Graphcore
Graphcore is a Bristol-based semiconductor company that designs Intelligence Processing Units (IPUs)—specialized AI accelerators intended as an alternative to NVIDIA's GPU dominance.
Profile
Design and sell specialized AI accelerator chips (IPUs) that compete with NVIDIA GPUs for training and inference workloads.
Graphcore is a Bristol-based semiconductor company that designs Intelligence Processing Units (IPUs)—specialized AI accelerators intended as an alternative to NVIDIA's GPU dominance. Founded in 2016 by Simon Knowles (CTO) and Nigel Toon (CEO), the company raised approximately $767 million before being acquired by SoftBank for roughly $500 million in July 2024. As a wholly-owned SoftBank subsidiary, Graphcore continues to operate independently.
The company's core product is the IPU architecture, with its second-generation Colossus MK2 (GC200) containing 59.4 billion transistors and offering significant performance-per-watt advantages for specific AI workloads. Graphcore achieved a $2.77 billion valuation in 2020 but struggled commercially, recording only $2.7 million in revenue in 2022 (down 46% from prior year) with $204.6 million in losses. The Microsoft Azure partnership—once positioned as a turning point—collapsed primarily due to software challenges rather than hardware defects.
Post-acquisition, SoftBank has doubled down: beyond the initial $500 million acquisition, it invested an additional $457 million in May 2026 and authorized up to £1 billion for a new AI Engineering Campus in Bengaluru, India, targeting 500 semiconductor engineering roles. Graphcore plans to double UK headcount to around 750 people, focusing on silicon design and software engineering. Co-founder Simon Knowles departed in August 2025, roughly one year post-acquisition, citing a desire to pursue other interests.
CEO Nigel Toon remains in place. The company's market pitch centers on IPU efficiency for large-scale AI training and inference, but faces persistent headwinds from GPU ecosystem lock-in and limited enterprise adoption relative to NVIDIA's entrenched CUDA ecosystem and software maturity.
Who buys this
- Cloud infrastructure providers (e.g., Microsoft Azure) evaluating alternative accelerators
- Academic and research institutions conducting AI research and drug discovery
- Financial services firms testing AI for natural language processing and risk modeling
- AI research labs and biotech companies requiring efficient parallel processing
Publicly disclosed clients
- Microsoft (Azure IPU preview partnership, though partnership reportedly ended)
- Oxford Nanopore (DNA sequencing using IPUs)
- University of Oxford
- Citadel
- J.P. Morgan (evaluating for NLP and speech recognition)
Strengths and what to watch
Strengths
- Proprietary IPU architecture with differentiated design for ultra-high throughput AI training workloads
- Sustained and escalating backing from SoftBank: $500M acquisition plus $457M additional investment (2026), signaling long-term commitment
- Second-generation Colossus MK2 processor delivers 8x performance improvement over first generation with 59.4 billion transistors
Watch for
- Co-founder and CTO Simon Knowles exited August 2025 (one year post-SoftBank acquisition); succession risk and brain drain signals
- Persistent cash burn: $767M raised, acquired for ~$500M, still loss-making; Bengaluru campus ($1B) and UK headcount doubling are execution bets
- GPU dominance entrenched via NVIDIA's CUDA ecosystem and developer familiarity; IPU adoption remains niche despite claims of superiority
Recent moves
Key Information
- Industry
- AI Hardware
- Founded
- 2016
- Headquarters
- United Kingdom
Frequently Asked Questions
What is Graphcore?
Graphcore is a Bristol-based semiconductor company that designs Intelligence Processing Units (IPUs)—specialized AI accelerators intended as an alternative to NVIDIA GPUs. Founded in 2016 by Simon Knowles and Nigel Toon, the company was acquired by SoftBank in July 2024 and operates as a wholly-owned subsidiary.
What are Graphcore IPUs and how do they differ from GPUs?
Graphcore IPUs (Intelligence Processing Units) are specialized AI accelerators designed for high-throughput training and inference workloads. Unlike GPUs optimized for graphics, IPUs feature a proprietary architecture delivering superior performance-per-watt for specific AI tasks, though GPU ecosystem dominance limits IPU adoption.
How much has SoftBank invested in Graphcore?
SoftBank acquired Graphcore for approximately $500 million in July 2024. Beyond the acquisition, SoftBank invested an additional $457 million in May 2026 and authorized up to £1 billion for a new AI Engineering Campus in Bengaluru, India, targeting 500 semiconductor engineering roles.
Did Graphcore's Microsoft Azure partnership succeed?
