Automates complex physics simulations to run around the clock
Frees engineering experts to refine core hardware designs
Verifies autonomous tool actions with dedicated review agent
Uncovers data gaps to continuously improve historical dataset
Proxima Fusion uses Gemini Enterprise Agent Platform to build engineering agents, accelerating the path to clean fusion energy.
Proxima Fusion uses Gemini Enterprise Agent Platform to build engineering agents, accelerating the path to clean fusion energy.
Moving past manual constraints to
design the future of fusion energy
Nuclear fusion has long been the holy grail of clean energy. By fusing light hydrogen atoms together under intense heat—replicating the process that powers the sun—it promises to create a safe, endless supply of carbon-free electricity. Dedicated to making this promise a reality is Munich-based Proxima Fusion, which aims to put fusion electricity on the grid by the 2030s, with its own working power plant.
To achieve this, Proxima Fusion is designing and building stellarators, which use complex magnetic coils to trap plasma heated to 100 million degrees Celsius.
To optimize these designs without building highly expensive physical prototypes, Proxima Fusion relies heavily on specialized simulations. However, running these simulations was previously a lengthy process, with individual simulations taking hours or even days to execute. While the company already used extensive automation around its simulations, it needed a way to multiply the number of simulations a user could orchestrate. Interpreting the data and relaunching simulations also required a considerable amount of expert input from the Proxima Fusion team, which slowed its cycle speeds down.
Proxima Fusion launched project AlphaClaw to accelerate this simulation process, with the aim of building a digital colleague that could execute simulations and carry out technical design tasks from start to finish. As a participant in the Google for Startups Cloud Program, the company chose to build this agentic process with Google Cloud.
“We chose Google Cloud because we needed a platform where our existing physics code could interact directly with state-of-the-art AI models,” says Christoph Neumaier, AI product architect at Proxima Fusion. “The Google Cloud team brought deep technical expertise in agent architecture, giving us the tools to safely expose our core engineering systems to AI.”
We chose Google Cloud because we needed a platform where our existing physics code could interact directly with state-of-the-art AI models. The Google Cloud team brought deep technical expertise in agent architecture, giving us the tools to safely expose our core engineering systems to AI.
Christoph Neumaier
AI Product Architect, Proxima Fusion
Designing a digital engineering colleague
with Gemini agents
To bring AlphaClaw to life, Proxima Fusion engineers worked side by side with Google technical specialists over a four-week project. Operating as a single team, they ran daily standups to align their goals and held a dedicated hackathon at the Google Munich office. Today, the resulting agent operates natively right where the engineering team works—inside their Slack workspace—enabling physicists to instantly run and analyze simulations.
We found Gemini models highly capable at navigating technical code frameworks and executing complicated developer steps. They reason through our parameters accurately, which means our engineers can let the agent manage long-running simulation tasks knowing each run is tightly controlled and orchestrated.
Christoph Neumaier
AI Product Architect, Proxima Fusion
When a researcher types a natural language request into Slack to initiate a simulation, Cloud Run acts as a responsive gateway, processing the message in under a second. To manage the heavy computational workloads that complex physics models require, Pub/Sub routes the task into an organized queue. This queue triggers high-performance Compute Engine virtual machines configured to handle long-running mathematical models.
Meanwhile, Gemini Enterprise Agent Platform drives the automation by running Gemini models through the Gemini Command Line Interface (CLI) in headless mode. Operating in the background, this headless engine works directly within internal code repositories to automatically build parameters, script execution files, and cross-reference three years of history across Cloud SQL for the database and Cloud Storage for associated data files, where both historical and new data live. The agent saves all final simulation data in Cloud Storage and posts detailed charts, plots, and visual summaries straight back into the Slack thread. It also posts pull requests with the code it used to generate the data, which is essential for reproducibility.
Because the agent operates directly on internal engineering systems, monitoring for safety and security issues, like leaking intellectual property or compromising data, is essential. Proxima Fusion used Gemini to build a secondary, independent reviewer agent on the platform. This model acts as a built-in safeguard, inspecting every command the agent intends to execute. If an action looks unfamiliar or attempts to alter a critical database, the reviewer agent pauses the run and messages a human engineer in Slack for explicit approval, ensuring researchers maintain complete control over the design loop.
“We found Gemini models highly capable at navigating technical code frameworks and executing complicated developer steps,” Neumaier explains. “They reason through our parameters accurately, which means our engineers can let the agent manage long-running simulation tasks knowing each run is tightly controlled and orchestrated.”
A brighter timeline for grid-ready power
with Google Cloud
AlphaClaw is currently in a live beta phase with 30 researchers, dramatically accelerating the team’s design timeline. As a shared, collaborative workspace, AlphaClaw multiplies individual capacity by running parallel simulations remotely from a single Slack thread without consuming local machine power. This drops entry barriers by offering access to the full simulation stack anywhere with zero setup, while compounding hard-won know-how into permanent memory pools.
Because every session is centrally governed behind a firewall, the system maintains strict security and cost visibility across all runs.
Proxima Fusion now plans to expand the rollout to 80 engineers. The company is also working with Google Cloud to introduce an advanced memory system using enterprise context storage capabilities on Gemini Enterprise Agent Platform. This will automatically organize key details from context history, converting successful design approaches into permanent engineering skills the agent can reuse.
To maintain control over its proprietary physics code, Proxima Fusion runs these workloads entirely within European data zones on Google Cloud. This lets the company keep data strictly within local boundaries as it works toward putting clean fusion electricity on the grid.
"At Proxima we attack every problem we face with advanced technology," concludes Martin Kubie, co-founder and Chief Engineer at Proxima Fusion. "In the next era of simulation, the search for design solutions will be accelerated by AI. AlphaClaw is that idea in practice: with Google Cloud, we've embedded the most advanced autonomous agents inside our engineering workflow. Through accelerating simulation, we're shortening the path to getting fusion onto the grid."
In the next era of simulation, the search for design solutions will be accelerated by AI. AlphaClaw is that idea in practice: with Google Cloud, we've embedded the most advanced autonomous agents inside our engineering workflow. Through accelerating simulation, we're shortening the path to getting fusion onto the grid.
Martin Kubie
Co-founder and Chief Engineer, Proxima Fusion
Fusion energy company Proxima Fusion is pairing deep physics engineering with advanced AI systems to design and build a stellarator power plant capable of delivering clean, safe, and abundant base-load power to the global energy grid.
Industries: Utilities, Technology, Startup
Location: Germany
Products: Cloud Run, Cloud SQL, Cloud Storage, Compute Engine, Gemini, Gemini Command Line Interface (CLI), Gemini Enterprise Agent Platform, Google Cloud, Pub/Sub
