--- title: NVIDIA Picks British Startup Ineffable Intelligence to Co-Design Its Reinforcement Learning Infrastructure description: NVIDIA partners with Ineffable Intelligence, the London AI lab founded by AlphaGo creator David Silver, to co-design infrastructure for reinforcement learning after Europe's largest seed round. author: Darie Nani (Editor-in-Chief) date: 2026-05-16T10:47:14.682Z updated: 2026-05-16T10:47:39.587Z canonical: https://www.sovereignmagazine.com/article/nvidia-british-startup-ineffable-intelligence-reinforcement-learning image: https://cdn.nanimediahouse.com/nvidia-ineffable-london-datacentre.webp categories: Artificial Intelligence, Startups content_type: News region: United Kingdom publication: Sovereign Magazine about: - type: Organization name: Ineffable Intelligence description: London-based AI laboratory founded by David Silver, creator of AlphaGo. Raised $1.1 billion in seed funding in April 2026 at a $5.1 billion valuation, the largest seed round in European history. Focused on reinforcement learning systems that learn from experience rather than human-generated data. url: https://ineffable.ai foundingDate: 2025-01-01T00:00:00.000Z industry: Artificial Intelligence sameAs: - https://www.linkedin.com/company/ineffable-intelligence/ --- NVIDIA has chosen a British startup to help design the next generation of AI training hardware. Ineffable Intelligence, the London-based laboratory founded by David Silver, will work alongside [NVIDIA's](https://www.nvidia.com/) engineers to build computing infrastructure purpose-built for reinforcement learning. The collaboration will begin on NVIDIA's Grace Blackwell hardware and extend to its forthcoming Vera Rubin platform, which Ineffable will be among the first outside groups to use. Silver, a professor at University College London, spent more than a decade leading reinforcement learning research at Google DeepMind. He created [AlphaGo](https://deepmind.google/research/breakthroughs/alphago/), the system that defeated the world champion Go player in 2016, and went on to develop AlphaZero and AlphaProof. He left DeepMind in 2026 to found [Ineffable Intelligence](https://ineffable.ai/), which in April [raised $1.1 billion in seed funding](https://techcrunch.com/2026/04/27/deepminds-david-silver-just-raised-1-1b-to-build-an-ai-that-learns-without-human-data/) at a $5.1 billion valuation, the largest seed round in European history. ## What Is Reinforcement Learning and Why Does It Need Different Hardware Most large language models today are trained through a process called pre-training: feeding vast quantities of text, images, or code into a neural network and asking it to predict what comes next. The technique has produced capable systems, but it is limited by the supply of high-quality human-generated data. Reinforcement learning takes a different approach. Instead of studying existing material, an RL system learns by acting, observing the outcome, scoring itself, and adjusting. It operates in continuous loops rather than consuming static datasets in a single pass. Silver described the distinction in plain terms: "Researchers have largely solved the easier problem of AI: how to build systems that know all the things humans already know. But now we need to solve the harder problem of AI: how to build systems that discover new knowledge for themselves." The hardware implications are substantial. Pre-training is a batch process: data flows through the network in large, predictable blocks. RL training is iterative and tightly coupled, placing different demands on memory bandwidth, interconnect speed, and orchestration. [Scale AI](https://scale.com/), the data labelling company, [reported in February](https://www.bigtechnology.com/p/most-ai-training-is-moving-to-reinforcement) that more than half of all its training work now involves reinforcement learning, up from less than a quarter six months earlier. The shift is well under way, but the infrastructure has not caught up. ## How NVIDIA GPU Architecture Supports AI Training at Scale NVIDIA's current flagship for large-scale AI work is the GB200 NVL72, a rack-scale system containing 72 Blackwell GPUs and 36 Grace CPUs connected via fifth-generation NVLink running at 1.8 terabytes per second per GPU. The entire rack functions as a single accelerator delivering 1.4 exaflops of AI performance with 30 terabytes of fast memory. NVIDIA says this configuration offers 30 times the inference throughput of its previous-generation H100 on trillion-parameter models. The company has described demand for Blackwell hardware as sold out through mid-2026, with a backlog of 3.6 million units. Jensen Huang, [NVIDIA's chief executive](https://www.sovereignmagazine.com/article/constant-direct-communication-why-nvidia-s-ceo-drops-everything-for-his-team), called demand "insane" during recent earnings commentary. NVIDIA itself estimates that developing derivative models from a single foundation model through post-training methods, including reinforcement learning with human feedback, now requires roughly 30 times the compute used to pre-train the original. Test-time reasoning, where a model takes longer to think through difficult queries, can demand more than 100 times the compute for challenging tasks. These multipliers explain why hardware designed for the pre-training era may not be adequate for what follows. ## Why NVIDIA Is Investing in Reinforcement Learning AI Companies The Ineffable Intelligence