strategic capability. If autonomous agents expand the range of enterprise tasks that depend on deep domain knowledge, the segment that needs custom training grows with it.

Training on the Job

Forge packages the training methodology that Mistral’s own scientists use to build the company’s flagship models. The platform supports the full lifecycle: pre-training on internal datasets, continued pre-training, supervised fine-tuning, direct preference optimization, and reinforcement learning. It generates synthetic training data tailored to enterprise workflows and evaluates model quality against customer-defined KPIs, with regression suites and drift detection to catch degradation over time. Customers that run training on their own GPU clusters pay a platform license fee, and Mistral embeds forward-deployed engineers with clients that need hands-on support, a delivery model borrowed from Palantir and IBM.

Mistral’s architectural argument for this approach centers on agentic reliability. When domain knowledge lives in model weights, every step in a multi-step agent workflow inherits that understanding without a separate retrieval call. RAG-augmented agents face characteristic failure modes: wrong documents retrieved, stale context from outdated embeddings, and terminology that falls outside a vector store’s coverage. Embedding the domain in weights reduces those risks structurally.

Forge also exposes interfaces for autonomous agents to launch training jobs, optimize hyperparameters, schedule runs, and generate synthetic data. Salamanca has described testing how Mistral’s own Vibe coding agent can run training experiments autonomously, positioning model improvement as an ongoing operational process. The vision is a closed loop in which agents run on custom models and custom models improve through agent-driven training.

Building the Alternative

Forge did not arrive alone. At GTC two weeks ago, Mistral made four major announcements. Mistral Small 4 is a 119-billion-parameter mixture-of-experts model that activates only six billion parameters per token, unifying chat, reasoning, vision, and agentic code capabilities in a single deployment under the Apache 2.0 license. Leanstral is an open-source formal verification agent for the Lean 4 proof assistant, built on the same efficient MoE architecture, which outperformed much larger models on proof benchmarks at roughly one-fifteenth the cost. Mistral also joined the NVIDIA Nemotron Coalition as a founding member, co-developing the base model for NVIDIA’s upcoming Nemotron 4 family on DGX Cloud alongside Perplexity, Cursor, LangChain, and Black Forest Labs.

Mistral closed a €1.7 billion Series C in September 2025, led by ASML at an €11.7 billion valuation. Annual recurring revenue reached an estimated $400 million in January 2026, having grown roughly 20x in twelve months, and CEO Arthur Mensch has publicly targeted exceeding €1 billion for the year. Mistral Compute, a European data center running 18,000 NVIDIA Grace Blackwell chips, is under construction for launch later this year.

Taken together, the GTC announcements reposition Mistral from model maker to infrastructure provider. Mistral is wagering that a critical mass of large organizations will choose to invest in owning their AI, and that European data sovereignty regulation will accelerate this shift. Agentic AI must mature fast enough for operational domain knowledge to become a decisive advantage, though, because without that pressure, the convenience of renting frontier models remains difficult to displace.

Early Days Yet

Enterprise infrastructure tools frequently follow a common trajectory: expensive and narrow at launch, broader as the technology matures and the cost of entry drops. Forge sits at the early stage of that curve, serving organizations whose domain specificity already justifies the investment.

Most enterprises will continue renting frontier models for the foreseeable future. Each new operational task assigned to an agent, though, strengthens the case for domain knowledge embedded in weights. The market that Forge addresses is a function of how fast agentic AI matures, and that process is visibly accelerating in 2026.