Graphcore

Graphcore is a British semiconductor company that designs the Intelligence Processing Unit (IPU), a processor architecture purpose-built for machine learning rather than adapted from graphics or general-purpose designs. Founded in 2016 in Bristol, England, it became the United Kingdom's most prominent AI chip unicorn before being acquired by SoftBank Group in 2024.^[1]

History

Graphcore was founded in 2016 by Simon Knowles and Nigel Toon, veteran chip designers who raised hundreds of millions of dollars to build a new class of AI processor. The company attracted early investment from technology firms including Dell Technologies Capital, Samsung, and BMW, and its funding rounds made it one of the most valuable private semiconductor startups in Europe — at its peak a so-called "unicorn" valued above US$1 billion.^[2]

The first-generation IPU products shipped to datacentre customers and researchers from 2019 onward, positioned as an alternative to NVIDIA GPUs for training and inference. In July 2024 SoftBank Group announced its acquisition of Graphcore, stating the deal would strengthen its ability to build "next-generation of AI compute" alongside its Arm holdings.^[1] In May 2026 SoftBank confirmed a further investment of roughly US$450 million into the company, with Graphcore continuing to operate under its own brand and its CEO Nigel Toon remaining at the helm.^[3]

Technology

The IPU differs from GPUs in several structural ways:

• Massively parallel MIMD cores. Each IPU contains thousands of small cores that each execute their own instruction stream, suited to the irregular control flow of machine learning workloads, rather than the lock-step parallelism of GPU shader cores.
• On-chip memory (SRAM). IPUs place large amounts of memory next to the compute, accessed by the cores directly, to reduce the memory-bandwidth bottleneck that limits many AI accelerators. Models too large for on-chip memory use a managed mechanism called IPUExchange that streams tiles of data in and out during computation.
• Bulk-synchronous programming. The Poplar software model organises computation into explicit phases (compute, exchange, synchronise), giving programmers fine control over data movement — powerful but steeper to learn than CUDA.^[2]

Graphcore's later products, the Bow IPU (2022), which added compute-in-memory boosts directly into the silicon, and the MK2000, targeted larger training workloads. Despite technical differentiation, the company struggled against NVIDIA's CUDA ecosystem lock-in and the scale of GPU supply chains, and reportedly explored a sale before the SoftBank deal.^[2]

Market position

Graphcore belongs to the broader wave of AI accelerator startups — including Cerebras Systems, Groq, and SambaNova — that have tried to challenge GPU dominance with specialised architectures. Its acquisition by SoftBank reframes it from an independent merchant chip vendor into part of a larger vertically integrated vision spanning Arm CPUs, chips, and cloud infrastructure. Analysts generally view the deal as an acknowledgement that competing with NVIDIA as a standalone company required capital and distribution on a scale few startups can sustain.

>Key Takeaways

• Graphcore designs the IPU, a machine-learning-specific processor with on-chip SRAM memory and massively parallel MIMD cores.
• It was the UK's leading AI chip startup before its 2024 acquisition by SoftBank Group, which injected a further ~US$450M in 2026.
• Its struggles illustrate the difficulty of challenging NVIDIA's CUDA ecosystem as an independent vendor.
• For Malaysia, its story connects to the country's strength in semiconductor packaging and its push up the value chain.

See Also

• NVIDIA
• Cerebras Systems
• Groq
• AI Accelerators
• Penang Semiconductor and AI Ecosystem