ITU roundtable links AI infrastructure access to standards and digital trust

Who took part and what was announced

The roundtable, titled "AI foundation: Digital trust and AI infrastructure for all", was held on the opening day of the summit. According to the ITU, it brought together more than fifty senior representatives from government, industry, standards bodies and academia. They included eighteen ministers and senior government representatives, more than twenty-four private-sector companies and twelve international organizations and academic institutions.

The session ran under the Chatham House Rule, so ideas could be reported without naming speakers. In public opening remarks, Seizo Onoe, Director of the ITU Telecommunication Standardization Bureau (TSB), said innovation should serve the public interest and that every country should be able to benefit from AI, develop domestic innovation and build solutions for local needs. He also announced an ITU standards initiative focused on identity and trust marks for AI. Bilel Jamoussi, ITU TSB Deputy Director, moderated the talks on three themes: AI infrastructure and inclusion, AI standards for industry, and agentic AI and digital trust.

Concentrated computing and the sovereignty debate

The first discussion asked who builds, owns and benefits from AI infrastructure. Participants cited figures showing how concentrated computing capacity is: only sixteen countries have large-scale computing infrastructure, while nearly two billion people remain offline. High-income countries account for 86% of the world's top 500 high-performance computing systems and 97% of their capacity. Developing economies outside China represent half of global internet users but less than 10% of global data center capacity.

Energy was named as a central constraint. Participants cited International Energy Agency estimates that data center electricity consumption could reach around 945 terawatt-hours by 2030, approximately double current levels. In Africa, where roughly 90% of data is stored outside the continent, electricity can be a greater limitation than connectivity when deciding where infrastructure can be hosted.

Views on AI sovereignty differed. Several participants cautioned against defining it as national ownership of every part of the AI stack, which could increase fragmentation and investment costs. Others argued that countries need control over selected layers of infrastructure, data or computing capacity. Regional computing pools, cloud-first policies, shared infrastructure and data embassy frameworks were put forward as possible approaches. The moderator summarized that countries want greater AI sovereignty, but not at the cost of international cooperation.

Barriers to participation included limited access to devices, computing capacity, advanced hardware and locally representative language models. Examples shared included investments in AI factories, gigafactories, supercomputing centers and cloud infrastructure, as well as national AI laws and subsidized access to computing resources.

Aligning horizontal and sector-specific standards

The second theme looked at fragmentation between horizontal AI standards, which set requirements across industries, and sector-specific standards for areas such as health care, telecommunications, medical devices and transportation. Participants agreed both are needed. The concern was that developing them separately can create overlapping requirements, inconsistent regulatory expectations, duplicated assessments and barriers to trade.

Proposed work included mapping horizontal standards against high-impact sector use cases, improving coordination between AI specialists and sector experts, and developing shared guidance and conformity assessment pathways. These steps could help regulators, manufacturers and users see which requirements apply and how to demonstrate compliance. The work builds on the World Standards Cooperation of ISO, IEC and ITU, which has worked with the wider United Nations system on sector-specific AI standards for nine years, including in health care, education and agriculture.

Trust requirements for AI agents

The final theme addressed AI agents acting across platforms, organizations and jurisdictions. Participants noted that many existing regulatory frameworks assume a system whose intended purpose and reasonably foreseeable behavior can be defined in advance. Agents may instead negotiate, carry out transactions, coordinate with other agents and pursue goals across several systems, which raises questions about oversight and accountability.

• Agents need verifiable identities and ways to show they act for a particular person or institution.
• Their authority and permissions should be clearly defined, machine-readable and capable of being limited, suspended or revoked.
• Portable consent, audit trails, provenance and verification should allow an agent's actions and consequences to be reconstructed.
• Accountability should stay with a human or institutional principal rather than shift to the AI system.

Participants said model developers alone cannot establish these mechanisms when agents work across vendors, sectors and national systems, and stressed interoperable standards. They also raised the need to distinguish agents operating inside an organization from those acting on the open internet.

What it means

According to the ITU, the roundtable found convergence on interoperability, shared infrastructure, clear accountability and international cooperation. It also recorded continuing differences over sovereignty, localization and regulation. The announced identity and trust marks initiative signals that ITU intends to carry the agent trust discussion into formal standards work.