Skip to content
Annuaire
Sections
Innovation

Remanence: ex-H Company team raises $30M for AI models that companies own

L’essentiel

Remanence, founded by three former H Company staff, raises $30M to help companies own their own AI models. What it changes.

À retenir

Remanence, founded by three former H Company staff, raises $30M to help companies own their own AI models. What it changes.

Three former employees of H Company, the young Paris-based start-up specialising in AI agents, announced on 17 November 2026 a $30 million seed round (around €25.5 million) for their new company, Remanence. Their idea: turn the traces left by a company’s AI agents into specialised models that the company itself owns. The deal is part of a wider debate about how much of their models companies will want to control themselves.

Who is funding it, and how much

According to FrenchWeb and Tech.eu, the round is co-led by EQT Ventures and FirstMark, with participation from Zero Prime, Emblem, Drysdale, Global Founders Capital and Yellow. Several tech entrepreneurs are also investing their personal money: Olivier Pomel (Datadog), Philippe Corrot (Mirakl), Romain Niccoli (Pigment), Victor Riparbelli and Steffen Tjerrild (Synthesia) and Mehdi Ghissassi (AI71), according to ActuIA. No valuation has been disclosed.

FirstMark is no newcomer to this story: the fund took part in the $220 million round with which H Company launched in May 2024, FrenchWeb recalls. It is therefore a shareholder in both the original company and the one founded by its former executives.

Founders from the Holo models

Pierre-Louis Cedoz, presented as CEO in the press release, Hamza Benchekroun and Axel Moyal led the post-training of the open Holo models at H Company. Their careers also include Google DeepMind, InstaDeep and Meta. FrenchWeb notes a discrepancy between the documents: the legal notice of incorporation of the simplified joint-stock company (SAS), created in September 2026 in Paris, names Hamza Benchekroun as chairman and Axel Moyal as managing director, while the press release presents Pierre-Louis Cedoz as CEO.

The case of Remanence shows a team trained at a young AI start-up going on to found its own company, in a narrower niche, with the backing of an investor that already knows it.

What Remanence actually does

The starting point is the tasks that a company’s AI agents already perform. FrenchWeb gives the example of a customer support agent that checks an order and then offers a refund. Remanence plugs into the models in use, converts these execution traces into training environments, then specialises an open-weight model through reinforcement learning: the system rewards successful behaviours and adjusts the parameters accordingly, explains Tech.eu. The customer could then retrain this model when its procedures change, according to FrenchWeb.

The customer keeps the model weights, can keep its data on its own infrastructure and chooses where to train and deploy it, according to ActuIA. One nuance matters, however: the ability to export the model depends on the licence of the original open model, FrenchWeb points out.

What it changes for companies

Remanence is betting that companies will increasingly want to own cheaper models tailored to their procedures, rather than permanently renting the proprietary models of the big American labs. Tech.eu notes that some companies are already seeking to reduce their dependence on OpenAI or Anthropic for cost reasons. Matt Turck, partner at FirstMark, sums up this view: in his opinion, enterprise AI “is becoming hybrid”, with closed models coexisting with models controlled in-house (remarks reported by ActuIA).

For a European IT department, the stakes are twofold. On the one hand, sovereignty: keeping one’s data and weights in-house meets confidentiality requirements and limits dependence on a single supplier. ActuIA also cites a TCS study published in early November according to which 94% of French executives consider independence from models essential. On the other hand, reliability on multi-step tasks: the company claims that its specialised models handle them more reliably than general-purpose models, and possibly at lower cost, reports Tech.eu; this is a promise from the company, not a publicly measured result.

The limits to keep in mind

  • No named customers: the company mentions design partners, in Europe and the United States, from customer support, law and enterprise software publishing, but discloses neither their names, nor figures on results, nor pricing.
  • An already contested niche: FrenchWeb cites Adaptive ML, which claims AT&T as a customer, Veris AI and OpenPipe’s open source ART framework.
  • Heavy spending: a large share of the funds will go to a dedicated GPU cluster, training environments and evaluation tools. The team is to grow mainly in Paris, and the company also plans to set up in London and San Francisco.

At this stage, investors are therefore mainly betting on a team and a thesis. The first production deployments will show whether the “owned” model lives up to its promises against the turnkey offerings of the big labs.

Frequently asked questions

What is post-training of an AI model?

It is the stage that follows the initial training of a large language model: it is refined on specific tasks, for example through reinforcement learning, so that it better follows given instructions or procedures. Remanence makes this its core business, applied to the traces of enterprise agents.

Is Remanence linked to H Company?

Its three founders worked there on the Holo models, and the fund FirstMark has invested in both companies. Remanence nevertheless remains a separate company, created in Paris in September 2026.

Can a small or medium-sized business use this type of service?

Nothing has been announced to that effect: no price list is public and the partners mentioned are companies that already operate AI agents. The approach also requires having enough execution traces to train a useful model.

Sources

Sur votre appareil

Comprendre cet article

L’analyse utilise l’intelligence locale du navigateur lorsqu’elle existe, sinon un résumé extractif. Le texte n’est envoyé à aucun service extérieur.

Facebook X LinkedIn

Ensuite A lire aussi

Free, no spam, one-click unsubscribe.