Why a Nation Can't Outsource Its Frontier AI - Alistair Pullen (Cosine AI)

What It Covers

Alistair Pullen, CEO of Cosine AI, explains how his UK-based frontier lab secured government compute allocation on the ISAMBARD supercomputer to build Britain's first sovereign LLM, why nations cannot rely on US AI access, and how constrained resources force architectural and algorithmic innovation in model training.

Key Questions Answered

• •Sovereign AI compute strategy: Cosine AI received government-allocated compute on the ISAMBARD supercomputer cluster in Bristol through the UK's Sovereign AI Unit, eliminating the largest startup barrier to frontier model training. Without this allocation, a comparable project would consume a significant portion of a $50–100M raise, making the effort financially unviable for a company of Cosine's size.
• •Inference-free licensing model: Cosine deploys model weights directly to customer GPUs or customer-managed cloud environments like Azure and AWS, rather than hosting inference themselves. This means revenue comes from technology licensing, not token margins, dramatically reducing infrastructure costs and making frontier model development viable without the billions Anthropic or OpenAI spend on inference data centers.
• •Active parameter count over total parameter count: Open-source models like DeepSeek V3 have 1.6 trillion total parameters but only 30–50 billion active per token, optimized for inferability rather than raw performance. Closed models like Claude Sonnet likely run 100+ billion active parameters from a 1.1–1.5 trillion total MOE architecture, which Pullen argues explains the persistent performance gap between open and closed frontier models.
• •Credit attribution in RL training: Standard RL assigns uniform weight to every token in a 256,000-token rollout based solely on a final pass/fail signal, reinforcing both correct decisions and irrelevant filler equally. Cosine targets high-entropy decision points within trajectories, attributing advantage disproportionately to those tokens, producing more targeted weight updates, better sample efficiency, and models that learn reusable abstractions rather than surface-level patterns.
• •Synthetic grader generation for coding RL: Most real-world software engineering tasks — feature work, refactoring — lack built-in test suites, making RL reward signals impossible without additional infrastructure. Cosine built an 18-month pipeline using custom post-trained models to synthesize implementation-agnostic graders for arbitrary codebases, enabling RL data generation across languages including Java, Fortran, C++, Verilog, and SysVerilog without requiring pre-existing test frameworks.