← Back to feed

Ineffable Intelligence has raised $1.1 billion in seed funding — the largest seed round in European history — and announced a strategic infrastructure partnership with Google Cloud. The announcement was made at Google Cloud Summit London ‘26. The company is building what it calls a “superlearner”: an AI system that discovers knowledge entirely from self-generated experience, without relying on human-authored training data.

The lab was founded by David Silver, the DeepMind researcher who led the AlphaGo and AlphaZero programs. Silver’s work on reinforcement learning — specifically on agents that develop expert-level skills through self-play rather than imitation — is the direct intellectual precedent for what Ineffable is attempting at scale. AlphaGo required human game records to bootstrap. AlphaZero did not. The Ineffable bet is that the same principle extends beyond games to the full range of cognitive tasks.

What a Superlearner Is

The lab’s stated mission is to build AI that “rediscovers and then transcends the greatest inventions in human history — language, science, mathematics.” The framing is reinforcement learning from first principles: the system generates its own experience, evaluates outcomes, and learns from the gap. No pretraining on internet text, no supervised labels, no imitation of human demonstrations.

This is meaningfully different from the dominant paradigm. Every major frontier lab — Anthropic, OpenAI, Google DeepMind, Meta — builds on large-scale pretraining over human-generated corpora. The architecture can then be fine-tuned with RL, but the foundation is imitation. Silver’s argument, implicit in the AlphaZero line of work, is that imitation puts a ceiling on capability: you cannot exceed what the training distribution contains. Pure RL from experience, given sufficient compute, has no such ceiling.

The practical challenge is that language and science are not games. The state and action spaces are vastly larger, the reward signal is ambiguous, and the compute requirements for meaningful RL at that scale are enormous. Ineffable’s $1.1B buys runway to find out whether those gaps are engineering problems or fundamental ones.

The Compute Setup

Ineffable Intelligence has selected Google Cloud as its preferred cloud provider and will deploy one of the largest clusters of A5X instances, powered by NVIDIA Vera Rubin NVL72 GPUs, on Google Cloud’s infrastructure. Vera Rubin NVL72 is NVIDIA’s current-generation flagship for frontier training, delivering 50 PFLOPS FP4 per rack with Jupiter networking.

Google Cloud CEO Thomas Kurian stated Ineffable is using the “full-stack AI Hypercomputer” — Jupiter networking, optimised storage — rather than a commodity GPU rental. The distinction matters for experience-based learning systems, which place different demands on infrastructure than static dataset training: continuous generation of trajectories, evaluation, and gradient updates require tight coupling between training and inference capacity.

Why $1.1B at Seed

The size of the round is unusual. Seed funding of this scale has previously appeared only for US-based frontier labs in growth stages, not pre-product European startups. Several factors explain it:

First, Silver’s track record. AlphaGo defeated Lee Sedol in 2016, AlphaZero surpassed all human chess and Go records in 2017, and AlphaFold 2 solved the protein structure prediction problem in 2020. Each was a step-change result, not an incremental improvement. Investors are betting on the researcher, not a product roadmap.

Second, the compute requirement. Frontier RL from scratch — without pretraining as a shortcut — requires more compute than any existing model training run. A $1.1B seed is not large relative to that target; it is the capital required to build enough infrastructure to generate a credible result at small scale before the next raise.

Third, European positioning. The EU’s Chips Act 2.0 and Tech Sovereignty Package are directing capital toward non-US AI infrastructure. A London-based frontier lab with credible scientific leadership is a structurally attractive recipient for that capital.

What This Means for the Frontier

The frontier lab count has been effectively fixed for the past year: Anthropic, OpenAI, Google DeepMind, Meta, xAI, Mistral, DeepSeek. Ineffable represents a new entrant with a genuinely different approach — not a variation on transformer pretraining with RL fine-tuning, but an attempt to build capability from pure experience.

The research question is whether Silver’s approach — proven at the scale of board games — can be made to work for open-ended domains. If it can, it would produce a model that has not been trained on the internet and carries none of the distributional biases, copyright entanglements, or knowledge cutoffs of current frontier models. That is a meaningful capability and commercial differentiation, if it works.

Ineffable is building in London. The first compute cluster is on Google Cloud. No model has been announced.