"Sovereign AI" Repeats the Dependency It Promises to End

The parade of sovereign-AI announcements — Palantir pitching Washington, Cohere pitching European governments, Ukraine building for survival, India building after Washington cut off access — reflects a real anxiety: no one wants their intelligence supply controlled by three American companies. But the pattern repeats every time. The sovereignty you buy is itself a vendor's product, and the dependency you were escaping shows up wearing a different logo. Palantir's Alex Karp has the sharpest pitch. He attacks OpenAI and Anthropic for per-token pricing, calls it a "wealth tax," and tells governments that controlling their own model weights is controlling their fate ^[1].

Your AI sovereignty dictates your institution's future. — Palantir

The product behind the pitch is a Palantir software layer running on Nvidia's open-source Nemotron models, sold to US agencies as the freedom to switch between AI systems ^[2]. The buyer leaves OpenAI. The buyer arrives at Palantir and Nvidia. Cohere's April acquisition of Aleph Alpha, the German AI firm, is the same pattern in corporate form — pitched as giving governments "uncompromising control over their AI stack" ^[3].

Organisations globally are demanding uncompromising control over their AI stack. — Coherence

The €500-600 million in financing comes from Schwarz Group, the German retail giant, and the whole thing runs on Schwarz's Stackit cloud ^[3]. A government wanting independence from American AI labs gets a Cohere model, Aleph Alpha's research, and a retailer's infrastructure. The dependency moves from one consortium of vendors to another. Ukraine's case is the starkest because the motive is survival. A country at war is building sovereign AI infrastructure with its telecom partner Kyivstar, whose parent company VEON is headquartered in Amsterdam, on top of Google's open-source Gemma model ^[4]. Even the most security-driven sovereign-AI buyer still requires foreign financing and a foreign model. The Pentagon's sovereign AI is a procurement decision that produced a single-vendor lock. Google's Gemini was integrated into classified military networks in April, while Anthropic was designated a supply-chain risk for refusing similar terms ^[5]. The government chose one vendor and banned another. National-security AI became a single-supplier commitment — the opposite of sovereignty. OpenAI's launch on Amazon's AWS GovCloud in June looked like giving government customers a second cloud option ^[6]. It traded one cloud vendor for another. And Microsoft's legal threat over an exclusivity violation revealed the constraint underneath: the freedom to switch providers is itself governed by the commercial agreements those providers write ^[6].

By making OpenAI capabilities available within familiar AWS environments, organizations can spend less time navigating operational barriers and more time building. — OpenAI

India's sovereign AI push, triggered when US export controls forced Anthropic to disable models for foreign nationals, names the paradox most candidly ^[7].

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. — Anthropic

India's response is a hybrid — a Jabil manufacturing plant (Jabil is US-based), roughly twenty domestic models, 88,000 GPUs allocated to startups ^[8]. The hybrid still embeds foreign manufacturing partnerships and foreign-trained talent. Dependency relocates, again. Every purchase path leads to the same destination. The sovereign stack is always someone's product. The one instrument that breaks vendor lock-in by force is not a purchase. The European Union's Digital Markets Act — a competition law — has produced the only binding orders that force AI platforms open. In June, the Commission ordered Meta to give rival AI assistants free access to WhatsApp's API, the first such interim measure in 17 years ^[9]. In April, it ordered Google to open Android to competing AI assistants through custom wake words and deep-level task access ^[10]. In June, it designated AWS and Azure as cloud gatekeepers, requiring portability and banning self-preferencing ^[11]. And in May, it prepared a record DMA fine against Google — up to 10 percent of global revenue, with a separate proceeding on Gemini's access to Android ^[12]. These are legal orders, not procurement choices. The gatekeepers do not get to sell sovereignty. They are told to open the gate. The distinction matters: every binding action forcing AI platforms open has come through competition law, not through AI-specific regulation. Whatever the EU's AI Act may eventually deliver, the enforceable instruments breaking vendor lock-in right now are antitrust orders. Then there is the one country that genuinely escaped US vendor dependency. China did not buy sovereignty. It built it. Zhipu AI trained its GLM-5.2 model on 100,000 Huawei Ascend 910B processors ^[13]. DeepSeek launched V4 optimized for Huawei chips ^[14]. ByteDance, Tencent, and Alibaba are pivoting to Huawei's Ascend 950 ^[15]. Domestic silicon, domestic models, domestic deployment — a stack with no American company inside it. The trade is straightforward: China replaced US dependency with Huawei dependency, and the whole project is bottlenecked by US export restrictions on the chipmaking tools that make the Ascend line possible. This path is real, but it requires a domestic chip industry and a willingness to trade foreign dependence for domestic concentration. Three paths to AI sovereignty. Purchase, which relocates the dependency — the path dominating the headlines and most aggressively marketed by Palantir, Cohere, and the cloud providers. Antitrust law, which breaks the lock by force but requires a regulator with DMA-level authority and the will to use it. Industrial substitution, which builds a replacement but requires chipmaking ambition and trades one dependency for another. Only the EU's path is available to a country that cannot make its own chips. The sovereign-AI buyers now signing procurement contracts will discover what they bought soon enough. The question is what they reach for next — antitrust law, industrial policy, or simply a different vendor's logo on the same dependency.