Graphcore: How Bristol's AI Chip Pioneer Built a New Class of Processor — and Why It Matters

Back to: Top 10 British Startups 2026

The history of computing is punctuated by moments when a new type of processor opened capabilities that its predecessors could not efficiently support — from the shift from mainframes to microprocessors to the rise of GPUs as the workhorses of graphics and, eventually, machine learning. Graphcore, a company founded in Bristol in 2016 by Nigel Toon and Simon Knowles, has spent the better part of a decade making the case that artificial intelligence workloads are sufficiently distinct from conventional computation to warrant another such architectural leap.

Company Overview

Graphcore's founders brought to the company decades of combined experience in semiconductor design, having previously built companies including ICERA and Element 14, both of which were acquired by global technology firms. Drawing on this background, they conceived an entirely new processor architecture — the Intelligence Processing Unit — designed from the ground up for the sparse, irregular computation patterns characteristic of machine learning workloads. The result is a chip that processes certain AI operations substantially faster and more efficiently than GPU-based systems, while consuming significantly less power — an advantage of growing importance as the energy cost of AI training becomes a major operational and environmental concern.

At its peak, Graphcore achieved a valuation of $2.8 billion, making it one of the most valuable deep-tech companies in British history and attracting investment from leading technology firms including Microsoft and Samsung. Its 2024 acquisition by SoftBank brought new financial resources and strategic alignment with a portfolio of AI companies that includes ARM Holdings, giving Graphcore access to the global distribution and partnership network that a challenger chip company needs to compete with established players.

Business Model

Graphcore sells its IPU hardware to hyperscale cloud providers, AI research organisations, and enterprise customers with demanding machine learning workloads. The company also offers IPU-as-a-Service through partnerships with cloud platforms, making its technology accessible to researchers and developers without requiring the capital investment of hardware purchase. Software is an increasingly important component of the revenue model: Graphcore's Poplar SDK and the PopVision profiling tools give developers the optimisation capabilities needed to extract maximum performance from IPU hardware, and the quality of this software stack has been a significant factor in competitive evaluations.

Post-acquisition, Graphcore's business model is evolving to take advantage of SoftBank's extensive relationships in the telecommunications, robotics, and enterprise technology sectors. These relationships open market opportunities — particularly in the deployment of AI inference at the network edge — that would have been difficult for an independent company to access at scale.

Innovation Factor

The IPU's fundamental innovation is its memory architecture. Conventional processors, including GPUs, are designed around the assumption that computation will be performed on large, densely structured datasets — an assumption that reflects the original applications of these chips in graphics rendering and scientific simulation. Machine learning, by contrast, frequently involves small, irregularly structured computations performed across very large models with parameters distributed across memory in patterns that conventional memory architectures handle inefficiently.

Graphcore's in-processor memory architecture — which places very large amounts of SRAM directly on the processor die, close to the compute units — eliminates the memory bandwidth bottleneck that limits GPU performance on many ML workloads. For certain classes of model, particularly those involving graph neural networks, sparse transformers, and recommendation systems, this architectural advantage translates into performance improvements of an order of magnitude or more compared with GPU-based alternatives.

Market Position

The AI chip market is fiercely competitive, with NVIDIA holding an extraordinary market share in GPU-based training hardware, and challengers including Google's TPUs, Amazon's Trainium, and a constellation of startups competing for specific workload niches. Graphcore's IPU occupies a defensible niche in this landscape — workloads where its memory architecture delivers unambiguous performance advantages — and the SoftBank acquisition provides resources and relationships to expand that niche aggressively. See also: Wayve's AI computing needs and Faculty AI's AI infrastructure work.

What's Next

Graphcore is developing its next-generation IPU architecture, incorporating lessons from commercial deployments and advances in semiconductor manufacturing to deliver further performance and efficiency improvements. The company is also expanding its software ecosystem, investing in frameworks and tools that make IPU programming accessible to a broader developer community — a critical prerequisite for achieving the adoption scale needed to compete with NVIDIA's entrenched ecosystem advantages. Visit graphcore.ai for more.