5min read

Authors
Tom Mendoza
Pierrick Cudonnec
Nitya Hinduja
Alexander Fred-Ojala
EQT Ventures is co-leading Remanence’s $30 million seed round alongside FirstMark.
We are investing because the most valuable AI asset a company will own is not the model it rents. It is the loop that makes its own model better every time it runs.
AI with no remanence
In physics, remanence is the magnetism a material keeps after the field that created it is switched off. Most enterprise AI today has none.
A company running agents in production generates thousands of execution traces a day: every tool call, every failure, every correction. It is the most specific data the company will ever produce about its own work, and almost all of it is thrown away. The next run starts from the same general-purpose model, makes the same mistakes, and is billed at the same rate.
With agents, that inefficiency becomes a margin problem. Per-token prices keep falling, but agentic workloads consume far more tokens per task, so total spend keeps rising. That spend goes to a handful of providers who can change price or model at any time. Performance also plateaus, because prompt and harness engineering only go so far when the model never learns the company’s workflows, tools and edge cases.
Open weights changed the math. Environments are the bottleneck.
Open-weight models have closed most of the gap with the frontier. The share of US enterprises running them in production rose from 23% to 67% in a single year (Infrared 2026). The leaders have already acted on this: Cursor and Ramp train models with reinforcement learning on their own usage, and beat frontier APIs on the tasks that matter to them.
Most companies cannot follow. Post-training talent is concentrated in a few hundred people worldwide. The hardest part is not the training run but the environment: a sandbox that mirrors the company’s real systems where agents can practise a task safely, with verifiers that score whether the agent did the job right. Together they are what reinforcement learning needs to improve a model. Today these are built by hand, one customer at a time, over months. That keeps post-training a services business.
Turning traces into training grounds
Remanence automates the hard part. By learning from production execution, the platform identifies where agents succeed, where they fail, and where they can improve. It automatically generates realistic training environments and reliable evaluation mechanisms, including scenarios that agents have yet to encounter. These environments are then used to train specialized AI models that become increasingly capable, reliable, and cost-efficient over time.
The customer owns the weights, its traces train only its model, and its data never has to leave its infrastructure. As traces accumulate and stronger open-weight models ship, the model is retrained on the same environments.
This is why we see environment synthesis as the heart of the business. Models can be swapped as stronger open-weight alternatives emerge, while the underlying training infrastructure remains a durable asset. And because it is generated rather than built by hand, the cost of each new deployment falls instead of scaling with headcount.
The team that took open-weight models to the top of computer use
Pierre-Louis Cedoz (CEO), Hamza Benchekroun (CTO) and Axel Moyal (CPO) led the post-training of H Company’s open-weight Holo models. In March 2026, Holo set the state of the art on OSWorld-Verified, one of the hardest long-horizon agentic benchmarks, ahead of proprietary frontier models at a fraction of their cost. Their work also secured H’s entry into NVIDIA’s Nemotron Coalition.
Pierre-Louis was H’s founding researcher, then its CTO, and built its research team from scratch after working on Gemini post-training at Google DeepMind. Hamza led post-training at H and scaled its asynchronous RL pipeline to very large models, after deploying RL for industry and biology partners at InstaDeep. At H Company, Axel led projects spanning agents, evaluations, and environment synthesis across web, desktop, and mobile applications.
At EQT Ventures, we've been following reinforcement learning and open-weight models closely. When we met Pierre-Louis, Hamza and Axel, their vision immediately aligned with our conviction that proprietary training environments will become a critical part of the enterprise AI stack.
Building and hiring SF, London and Paris, for companies that want to own their AI
The round funds a dedicated GPU cluster in Europe. It also pays for research and engineering hires, mainly in Paris with an early presence in London and San Francisco, and for scaling design partnerships across Europe and the US.
Beyond capital, Remanence can benefit from EQT's broader ecosystem, including access to expertise across businesses with complex, high-volume workflows, as well as experience in digital infrastructure and scaling technology companies.
“Every company running agents is generating its most valuable training data and throwing it away. Pierre-Louis, Hamza and Axel have already shown they can turn open-weight models into best-in-class agents. Remanence brings that to companies that want to own its intelligence rather than rent it,” said Tom Mendoza, Partner, EQT Ventures.
“AI companies spend more on intelligence every quarter and own none of it. We turn the experience their agents already generate into a model they own, one that improves every time it’s retrained,” said Pierre-Louis Cedoz, co-founder and CEO of Remanence.
Models that keep their magnetism
The next generation of enterprise AI won't be defined solely by who has access to the most powerful models. It will be defined by who can turn proprietary experience into intelligence that improves over time. That's the future Remanence is helping build, and why we're excited to back the team.
This is a marketing communication. Please refer to the information document of EQT Ventures before making any final investment decisions.




















































































































































































