Sovereign AI is the push by nations and organizations to keep AI under their own control rather than someone else's. It runs on a few hard realizations: you can borrow a model but not the data, the ability to verify frontier AI sits in only a handful of countries, and policy capital is flowing toward building autonomous models on national languages and national data. This hub gathers Pebblous's writing on that terrain.

Read across five axes and the map sharpens. ① National AI autonomy is the attempt to build home-grown models on national languages, data, and policy funding. ② International AI governance is the question of who writes the rules and how — WAIC, the UN, and national bills belong here. ③ Data sovereignty is the design of borrowing a model while keeping the data locked inside. ④ Compute and verification sovereignty covers the concentration of the power to build AI and to verify its output. ⑤ Distillation and autonomy tech asks whether shortcuts like distillation can actually localize a model.

Pebblous cares about this terrain for a simple reason: dig into any of the five axes and you end up back at data. Autonomy comes from data, the trust in governance comes from verifiable data, and sovereignty is decided by control over data. From the view that data quality and provenance are the underlying infrastructure of AI sovereignty, reading these pieces brings the whole map into focus.

Series Guide — Five Axes

① National AI Autonomy — Language, Models, Policy Capital

Nemotron-Personas-Korea — A Starting Point for Korea's AI Autonomy

NVIDIA's 7-million synthetic-persona dataset for Korea, and why synthetic data grounded in national demographics is a starting point for autonomy.

You Can Rent the Model, But Not the Data

The policy mechanism behind the KRW 560B national-fund investment in Upstage, read against the global map of sovereign-AI capital.

The EU Ordered a 400B-Parameter AI. Maltese Data Is 0.03%.

The EU commissioned an open-source AI covering 24 official languages — but the real bottleneck was the scarcity of low-resource language data.

Gemma 4 Deep Report — Apache 2.0 Opens the Door to Sovereign AI

How Apache 2.0 open weights open the door for nations and organizations to build their own autonomous models, with the architecture laid out.

Gemma 4 31B Runs on a 24GB GPU — NVIDIA NVFP4 Quantization

Running a frontier-class model on local hardware through quantization — the practical conditions for autonomous infrastructure.

Korea's AI Action Plan and the Strategic Alignment of Pebblous AADS

The alignment between the national AI action plan's four core areas and the Pebblous AADS program, read through a data-quality lens.

② International AI Governance — Who Writes the Rules

Reading Xi's WAIC 2026 AI Plan by the Numbers

Xi's WAIC 2026 keynote tested against five neutral yardsticks separating declaration from delivery, turned inward as a self-check on Korea.

Only Two Countries Can Verify the World's Most Powerful AI

90% of top AI supercomputers sit in the US and China. The real pressure point is the concentration of verification power, not rule-making.

Evidence Was Laid Down Before the Rules

The UN's 193 members put a science panel's shared evidence base on the table before bargaining over AI rules.

What 193 Countries Did Before Making AI Rules

The same story told in plain terms anyone can follow.

Malaysia Defines Who Owns the Data Instead of Blocking AI

An ASEAN data-ownership approach that protects both training input and AI output as intellectual property.

The US Cut Access to a Three-Day-Old Anthropic Model

How the first AI export control exposed the risk of "model subscription" — the first case of cross-border AI control machinery.

The First US State Law to Put Frontier AI Under Annual External Audit

Illinois SB 315 mandates independent third-party annual safety audits for large frontier AI developers.

③ Data Sovereignty — Borrow the Model, Lock the Data

Borrow the Model, Lock the Data: The Sovereign AI Infrastructure of WWDC 2026

Apple borrows Gemini while keeping data inside the device and Private Cloud Compute — analyzed through a data-sovereignty lens.

6,000 People Deployed Inside the Customer

Microsoft placed AI engineers inside customer walls for $2.5B. Customer data sovereignty is the real stake in the deployment war.

John Deere's $99M Settlement Partly Concedes the Repair War

Who controls manufacturing-equipment data? The data-sovereignty war between makers and users.

④ Compute & Verification Sovereignty — The Power to Build and to Verify

Singapore Certifies AI Trust at the National Level

Budget 2026, the Punggol testbed, and the AI Verify Foundation, connected as policy → infrastructure → attestation.

The Surplus-Server Invoice: AI Overinvestment and a Missed Data Signal

Meta Compute's announcement wiped KRW 569T off the KOSPI in a day — the shock of an unverified compute signal on a national economy.

From an Engine That Wouldn't Run to Drawing a Leaf-Density Map

Reproducing an NVIDIA-first engine on Apple Silicon — a hands-on record of compute autonomy freed from single-vendor lock-in.

⑤ Distillation & Autonomy Tech — Can You Localize a Model by Shortcut?

You Can't Buy Sovereignty Through Distillation

Using the Anthropic–Alibaba distillation dispute as a hook, this piece digs into how distillation free-rides on the diversity and curation of the original data.

All five axes of AI sovereignty converge on data. Pebblous diagnoses and raises the quality of the data AI learns from. See our approach at DataClinic, where data trust and provenance are handled in one place.