No. Graphcore's Microsoft Azure partnership collapsed primarily due to software challenges rather than hardware defects. Despite IPU technical advantages, the partnership failure highlighted the persistent difficulty of displacing NVIDIA's GPU ecosystem lock-in and developer familiarity in competitive AI infrastructure markets.
Why did Graphcore cofounder Simon Knowles leave the company?
Simon Knowles, Graphcore's cofounder and CTO, departed in August 2025—roughly one year post-acquisition by SoftBank. He cited a desire to pursue other interests. His exit raised succession risk concerns and potential brain drain signals, coinciding with challenging commercial performance and market headwinds.
What are Graphcore's biggest challenges?
Graphcore faces entrenched GPU ecosystem dominance via NVIDIA's CUDA platform and developer familiarity. The company struggles with limited enterprise adoption, persistent cash burn despite SoftBank backing, and the challenge of proving IPU superiority in a market where switching costs are high.
How Graphcore compares
Direct head-to-head against 3 competitors. Picked by 7wData.
Graphcore
- Positioning
- Design and sell specialized AI accelerator chips (IPUs) that compete with NVIDIA GPUs for training and inference workloads.
- Customer segments
- Cloud infrastructure providers (e.g., Microsoft Azure) evaluating alternative accelerators
- Strengths
- Proprietary IPU architecture with differentiated design for ultra-high throughput AI training workloads
- Watch for
- Co-founder and CTO Simon Knowles exited August 2025 (one year post-SoftBank acquisition); succession risk and brain drain signals
- Recent moves
- SoftBank invests additional $457 million in Graphcore AI chip development
Cerebras Systems
- Positioning
- Wafer-scale silicon for large-model training and inference, targeting hyperscalers, foundation model labs, and sovereign AI programs.
- Customer segments
- Hyperscalers, foundation model labs, sovereign AI programs, and enterprises running large-scale training and inference workloads.
- Strengths
- Single-die Wafer Scale Engine eliminates inter-chip communication overhead, delivering high memory bandwidth on large model runs.
- Watch for
- IPO S-1 flagged 86% revenue from UAE-linked entities at filing; customer concentration is a disclosed material risk.
- Recent moves
- IPO on Nasdaq raised $5.55 billion at $185 per share, May 14, 2026; fully diluted valuation reached $56 billion.
SambaNova Systems
- Positioning
- Fifth-generation RDU architecture for cost-efficient agentic inference, targeting enterprise and sovereign AI production deployments.
- Customer segments
- Enterprises, sovereign AI programs, government organizations, and model providers requiring production-grade AI inference at scale.
- Strengths
- Reconfigurable Dataflow Unit maps model execution directly onto silicon, cutting memory access overhead for decode-heavy inference workloads.
- Watch for
- Abandoned Intel acquisition discussions then raised $350M instead, February 2026; pre-profitability chipmaker dependent on continued investor backing.
- Recent moves
- SN50 RDU launched for agentic inference, Intel multi-year partnership signed, $350M Series E closed, February 2026.
Groq
- Positioning
- Fast inference platform for production AI deployments; continues independently after NVIDIA licensed its LPU architecture, December 2025.
- Customer segments
- AI developers, cloud inference buyers, and sovereign AI programs requiring low-latency token generation via API or on-premises.
- Strengths
- LPU architecture delivers sub-second token-generation latency, consistently benchmarking above GPU-based cloud inference on throughput per dollar.
- Watch for
- NVIDIA December 2025 licensing deal absorbed core LPU intellectual property and both co-founders; standalone vendor viability is uncertain.
- Recent moves
- NVIDIA signed non-exclusive LPU architecture licensing deal worth roughly $20 billion, December 2025; founders Jonathan Ross and Sunny Madra joined NVIDIA.
Sources
- www.graphcore.ai — Company mission, products, geographic presence, career opportunities, blog
- techcrunch.com — SoftBank acquisition, deal terms, company history, investor list, Nigel Toon and Simon Knowles roles
- sifted.eu — 2022 revenue ($2.7M) and losses ($204.6M), Microsoft deal collapse, software challenges, employee departures, market positioning
- en.wikipedia.org — Founding date (2016), founders (Simon Knowles and Nigel Toon), funding rounds, employee count (~450 in 2024), Series D valuation
- sifted.eu — Simon Knowles exit August 2025, official statement, revenue trajectory (2023: $4M), acquisition price context
- www.graphcore.ai — £1 billion Bengaluru campus investment, 500 job creation, UK headcount doubling to ~750, SoftBank backing
- www.cnbc.com — SoftBank $457 million additional investment in May 2026, AGI collaboration, Graphcore subsidiary status
- www.glassdoor.com — Employee sentiment and workplace culture (4.7/5 rating, 92% recommendation rate, positive recent reviews on collaboration and innovation)