Graphcore

Graphcore is a semiconductor company that designs and markets the Intelligence Processing Unit (IPU), a processor built from the ground up for artificial intelligence and machine learning workloads. Unlike general-purpose GPUs that were retrofitted for AI, the IPU employs a massively parallel, graph-native architecture optimized for the sparse, irregular compute graphs common in modern neural networks. Its signature design choice places exceptional amounts of SRAM directly inside the processing tiles, minimizing the memory bottleneck that plagues traditional accelerators when handling fine-grained operations and probabilistic models.

The company sells its silicon in multiple form factors: PCIe accelerator cards, multi-IPU server blades, and turnkey systems such as the Bow Pod configurations. These products target data center operators, cloud service providers, and advanced AI research labs that need high-throughput training and inference outside of the incumbent GPU stack. Graphcore complements its hardware with the Poplar SDK, a proprietary software stack that compiles standard machine learning frameworks—primarily PyTorch and TensorFlow—into optimized graph programs for the IPU.

Headquartered in Bristol, Graphcore positioned itself as Europe’s answer to Nvidia, emphasizing energy efficiency, data sovereignty, and an alternative architectural philosophy. Following its 2024 acquisition by SoftBank, it is increasingly integrated into a vertically aligned strategy alongside Arm, with its roadmap expected to support next-generation AI data center infrastructure rather than purely independent commercial sales.

Graphcore was founded in 2016 in Bristol, England, by Nigel Toon and Simon Knowles, two serial entrepreneurs who had previously built Icera, a baseband processor company acquired by Nvidia. Their pitch was audacious: in a world where GPUs had become the default AI engine, a purpose-built “Intelligence Processing Unit” could deliver superior performance-per-watt for the graph-based mathematics underlying deep learning.

The company moved quickly, unveiling its first-generation Colossus IPU in 2018 and attracting venture backing from prominent firms. By 2020, Graphcore had achieved unicorn status after raising a substantial late-stage round led by Ontario Teachers’ Pension Plan, with participation from strategic investors including Microsoft and BMW. For a time, it was the great British hope in the semiconductor renaissance, expanding internationally and exploring paths to the public markets.

Yet commercial reality proved bruising. While the IPU won praise for its novel architecture, Nvidia’s CUDA ecosystem and relentless product cadence locked in hyperscalers and enterprises. A difficult 2023 brought restructuring and scaled-back operations. In 2024, SoftBank stepped in, acquiring Graphcore outright and ending its life as an independent company. The deal inserted Graphcore into SoftBank’s broader AI portfolio, pairing it with Arm to potentially co-design future accelerators and compute fabrics.

Graphcore competes in the crowded data center AI accelerator market against Nvidia’s dominant H100/H200/B200 GPUs, AMD’s Instinct MI series, Intel’s Gaudi/Habana line, and well-funded startups such as Cerebras and SambaNova. Where Graphcore distinguishes itself is in architectural philosophy: the IPU was designed natively for computational graphs and fine-grained parallelism, giving it an edge on certain sparse models, probabilistic workloads, and small-to-medium training tasks that do not map efficiently to massive GPU kernels. Its Poplar software stack provides a credible alternative compiler path for PyTorch developers, though it lacks the decade of ecosystem maturity that CUDA enjoys.

Where Graphcore loses is in raw market presence. Nvidia’s bundling of hardware, software, and networking (InfiniBand/NVLink) has created a moat that even AMD struggles to cross. Graphcore has not secured a tier-one hyperscale public cloud deployment comparable to AWS, Azure, or Google Cloud’s GPU offerings, and its post-acquisition status raises questions about whether it will remain a broadly available merchant silicon supplier or become a captive technology provider for SoftBank and Arm. If SoftBank leverages Graphcore to build differentiated AI compute fabrics, the IPU could find a defensible niche; otherwise, it risks marginalization against the incumbent GPU duopoly.