Safe Superintelligence Inc. (SSI), the startup co-founded by former OpenAI leader Ilya Sutskever, is breaking from common industry practice by choosing Google’s custom Tensor Processing Units (TPUs) as the backbone for its next-generation Artificial Intelligence model training. This move comes as Google intensifies competition with NVIDIA and AMD, offering substantial compute deals to startups through Google Cloud Platform (GCP).
SSI will access Google’s latest TPU v7p, codenamed Ironwood, which features 4,614 TeraFLOPS of FP8 precision, 192 GB of high-bandwidth HBM memory, and a custom ICI interconnect. These Ironwood TPUs can be deployed in massive pods—up to 9,216 chips per configuration—delivering a formidable 42.5 ExaFLOPS of computing power. The unprecedented scale and speed are designed to significantly reduce model training times, enabling faster research iterations and supporting the rapid evolution of SSI’s ambitious Artificial Intelligence research agenda.
SSI’s declared mission is the development of artificial superintelligence (ASI) with a strong emphasis on safety as a fundamental principle. The company maintains that safety and capabilities will be advanced in parallel and approached as core technical challenges, not afterthoughts. By maintaining a singular focus free from product-driven distractions or time-to-market pressures, SSI intends to ensure that safety standards are always ahead of progress. Their partnership with Google exemplifies the growing trend of aligning high-performance cloud hardware with deep, safety-centric Artificial Intelligence research.
