The Sovereign AI Movement Is a Vendor Swap

In June, the US government forced Anthropic to disable its Fable 5 and Mythos 5 models for all foreign nationals. It was the first live exercise of the diplomatic kill-switch, and it landed during the Pax Silica summit: a 35-nation initiative the US was hosting to secure the AI supply chain ^[1]. The Indian AI startup Sarvam AI had already issued the warning that the cutoff would validate.

For AI users, it is clear that you should not confuse access with ownership, or adoption itself as an advantage. — Sarvam AI

The ban triggered India's sovereign AI push directly ^[2]. But what followed was not a clean break from dependency. It was a redistribution of it, and the pattern now spans continents. Japan's "sovereign cloud," launched by SoftBank and Oracle, runs on Oracle's Alloy platform with more than 200 OCI services. The data stays in-country; the technology stack does not ^[3]. Saudi Arabia signed an open-source AI pact with China, bringing China Mobile, ZTE, and Inspur into its Vision 2030 program ^[4]. India's Sarvam AI built models that outperform Google Gemini 3 Pro on Indian-language documents, but they run on 4,000 NVIDIA H100 GPUs ^[5]. The BRICS+ coalition of Russia, China, India, Iran, and the UAE is developing a Sovereign AI Stack explicitly framed as resistance to "cognitive containment," accusing the US of using safety guardrails to enforce "a shallow regime of artificial homogeneity." The stack runs on Chinese open-weight models ^[6]. NVIDIA and Palantir are selling the sovereign infrastructure itself. In March they launched a joint sovereign AI operating system, and the sovereign AI market is projected to grow from $150 billion in 2025 to $600 billion by 2030 ^[7]. American companies are profiting from the movement to decouple from American cloud dominance. The pattern extends beyond the usual geopolitical blocs. Canada and Germany launched a Sovereign Technology Alliance in February, explicitly aimed at reducing "strategic dependencies on dominant technology platforms from the United States and China" ^[8]. African AI leaders frame sovereign AI as an anti-colonial resource fight.

We’ve seen what happened in oil and gas, we exported raw materials and imported refined products. We can’t make the same mistake with AI. — Olubayo Adekanmbi

France's Mistral AI is building open-weight models as a deliberate third path, describing a "decentralized model where AI is built and served closer to home" to enhance European "digital sovereignty" ^[9]. The US is running its own version of the same play. The Right2Compute Coalition, launched in July, frames compute sovereignty in language that mirrors the BRICS+ rhetoric.

If we do not build the power plants, grids, and data centers of the intelligence age at home, in America and in allied nations like the United Kingdom, our most sensitive data, our IP, and our economic future will be processed abroad, under someone else's laws. — Brittany Kaiser

The US simultaneously rejected UN global AI governance while launching the American AI Exports Program to promote its own version of AI sovereignty: American technology as the gold standard, with local data control ^[10]. The one-way valve is becoming two-way. China is now considering national security laws to restrict overseas access to its own advanced models from Alibaba, ByteDance, and Z.ai. The proposed measures include limiting API access, blocking weight downloads, and designating AI technology leaks as national security offenses ^[11]. The kill-switch the US pulled on Anthropic's foreign users is being replicated in Beijing. And Chinese models carry their own ideological switch already: DeepSeek R1 refuses roughly 85% of politically sensitive prompts under China's 2023 Interim Measures requiring models to uphold "core socialist values" ^[12]. So the sovereign AI movement has produced a global map where every flag of independence is pinned to a vendor's logo. But there is one architecture that actually dissolves the control point, and it is coming from a different direction entirely. On-device AI runs inference on a local chip. There is no cloud service to cut off, no API to revoke, no centralized control point for any government to exercise. Perplexity CEO Aravind Srinivas compared it to the mainframe-to-PC shift.

The biggest threat to a data center is if the intelligence can be packed locally on a chip that’s running on the device and then there’s no need to inference all of it on like one centralized data center. — Aravind Srinivas

Akamai is investing $2.8 billion in edge AI infrastructure targeting medium-sized models and inference at the edge, a commercial bet that pulls in the same direction ^[13]. Power grid bottlenecks in the US are accelerating the migration: Chinese open-weight models combined with local hardware like Nvidia's DGX Spark provide frontier capabilities at a fraction of cloud costs ^[14]. The architecture that truly dissolves the kill-switch also dissolves any government's ability to monitor or regulate AI use. A model running on a phone in a pocket answers to no cloud, no API key, and no jurisdiction. Sovereignty, taken to its logical endpoint, means ungovernability. That is the trade-off the sovereign AI movement has not yet had to face.