partnership sits within a broader pattern. NVIDIA has committed more than $40 billion in equity investments in AI companies, effectively becoming a venture investor alongside its role as the dominant hardware supplier. Sequoia Capital and Lightspeed Venture Partners led the Ineffable round, but NVIDIA itself invested alongside Google, DST Global, Index Ventures, and EQT. Sceptics note the circularity: NVIDIA funds startups that subsequently spend their capital on NVIDIA hardware. The arrangement creates demand for the company's own products. Supporters counter that the investments build an ecosystem, locking partners into NVIDIA's software and hardware stack before competitors can offer alternatives, and that the startups gain access to unreleased technology and direct engineering support. Huang framed the partnership in terms of the industry's direction: "The next frontier of AI is superlearners, systems that learn continuously from experience. We are thrilled to partner with Ineffable Intelligence to codesign the infrastructure for large-scale reinforcement learning as they push the frontier of AI." ## Europe's Largest AI Seed Round and What It Signals for London Ineffable Intelligence's $1.1 billion raise is notable as much for where it happened as for how much it raised. The company is headquartered in London, Silver holds his professorship at UCL, and the round included $20 million from the UK Sovereign AI Fund and the British Business Bank. European AI companies have historically struggled to raise capital at frontier-lab scale. [Nscale's $14 billion round](https://www.sovereignmagazine.com/article/reshaping-the-ai-infrastructure-game-how-uk-s-nscale-landed-a-14-billion-deal) earlier this year was one sign of change, but the assumption has broadly held that the largest research labs would remain in San Francisco and that European startups would either relocate or accept smaller rounds. Ineffable's raise challenges that assumption directly: the largest seed in European history, funding a lab whose stated ambition is superintelligence, with the founder choosing to stay in London. The UK government's co-investment is a deliberate attempt to keep frontier AI talent onshore. Silver was one of Britain's most prominent AI researchers; his departure from DeepMind to found an independent company in London, rather than in the Bay Area, is the kind of outcome the Sovereign AI Fund was designed to produce. ## NVIDIA AI Infrastructure and the Competitive Landscape NVIDIA holds between [80 and 90 per cent of the AI accelerator market](https://siliconanalysts.com/analysis/nvidia-ai-accelerator-market-share-2024-2026), depending on which analyst estimate one uses, and its share of the training market specifically exceeds 90 per cent. The company [crossed $5 trillion in market capitalisation](https://www.sovereignmagazine.com/article/nvidia-s-valuation-surpasses-germany-s-entire-economy-apparently-that-s-normal-now) in February 2026. Its data centre division generated [$193.7 billion in revenue](https://futurumgroup.com/insights/nvidia-q4-fy-2026-earnings-highlight-durable-ai-infrastructure-demand/) in the most recent fiscal year. [AMD's MI300X](https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html) has found traction with Microsoft Azure and Meta, and the MI350 is due to ship later this year. But AMD's overall share of AI data centre spending remains in single digits, and its ROCm software stack still trails NVIDIA's CUDA ecosystem in framework compatibility and tooling. The more substantive competitive threat comes from custom silicon built by the cloud providers themselves: Google's TPUs, Amazon's Trainium chips, and Broadcom-designed ASICs. These reduce the hyperscalers' dependence on NVIDIA but operate only within their own cloud environments. By co-designing infrastructure for reinforcement learning before any rival has equivalent offerings, NVIDIA is opening a new front. If RL does become the dominant training paradigm, the companies that built their systems around NVIDIA's Blackwell and Vera Rubin hardware will face steep switching costs, much as CUDA's dominance in pre-training created a lock-in that persists today. ## FAQ **Q: What is reinforcement learning in AI?** Reinforcement learning is a type of machine learning where a system learns by taking actions, observing outcomes, and adjusting its behaviour based on rewards or penalties. Unlike pre-training, which learns from existing datasets, RL systems generate their own experience through trial and interaction. **Q: What is the largest seed round in European history?** Ineffable Intelligence, the London-based AI lab founded by David Silver, raised $1.1 billion in April 2026 at a $5.1 billion valuation. The round was led by Sequoia Capital and Lightspeed Venture Partners, with participation from NVIDIA, Google, and the UK Sovereign AI Fund. **Q: What is the NVIDIA Vera Rubin platform?** Vera Rubin is NVIDIA's next-generation AI computing platform, expected to succeed the current Grace Blackwell architecture. Ineffable Intelligence will be among the first external organisations to work with the hardware, as part of its engineering partnership with NVIDIA. **Q: What share of the AI chip market does NVIDIA hold?** NVIDIA holds an estimated 80 to 90 per cent of the overall AI accelerator market and more than 90 per cent of the AI training market specifically. Its data centre division generated $193.7 billion in revenue in fiscal year 2026.