Graphcore was once one of the most promising AI chip startups, raising nearly $700 million and reaching a $2.8 billion valuation on the promise of its novel IPU (Intelligence Processing Unit) architecture. However, the company failed to achieve commercial traction against NVIDIA’s dominance and was acquired by SoftBank in 2024 at a significant discount to its peak valuation.
Core Products
The IPU was designed with a fundamentally different architecture from GPUs, featuring massive on-chip SRAM and a Bulk Synchronous Parallel computing model optimized for the irregular computation patterns in AI workloads. The Bow IPU was the second generation, and Poplar was the software development kit. The architecture showed promise for certain workloads but struggled with LLM training.
Competitive Position
Graphcore’s IPU showed genuine technical innovation but could not overcome NVIDIA’s ecosystem advantages. The company struggled to attract customers who were deeply invested in CUDA and unwilling to port their software stacks. As the AI industry shifted toward massive LLM training (which favored GPU memory architectures), the IPU’s advantages became less relevant.
Recent Developments
SoftBank acquired Graphcore in 2024 for reportedly around $600 million, a significant writedown from the company’s peak valuation. The acquisition is part of SoftBank’s broader AI strategy. The Graphcore team is now working on next-generation AI processors within SoftBank’s portfolio, potentially powering SoftBank’s planned AI data centers.
Outlook
Graphcore’s story illustrates the extreme difficulty of competing with NVIDIA in AI hardware. Even with $700 million in funding, strong technical talent, and a novel architecture, the company could not break NVIDIA’s ecosystem lock-in. Under SoftBank ownership, the technology may find new life as part of a larger AI infrastructure strategy, but the independent Graphcore story has ended.