In May 2026, the US government moved to halt NVIDIA AI chip exports to overseas subsidiaries of Chinese firms based in third countries like Malaysia — not just mainland China (Financial Times). Regulators are shifting from 'border-based' to 'entity-based' controls as Southeast Asian bypass routes surfaced. Japanese companies with supply chains or AI contract services in ASEAN may face additional export-compliance requirements.

Until now, US AI chip export controls primarily targeted entities on mainland China — the H100 ban, then H800, then H20. That framework left Southeast Asian subsidiaries of Chinese-linked firms in a gray zone, and procurement routes through Malaysia and Thailand became de facto workarounds. In May 2026, the Financial Times reported the US moving to close those routes, shifting focus from where goods land to who ultimately controls the receiving entity.

Japanese companies running data centers or AI contract services in ASEAN with Chinese-linked supply chains will likely need to reassess compliance. The logic of 'this isn't going to China' is becoming harder to sustain. For teams already running on purely US/EU-origin infrastructure, this changes nothing today. In the longer run, 'who ultimately controls this entity' will become a standard question in AI procurement — similar to how GDPR turned 'where is your data stored' into a default checklist item.

In May 2026, the G7 agreed to use a common vocabulary distinguishing open-source AI from open-weight AI, pledging to align terminology in each country's policy documents (French government documents / Phoronix). A shared framework for how much of training data, code, and weights are public makes regulation vs. promotion debates clearer. Since openness intersects with EU AI Act GPAI obligations, US export controls, and Japan's soft law, companies now need to know which category their model choices fall under.

For the past two years, the term open-source AI could mean anything from 'we released the weights' to 'we released training data, code, and weights' — the definition varied by organization and country. EU AI Act discussions wanted lighter compliance for open models but could not draw the line without a shared definition. In May 2026, the G7 aligned on common terminology: open-weight (weights released) vs. open-source (training data, code, and weights all released), giving policy documents a shared frame.

Practically, your next vendor-comparison deck or compliance memo can use open-weight and open-source with shared meaning across borders. For teams navigating EU AI Act GPAI obligations, it clarifies where certain model tiers sit. For teams evaluating Chinese open-weight models — already a geopolitical risk layer — this adds a definitional lens on top. Don't expect this to change any regulation overnight; the vocabulary gap that let policy debates talk past each other is now closing.

On May 31, 2026, SoftBank announced plans to build up to 5GW of AI data centers across 3 sites in northern France, with a maximum investment of €75B (approx. ¥14T). A front-loaded commitment of €45B by 2031 makes it the company's largest European AI infrastructure investment. It signals a move to form a European second pole after US Stargate, though the fact that some past large announcements remain unbuilt warrants caution. For Japanese companies, this could expand options for low-latency European training and inference infrastructure.

Until now, European AI infrastructure has been dominated by US hyperscalers: AWS, Azure, Google Cloud. The EU AI Act created a regulatory push toward European-hosted compute for GDPR-sensitive workloads, but sovereign European compute options remained limited. Stargate — OpenAI, Oracle, and SoftBank's US-centric initiative — announced hundreds of billions for US soil. As of May 31, SoftBank extended that logic to Europe: three sites in northern France, up to 5GW, €75B maximum investment by 2031.

For companies wanting low-latency AI compute inside the EU data boundary — GDPR compliance, regulated sectors, government workloads — this expands the medium-term supply picture. It won't change your cloud procurement this quarter; the sites need to be built first. The notable caveat: large commitments have a history of taking longer to materialize than press releases suggest. Worth tracking, not worth betting on yet. For teams already running on US hyperscalers with EU regions, this is a watch-list item, not an action item.

In May 2026, METI and NEDO launched GENIAC-PRIZE 2026, an AI competition with approximately ¥1 billion in prizes and computing resources combined. It focuses on physical AI to substitute human labor in care and logistics, and provides compute to students and researchers without GPUs so they can compete globally. Rather than adding regulation, this is a direct industrial-policy push consistent with Japan's soft-law approach, opening a new funding and compute pipeline alongside existing grants for PoC-stage firms and university spinoffs.

For the past two years, Japanese AI startups and university labs wanting to run serious training experiments faced a wall: cloud GPU costs in Japan are among the world's highest. The ABCI cluster at AIST helped some researchers, but access was limited and grant applications took months — by the time a budget arrived, the PoC window had often closed. METI's AI focus had leaned toward governance (the AI Promotion Bill) more than direct funding push. GENIAC-PRIZE 2026 shifts that: a competition format with prize money and compute bundled together.

For students, researchers, and early-stage startups working on physical AI in Japan — care robots, logistics automation — this is a new compute-and-funding route that didn't exist before. Contest format means the criteria are clearer than grant applications, reducing time on paperwork. The explicit focus on shortage-labor sectors signals where the Japanese government wants domestic AI to deploy first. For large firms that already have GPU access or offshore cloud options, the impact is minimal — this is aimed at teams that have been priced out of experimenting.