WED, 03 JUN 2026 · 18:32:05 UTC

Graphcore

Hardware

UK·HQ Bristol·Est. 2016

IPU silicon — UK's homegrown AI accelerator.

5.0

our score

Our take

UK AI chip pioneer acquired by SoftBank; now a strategic bet inside the Arm ecosystem against Nvidia.

At a glance

Best known for
Intelligence Processing Unit (IPU) AI accelerators
Biggest strength
Graph-native silicon architecture with experienced UK semiconductor team
Biggest risk
Competing for hyperscale adoption against Nvidia's entrenched CUDA ecosystem
Stage
Subsidiary (SoftBank)
Primary revenue
Sales of IPU datacenter accelerators, systems, and Poplar software licenses

What they do

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.

Origin story

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.

Key products

IPU (Intelligence Processing Unit)

Graphcore’s flagship AI processor built for graph-native computation, featuring massive on-chip memory and fine-grained parallelism for data center training and inference.

Bow IPU

An advanced processor using 3D wafer-on-wafer stacking technology to boost performance and power efficiency for large-scale AI pod deployments.

Poplar SDK

The proprietary software stack and compiler toolchain that enables developers to deploy PyTorch, TensorFlow, and other ML workloads on Graphcore hardware.

Leadership

  • NT

    Nigel Toon

    Chief Executive Officer and Co-founder

    Previously co-founded Icera (acquired by Nvidia); veteran UK semiconductor executive.

  • SK

    Simon Knowles

    Co-founder and Chief Technology Officer

    Architect behind the IPU’s graph-native design; previously co-founded Icera.

Funding history

Year
Round
Amount
Lead investors
  • 2020
    Series E
    $222M
    Ontario Teachers’ Pension Plan
  • 2024
    Acquisition
    Undisclosed
    SoftBank Group

Strengths & risks

Strengths

  • +Graph-native IPU architecture with high on-chip SRAM tailored for sparse AI models
  • +Experienced founding team with proven semiconductor exits (Icera)
  • +Strong UK/European positioning for sovereign AI and data residency
  • +Tight hardware-software integration via the Poplar SDK toolchain
  • +Strategic backing from SoftBank and alignment with Arm ecosystem
  • +Distinct memory architecture reducing data-movement bottlenecks for select workloads

Risks

  • Nvidia CUDA ecosystem lock-in makes hyperscale customer acquisition extremely difficult
  • Post-acquisition independence eroded; roadmap may serve SoftBank/Arm first
  • Limited public cloud availability compared to GPU instances from AWS/Azure/GCP
  • Uncertain commercial traction for next-gen silicon after 2023 restructuring
  • Talent retention risk in Bristol amid UK semiconductor competition

Recent moves

  1. SoftBank acquisition

    2024

    SoftBank acquired Graphcore outright, taking the British chip designer private and pairing its IPU roadmap with Arm’s infrastructure strategy.

  2. Company-wide restructuring and layoffs

    2023

    Amid stalled revenue growth and a difficult fundraising climate, Graphcore cut headcount and consolidated operations to preserve cash.

Competitive position

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.

What to watch

  • 01SoftBank’s capital commitments and stated IPU roadmap cadence under private ownership
  • 02Any Arm co-design announcements linking Neoverse CPUs to Graphcore accelerators
  • 03External customer wins outside the SoftBank/Arm sphere to prove merchant viability
  • 04Next-generation silicon tape-out and foundry partner selection
  • 05Engineering talent retention in Bristol versus UK and global chip rivals

Frequently asked questions

What is an IPU and how is it different from a GPU?

An Intelligence Processing Unit is a graph-native AI accelerator with massive on-chip memory and fine-grained parallelism. Unlike GPUs optimized for dense vector math, IPUs are architected for the sparse, irregular compute graphs common in advanced machine learning.

Who owns Graphcore now?

SoftBank Group acquired Graphcore in 2024. The company operates as a private subsidiary, with its strategy increasingly aligned alongside SoftBank’s other semiconductor asset, Arm.

Does Graphcore compete directly with Nvidia?

Yes, Graphcore positions the IPU as a data center alternative to Nvidia GPUs for AI training and inference, though it faces a significant ecosystem gap against CUDA.

What machine learning frameworks does Graphcore support?

Graphcore’s Poplar SDK supports popular frameworks including PyTorch and TensorFlow, compiling models down to optimized graph programs for IPU hardware.

Where is Graphcore headquartered?

Graphcore is headquartered in Bristol, United Kingdom, where much of its silicon design and engineering takes place.

How many employees does Graphcore have?

Following restructuring and the SoftBank acquisition, Graphcore employs an estimated 300–500 people, though exact figures fluctuate.

Is Graphcore planning an IPO?

Graphcore previously explored public-market options but was acquired by SoftBank in 2024 instead; an IPO is off the table for the foreseeable future.

What are Graphcore’s key products?

The company’s core offerings include the IPU and Bow IPU processors, along with the Poplar SDK that enables software integration with standard ML pipelines.

The bottom line

Graphcore’s future is now inextricably linked to SoftBank’s capital allocation and Arm’s product roadmap. As a private subsidiary, it no longer faces public-market scrutiny, but it also lacks the independent go-to-market urgency that once defined it. The coming years will reveal whether SoftBank positions Graphcore as a broad alternative to Nvidia or merely as an internal silicon supplier for its own data center and AI compute initiatives.

The IPU remains architecturally distinctive—its graph-native design, abundant on-chip SRAM, and disaggregated memory model offer genuine advantages for specific sparse and probabilistic AI workloads. However, the AI training market has consolidated ferociously around Nvidia’s CUDA ecosystem and, increasingly, AMD’s ROCm. For Graphcore to matter, it must secure a Tier-1 hyperscale deployment that is not owned by SoftBank, or it must become the de facto AI accelerator inside Arm’s Neoverse-powered infrastructure. Without such external validation, the company risks becoming a well-engineered niche asset rather than a category challenger.

Visit Graphcore

Key products

  • IPU
  • Bow IPU
  • Poplar SDK

Related companies

All companies →