SoftBank’s Graphcore Picks Bengaluru for £1bn AI Campus, 500 Jobs Incoming

Graphcore, now a wholly owned subsidiary of SoftBank Group, is making a major move in India with the launch of a new AI Engineering Campus in Bengaluru. Here's a quick breakdown of the announcement:

Graphcore's £1bn India Expansion

• Location: Bengaluru, India
• Investment: Up to £1 billion over the next 10 years
• Jobs Created: 500 new semiconductor roles
• Immediate Hiring: First 100 roles already open, spanning:
  • Silicon Logical Design
  • Physical Design
  • Verification
  • Characterization
  • Bring-up

This campus will be central to Graphcore’s efforts in building next-generation AI computing infrastructure, aligning with SoftBank’s broader ambition to lead in Artificial Super Intelligence platforms.

Graphcore: AI Chip Innovator

• Founded: 2016 in Bristol, UK by Nigel Toon and Simon Knowles
• Industry: Semiconductors and AI hardware
• Core Product: Intelligence Processing Unit (IPU) — a novel processor architecture designed specifically for machine learning workloads
• Mission: To enable innovators to build next-generation AI applications and democratize access to machine intelligence
• Ownership: Now a wholly owned subsidiary of SoftBank Group Corp, continuing to operate under the Graphcore name

Graphcore competes with companies like Nvidia in the AI compute space, and its IPU architecture is known for handling entire ML models inside the processor — a departure from traditional GPU-based systems.

AI Accelerator Comparison (2025) 

Feature / Chip	Graphcore IPU (Bow-200)	Nvidia Blackwell B200 GPU	Google TPU v6e (Trillium)	AMD MI350 GPU
Architecture	Massively parallel IPU tiles	GPU with Transformer Engine	Custom ASIC for ML workloads	GPU with unified memory
Memory	900 MB per IPU tile	180 GB HBM3e per GPU	32 GB HBM per chip	128 GB HBM3e
Bandwidth	~1.5 TB/s (system level)	Up to 8 TB/s	1.6 TB/s per chip	~5.2 TB/s
Compute (FP16)	~350 TFLOPS (system level)	4.5 PFLOPS	918 TFLOPS BF16	~2.5 PFLOPS
Compute (INT8)	Not optimized	9 PFLOPS	1.836 PFLOPS	~5 PFLOPS
Scalability	3D wafer-scale IPU pods	DGX B200 clusters	256-chip TPU pods	MI350 clusters
Target Workloads	Sparse ML, graph networks	Transformer-based LLMs	Large-scale ML training	HPC + AI inference
Power Efficiency	High for sparse workloads	Improved over H100	Optimized for datacenter	Competitive with Nvidia
Deployment	Graphcore IPU systems	Nvidia DGX platforms	Google Cloud TPU pods	Enterprise GPU servers