SoftBank Deploys World’s Largest NVIDIA Blackwell SuperPOD to Power Japan’s AI Ambitions

On July 22, 2025, SoftBank achieved a milestone in AI infrastructure by deploying a DGX SuperPOD equipped with DGX B200 systems housing over 4,000 NVIDIA Blackwell GPUs, the largest operational SuperPOD of its kind globally. The platform draws on Quantum‑2 InfiniBand networking, supports the NVIDIA AI Enterprise software suite, and joins SoftBank’s broader GPU footprint, now exceeding 10,000 GPUs total with a combined compute capacity of 13.7 exaflops.

The inclusion of NVIDIA Blackwell GPUs significantly boosts compute density and performance-per-watt. On benchmarks, an 8-GPU DGX B200 node achieves over 1,000 tokens per second per user on a 400B‑parameter model like Llama 4 Maverick, and historical DGX systems recorded over 250 TPS per user on even larger DeepSeek‑R1 models, reflecting Blackwell’s impressive inference efficiency. Additionally, AI-focused storage provider DDN confirms that DGX-B200 setups sustain 1 GB/s per GPU, ensuring data throughput keeps pace with compute demands.

SoftBank’s SB Intuitions subsidiary, focused on Japanese-language generative AI, is the initial beneficiary. Having developed 460‑billion-parameter LLMs in FY 2024, SB Intuitions plans to release a 70‑billion-parameter commercial model, “Sarashina mini,” by March 31, 2026. The new compute framework will support multiple parallel training cycles to iterate toward larger and more capable models. The group also unveiled the “Sarashina” brand identity, reinforcing its commitment to culturally-aware AI.

Japan’s Ministry of Economy, Trade and Industry certified the SuperPOD as part of its “Cloud Program” under the Economic Security Promotion Act, designating it as critical social infrastructure. SoftBank now aims to offer compute-as-a-service to third-party AI developers and researchers across Japan—a strategic pivot from internal use only and a push toward building sovereign AI capacity.

Infrastructure Evolution & Technological Context

This deployment represents SoftBank’s third major compute expansion:

1. September 2023: rolled out a DGX SuperPOD with 2,000 NVIDIA Ampere GPUs, delivering 0.7 exaflops.
   
2. October 2024: added 4,000 Hopper GPUs, boosting compute to 4.7 exaflops.
   
3. July 2025: upgraded to Blackwell GPUs, yielding 13.7 exaflops, an almost tripling of compute power in fewer than two years.

At 13.7 exaflops, SoftBank joins the global elite of AI compute providers, integral for training trillion-parameter and multimodal models. Blackwell’s high throughput (e.g., 1K+ TPS per user) accelerates interactive AI use cases and reduces latency.

Certification by METI positions the infrastructure as part of Japan’s national resiliency strategy, reducing dependency on foreign cloud services. Providing access to universities, startups, and enterprises supports broader AI innovation across Japan.

Overall, SoftBank’s computation leap positions it among global leaders in AI infrastructure. By matching next-gen hardware with Japan-specific language models and opening access to external users, SoftBank is operationalizing a national AI strategy – combining power, sovereignty, and ecosystem catalyzation. As SB Intuitions builds toward Sarashina mini and beyond, watch for announcements on commercialization, third-party customers, and model performance metrics in the coming year.