Nigel Toon

Graphcore designs the Intelligence Processing Unit, a processor architecture built specifically for machine learning workloads, offering a hardware alternative to NVIDIA GPUs for AI model training and inference. The Bristol-based company developed the IPU around a bulk synchronous parallel computation model that distributes model parameters across thousands of processor cores with local memory, achieving high efficiency for sparse and irregular neural network computations that GPUs handle inefficiently.\n\nThe company raised approximately $700 million across six funding rounds including a Series E that valued it at approximately $2.8 billion, with investors including Sequoia Capital, Microsoft, and Samsung Ventures. Graphcore processors are deployed in research institutions including Oxford, Cambridge, and the Rosalind Franklin Institute, as well as commercial AI platforms. The company has shipped multiple IPU generations including the MK2 IPU and Bow IPU, with the Colossus processor and IPU-POD system providing data centre scale AI compute.\n\nGraphcore competes directly against NVIDIA in the AI accelerator market, alongside AMD, Intel Gaudi, and other AI chip startups including Cerebras, SambaNova, and Groq. The AI accelerator market is projected to exceed $100 billion by 2027, driven by demand for model training compute. Graphcore faces the dominant position of NVIDIA and its CUDA software ecosystem as the primary barrier to adoption, requiring significant software investment to match the maturity of CUDA tooling that researchers and engineers have relied on for over a decade.