Japan: METI ¥10T AI Strategy / RIKEN FugakuNEXT / Rapidus

Physical hardware: SAKURA’s planned total after Hopper and B200 expansion, described in November 2024. Primary source · 2024-11-12

METI approved up to about ¥50.1 billion for its GPU cloud project in April 2024. Data Center Dynamics reported this subsidy ceiling as approximately US$324M on 19 April 2024. NVIDIA’s November 2024 account described an expansion from 2,000 to nearly 4,000 Hopper GPUs, plus HGX B200 systems, toward approximately 10,800 GPUs in total.

METI approved up to about ¥10.24 billion in April 2024. Data Center Dynamics reported this subsidy ceiling as approximately US$66M on 19 April 2024. NVIDIA reported that November that KDDI was building HGX systems and planning a liquid-cooled data centre with GB200 NVL72 racks.

Physical hardware: Planned GPUs across Highreso’s two AI data centres, described in November 2024. Primary source · 2024-11-12

Highreso and Highreso Kagawa jointly received approval for up to about ¥7.70 billion in April 2024. NVIDIA’s November announcement described a planned 1,600 GPUs across two AI data centres, including H200 systems at Kagawa.

Physical hardware: RUTILEA’s Hopper capacity described in November 2024. Primary source · 2024-11-12

RUTILEA and AI Fukushima jointly received approval for up to about ¥2.56 billion in April 2024. NVIDIA’s November account described an expansion to more than 1,000 Hopper GPUs; AI Fukushima’s Okuma data centre had begun operating that September.

METI approved up to about ¥1.93 billion in April 2024. NVIDIA’s November announcement described GMO GPU Cloud’s H200 systems with Spectrum-X networking and BlueField-3 processors.

Selected for GENIAC in February 2024 as one of seven institutions. Sakana’s announcement describes access to a GPU supercomputing cluster for several months in 2024.

• The selection announcement gives no GPU count for Sakana’s allocation.

Selected in GENIAC’s first three rounds, beginning in February 2024. Its July 2025 project develops an in-vehicle foundation model for autonomous driving, following work on the HERON vision-and-language model.

• The July 2025 selection announcement gives no GPU allocation for Turing.

Released PLaMo-100B-Pretrained in October 2024 following GENIAC’s first round. An August technical report described a separate pre-training experiment with a one-trillion-parameter mixture-of-experts model; the team reported stable training but insufficient evaluation of model performance.

• The October 2024 release gives no GPU allocation for Preferred Elements.

Selected for GENIAC twice in 2024. Its October announcement describes Japanese speech-model training with multilingual audio data and the release of the Kotoba SpeechGen beta; earlier Kotoba-Speech work addressed real-time Japanese–English translation.

• The company’s October announcement gives no GPU allocation.

Selected in October 2024 for a GENIAC dataset study with Macromill, Valright Advisory and Nippon Total Telemarketing. Their February 2025 announcement targeted roughly one million question-and-answer dialogue records from around 30,000 participants, alongside research on data management and access.

• The cited selection concerns a dataset study; no GPU allocation is specified.

Selected in February and October 2024 for GENIAC work on smaller language models. Company tests reported in January 2025 placed its 32-billion-parameter model above GPT-4 on mtbench and elyza-tasks.

• The model announcement gives no GPU allocation for ABEJA.

Released Tanuki-8×8B in August 2024 with GENIAC support. The laboratory first ran a seven-team development competition, then continued with the winning team to train the larger model.

• The release announcement gives no GPU allocation for the laboratory.

Selected for GENIAC in February 2024 and released Stockmark-LLM-100b that May. It trained the model on Japanese business information and designed it to withhold answers when it lacked the required knowledge, with planned retrieval-based integration of companies’ internal information.

• The release announcement gives no GPU allocation for Stockmark.

Project allocation: H100 capacity for a six-month GENIAC project announced in October 2024. Primary source · 2024-10-11

Selected for GENIAC in October 2024. Ubitus announced a six-month project using 256 H100 GPUs to adapt Llama 3.1 405B for Chinese, Japanese and Korean tourism content.

Selected in October 2024 to develop models that structure unstructured documents, with planned use in DX Suite. AI inside’s CTO later reported using Amazon EC2 P5 instances for this research; AWS’s July 2025 account covers support for 12 GENIAC participants.

• The AWS account reports shared programme usage; it gives no GPU count specific to AI inside.

Selected for GENIAC’s third round in July 2025. Rakuten proposed a Japanese-language foundation model with greater memory of prior interactions, intended for personalised AI agents, alongside work on training and inference efficiency.

• The selection announcement gives no GPU allocation for Rakuten.

Selected for GENIAC’s third round in July 2025. The drug-discovery project combines Alivexis’s ModBind simulation technology with model training to predict the biological activity of small molecules.

• The selection announcement gives no GPU allocation for Alivexis.