SoftBank-backed Graphcore Edge IPU Chip Licensing Model

📌 Background

Graphcore became a wholly-owned subsidiary of SoftBank in July 2024. In 2026, SoftBank injected an additional $450 million, integrating the company into the 'Project Izanagi' chip strategy to drive the transition of IPUs from data centers to edge devices. Amidst the intense AI computing competition in 2026, SoftBank is leveraging its Stargate infrastructure and Arm ecosystem to build its own proprietary computing landscape.

👤 Target Customers

Edge device manufacturers, hyperscale data center clients, and AI infrastructure projects within the SoftBank ecosystem.

💰 Revenue Streams

1) Sales of IPU chips and IPU-M2000/IPU-POD systems; 2) Licensing of edge chip IP; 3) Supply of accelerator chips to hyperscale AI infrastructure projects via Stargate deployment pipelines.

🧮 Cost Structure

Chip R&D expenses, wafer foundry and 3D stacked packaging costs, labor costs for the 500-engineer team in Bangalore, India, and maintenance of the Poplar software stack.

🛡️ Moat

The world's first processor-memory architecture IPU and 3D wafer-stacked Bow IPU (65TB/s memory bandwidth, 10x sparse workload throughput compared to GPUs), combined with SoftBank's capital, the Arm ecosystem, and synergies from the Izanagi strategy.

🔑 Keys to Success

• On-time delivery of next-generation chips per the Izanagi roadmap
• Securing stable shipments by binding to hyperscale projects like Stargate
• Scaling the edge IP licensing ecosystem

⚠️ Risks

• Difficulty for third-party accelerator cards to break through in an NVIDIA-dominated market
• Potential strategic shifts or budget cuts by SoftBank
• Edge AI demand falling short of expectations

🏢 Cases

• SoftBank's $450 million capital injection in April-May 2026 to accelerate IPU edge deployment
• Bow IPU achieving 10x the sparse workload throughput of GPU solutions via 3D wafer stacking
• Dell DSS8440 servers utilizing 8 Graphcore C2 IPU cards to deliver 1.6 petaflops of computing power

📊 SWOT Analysis