Graphcore IPU architecture and the SoftBank acquisition aftermath

Graphcore’s Intelligence Processing Unit uses a massively parallel MIMD architecture to target sparse machine learning workloads. After facing significant losses, the company was acquired by SoftBank for an estimated $500 million to $600 million to support a vertical…

Graphcore IPU architecture and the SoftBank acquisition aftermath

The IPU architecture and the software barrier

Nigel Toon and Simon Knowles founded Graphcore to build processors that mimic human brain intuition. They previously sold Icera to NVIDIA in 2011. In 2012, they began discussions in a Bath pub about how to overcome hardware limitations in AI. They decided to design the Intelligence Processing Unit (IPU) to improve efficiency and reduce energy usage. The founders believe that overcoming hardware limitations is more important for advancing AI than focusing on complex software. The company incorporated in Bristol in 2016. The IPU uses a Bulk Synchronous Parallel (BSP) execution model that organizes computation around thousands of independent processor tiles. Each tile contains its own local SRAM to prevent the memory bottlenecks that plague traditional GPUs. This design targets the sparse, irregular compute patterns of machine learning workloads. The Colossus MK2 GC200 uses 1,472 independent processor cores and 900 MB of in-processor memory to run nearly 9,000 independent parallel program threads across its massively parallel MIMD architecture to improve throughput. This hardware delivers 560 TFLOPS FP8, 280 TFLOPS FP16, and 70 TFLOPS FP32 performance at 185W. The company also planned $120 million "brain-scale" supercomputers in 2024. The Bow IPU uses TSMC’s wafer-on-wafer 3D stacking to increase SRAM density. The Bow chip delivers 250 TFLOPS at FP16 and achieves 7.5 TB/s on-chip memory bandwidth.

Financial contraction and the SoftBank buyout

Graphcore attracted major investors like Samsung, Dell Technologies, and Sequoia Capital. In 2016, interest in AI chips surged after Intel acquired Nervana and Google announced its own chip development. In 2019, Matt Miller from Sequoia pushed for investment despite the company not actively raising funds. In December 2018, a $200 million Series D round brought the valuation to $1.7 billion. By 2020, a $222 million Series E round pushed the valuation to $2.8 billion. However, the company struggled to gain commercial momentum against NVIDIA. The NVIDIA CUDA ecosystem created high switching costs for developers because most machine learning workflows rely on it. Graphcore’s Poplar software stack lacked the same adoption levels. This software deficiency led to the collapse of a significant deal with Microsoft. In 2022, the company reported $2.7 million in revenue against $205 million in losses. Because of this, the gap between architectural correctness and commercial success exceeded $200 million per year. The company also cut its headcount by 20% and closed operations in Japan, South Korea, and Norway. In July 2024, SoftBank acquired Graphcore for an estimated $500 million to $600 million. This price was significantly lower than the company’s $2.77 billion peak valuation and was less than the total capital raised.

Metric Colossus MK2 GC200 Bow IPU
Process Node 7 nm TSMC 7 nm TSMC
Cores 1,472 1,472
Threads 8,832 8,832
Memory Bandwidth – 7.5 TB/s
FP16 Performance 280 TFLOPS 250 TFLOPS

The Silicon Trinity and Izanagi deployment

SoftBank is now integrating Graphcore into a vertical AI stack. This strategy combines Arm architecture, Ampere CPUs, and Graphcore IPUs. This effort is part of Masayoshi Son’s plan to build a complete alternative to the x86 and NVIDIA pairing. SoftBank acquired Ampere Computing in November 2025 for $6.5 billion. The company is co-developing the Izanagi chip with Ampere to support the $500 billion Stargate hyperscale infrastructure project. This project is a partnership involving OpenAI and Oracle. SoftBank also injected $457 million into Graphcore in 2026. SoftBank owns Arm and has a significant stake in the Stargate consortium, which gives Graphcore guaranteed deployment targets. Graphcore is expanding its engineering presence with a £1 billion campus in Bengaluru, India. The company plans to hire 500 semiconductor engineers in India, starting with an immediate 100 roles in silicon logical design, physical design, verification, characterization, and bring-up. You should monitor whether the Poplar-Ampere SDK provides enough pull to overcome NVIDIA’s software lock-in. The Izanagi chip combines IPU parallel processing with Ampere’s Arm-based server CPU architecture. Will the Izanagi chip achieve sufficient volume to challenge the status quo in hyperscale data centers?

airtrain.ai
airtrain.ai

The airtrain.ai newsroom covers AI research, models and the tools built on them.

More on this topic

Stay ahead of AI

Get the week's most important AI stories delivered to your inbox every Monday.

No spam. Unsubscribe anytime.

More Stories