NVIDIA and London-based Ineffable Intelligence have begun an engineering collaboration to build infrastructure for large-scale reinforcement learning, marking a sharper industry push towards AI systems that learn from experience rather than relying mainly on static human-generated data.The partnership, announced on 13 May, brings together the world’s dominant supplier of AI computing systems and a new frontier AI lab founded by David Silver, the researcher best known for his work on AlphaGo and AlphaZero. The companies will co-design hardware and software pipelines intended to support agents that act, observe, receive feedback and update their behaviour continuously.
Jensen Huang, NVIDIA’s founder and chief executive, described the next frontier of AI as “superlearners — systems that learn continuously from experience”, saying the two companies would work on infrastructure for large-scale reinforcement learning as Ineffable seeks to develop a new generation of intelligent systems.
The collaboration will begin on NVIDIA’s Grace Blackwell platform and is expected to be among the first to explore the company’s upcoming Vera Rubin architecture. The technical goal is to understand how next-generation computing systems should be designed when AI workloads move beyond training on fixed datasets and towards models that generate knowledge through simulation, trial and error.
Silver has argued that AI research is entering a harder phase. “Researchers have largely solved the easier problem of AI: how to build systems that know all the things humans already know,” he said. “But now we need to solve the harder problem of AI: how to build systems that discover new knowledge for themselves. That requires a very different approach — systems that learn from experience.”
Reinforcement learning is not new. It has powered breakthroughs in games, robotics and control systems, where models learn by pursuing goals and receiving rewards or penalties. What is changing is the scale of ambition. Ineffable is seeking to apply the method to broader domains of intelligence, where systems may need to generate their own data, test strategies in rich environments and improve without constant human labelling.
That shift places heavy demands on infrastructure. Large language models are typically trained by moving vast stores of text, images and code through a system. Reinforcement learning systems must instead operate in tight loops: an agent takes an action, observes a result, receives a score and updates its model. These cycles can put intense pressure on interconnects, memory bandwidth, inference serving, simulation engines and orchestration software.
For NVIDIA, the partnership reinforces its effort to frame AI infrastructure as a full-stack problem rather than a market limited to graphics processors. The company’s AI platforms increasingly combine chips, networking, software, model libraries and reference architectures designed for what Huang calls “AI factories”. Reinforcement learning adds another layer of complexity because the system must support training, inference and environment interaction at once.
Ineffable’s emergence has also sharpened London’s position in the global race to build frontier AI companies. The company has drawn attention because of Silver’s track record at DeepMind, where reinforcement learning became central to some of the field’s best-known milestones. Its reported $1.1 billion seed financing and multibillion-dollar valuation have placed it among Europe’s most closely watched AI start-ups.
The timing is significant for the wider AI industry. Model developers are facing questions over whether simply scaling larger datasets and bigger models will continue to deliver the same performance gains. High-quality human data is finite, while costs for compute, energy and expert labour continue to rise. Reinforcement learning offers one route around that constraint by allowing systems to create useful training signals through interaction.
The approach also carries risks. AI systems that learn through self-generated experience may produce strategies that are powerful but difficult to interpret. That raises challenges for safety testing, auditability and governance, especially if such models are deployed in scientific research, finance, logistics, cybersecurity or autonomous systems. Infrastructure for reinforcement learning will therefore need monitoring, evaluation and control mechanisms built into the training loop, not added after deployment.
Topics
Technology