Page 11 - The Annual AI Governance Report 2025 Steering the Future of AI

The Annual AI Governance Report 2025: Steering the Future of AI

2025 also saw increased attention to frontier model development outside the traditional
hubs. The release of DeepSeek-V2, China’s highest-performing open-weight model, marks
a significant milestone in the global development of frontier-scale AI systems. DeepSeek’s
capabilities prompted reflection on the minimum levels of compute, engineering expertise,
and energy resources needed to build frontier systems. At the same time, models like Mistral
and Falcon reinforced that open-source leadership is emerging across a range of geographies.

AI systems consume energy across two main stages: (a) Training, which requires high energy
input at a single location over a limited time. (b) Inference, which involves ongoing energy use
as models are deployed and used by millions daily. Current typical chatbot interactions have a
negligible per-use energy footprint (~0.2–0.34 Wh per query): assuming 100 interactions with
a chatbot (corresponding to 10 chats with 10 back-and-forth messages) every single day of a
year, corresponds to roughly 10 kWh for the annual chatbot use, or the equivalent of a 10 km car
drive, or 5 hot showers of 5 minutes, or 2 hot baths. However, future scenarios – where billions
of users each employ multiple AI agents – could dramatically increase global energy demand,
potentially reaching on the order of TWh/day or more. This would represent a significant share of
the world's total electricity consumption (~80 TWh/day, which itself is about 20% of the world’s
total energy consumption). The energy infrastructure for AI is rapidly expanding: Globally, AI
data centers may need 68 GW by 2027 and 327 GW by 2030, comparable to major U.S. states
like California.
The global AI standards landscape is shaped by international Standards Development
Organizations – ITU, ISO, IEC – and practitioner-led bodies like IEEE. Together, they are building
a layered system of standards, spanning technical, managerial, and socio-technical domains.
The AI and Multimedia Authenticity Standards (AMAS) initiative, launched under the World
Standards Cooperation (the partnership of ITU, ISO and IEC), is developing tools and guidance
to support transparency, accountability, and human rights in AI. Its initial deliverables – mapping
the current landscape and identifying standardization gaps – are being published at the AI
for Good Summit 2025, with more to follow. A new AI Standards Exchange Database will be
introduced at the same summit, consolidating standards from ITU, ISO, IEC, IEEE, and IETF.
ITU, ISO, and IEC also launched the International AI Standards Summit series in 2024, with the
second edition planned for Seoul in December 2025, aligning with the goals of the UN Global
Digital Compact. Regionally, CEN-CENELEC’s JTC 21 in Europe is advancing sector-specific
AI standards, notably in healthcare and mobility, supporting implementation of the EU AI Act
and promoting cross-border interoperability through trustworthiness metrics and regulatory
alignment. Nevertheless, adoption of formal AI standards remains the exception rather than the
rule as firms are faced with a patchwork of faster – but unofficial – industry frameworks, resulting
in inconsistency and forum-shopping risks.

The Global AI Summit Series (UK 2023, Seoul 2024, France 2025, India 2026) has evolved
from ad hoc gatherings to structured international cooperation. Nevertheless, AI governance
efforts remain fragmented and politically uneven. While a number of countries have launched
AI safety institutes and initiated risk assessment frameworks, global coordination remains limited
to a handful of Track 2 dialogues – informal, non-government channels who meet to explore
ideas and issues that their governments (the “Track 1” diplomats) may be unable or unwilling
to discuss in public – and regional initiatives. Track 2 diplomacy between the US and China has
grown more substantive. The Ditchley Statement (2023) and the Beijing Dialogue (2024) reflect
expert consensus around compute thresholds, model registration, and dual-use risk evaluation.

2