The Stack That Isn't One

In October 2025, the International Criminal Court in The Hague began replacing Microsoft software across 1,800 workstations with German open-source OpenDesk. The trigger was not a BRICS directive but sanctions the Trump administration imposed on ICC officials. The kill switch is not hypothetical: the sanctions showed that a Western government would use its leverage against the court, and the court treated the next step — an American firm cutting cloud access on government order — as a live threat, not a distant one. The institution that flinched was not a member of the rival coalition. It was a court headquartered in the Netherlands, funded substantially by Western democracies, enforcing a statute the United States has never ratified. The catalyst that set the sovereign-stack project in motion was demonstrated first against the West's own architecture. ^[1] That question — if the ICC cannot trust the stack it sits on, who can? — has produced an answer: a multi-layered project to build alternatives to Western AI and cloud infrastructure. But the project is not what its architects describe. It is not a single coordinated architecture making Western infrastructure irrelevant. It is a collection of parallel and sometimes competing stacks, each at a different stage of maturity, and the fractures that limit it run not between the West and its rivals but through the rival coalition itself. The software layer is the most mature, and China leads it. In May, DeepSeek permanently cut V4-Pro prices by 75 percent, making the model 12 to 19 times cheaper than OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7. ^[2] Chinese open-weight models have surpassed American models in downloads on Hugging Face. ^[3] A US-China Economic and Security Review Commission report found that roughly 80 percent of US AI startups now use Chinese open-source base models, and warned that export controls targeting the "digital loop" were "not well suited" to addressing China's "physical loop" of deployment-driven data accumulation. ^[4] The software lead is not accidental. In February, DeepSeek deliberately excluded Nvidia and AMD from the V4 model optimization phase, giving Huawei a multi-week head start to tune software for Chinese processors — a deliberate decoupling of the software layer from Western hardware. ^[5] The institutional layer is being built just as deliberately. In June, China's Global Governance Initiative white paper called for establishing a World AI Cooperation Organization and hosting the 2026 World AI Conference in Shanghai. ^[6] In July, APEC ministers adopted the Chengdu Statement under Chinese leadership, establishing a cooperation framework for open-source digital ecosystems across the Asia-Pacific. ^[7] At the SCO summit in Tianjin, Putin declared that "whoever leads in AI will rule the world." ^[8] The China-Africa partnership launched the Nyansapo Africa Knowledge Q&A System, embedding Chinese-built AI into the Global South's governance infrastructure. ^[9] Brazil and China expanded technology partnerships, with both nations affirming they "share similar positions on upholding multilateralism and improving global governance." ^[10] The architecture is Chinese-led, not collectively governed, and the partners appear to be receiving Chinese-built systems more than co-designing them. At the UN Global Dialogue on AI Governance in Geneva, both India and China advocated for Global South digital sovereignty — China's representative called for the right of countries to "independently choose AI products without being coerced into taking sides" — but the convergence was on principle, not on a shared stack. ^[11] Then there is the hardware layer, the floor beneath everything else, and it is here that the project remains most dependent on the very infrastructure it aims to escape. Huawei plans to double Ascend 910C production to 600,000 units in 2026, with a three-year roadmap through the Ascend 970 in 2028. ^[12] China developed a prototype EUV lithography machine through a state-funded initiative, aiming for entirely domestic advanced chips by 2028 to 2030. ^[13] Chinese chip self-sufficiency has risen from 10 percent in 2021 to 41 percent in 2026. ^[14] Domestic chip startups have launched products with "no foreign technology licences." ^[15] But Huawei's Ascend 910C was built from nearly three million smuggled TSMC dies and stockpiled older-generation Samsung and SK Hynix memory chips. ^[16] The company remains stalled at 7-nanometer fabrication while Nvidia operates at 4-nanometer, with persistent thermal management and yield-loss problems. ^[12] The sovereign stack's hardware floor is, for now, laid on smuggled Western foundations. The deeper problem is not technical. It is that the stacks are parallel, not coordinated — and the fractures run through the coalition that was supposed to make them an alternative order. India is building its own sovereign AI stack, not joining China's. Sarvam AI, funded under the IndiaAI Mission, launched models that outperform Google Gemini and OpenAI on Indian-language tasks, with the stated goal of building "a full-stack sovereign platform." But the platform runs on 4,000 NVIDIA H100 GPUs — the very Western hardware the sovereign-stack project is meant to escape. ^[17] At the UN General Assembly, India's foreign minister Jaishankar positioned India as a "democratic alternative to China's influence" and advocated "diversified, resilient supply chains to reduce dependence on any single market." ^[18] India is not China's partner in the sovereign-stack project. It is its rival. The fractures run deeper still. In March, India, as BRICS chair, refused to condemn US-Israeli strikes on Iran that killed Supreme Leader Khamenei. Modi declared that "India stands with Israel, firmly, with full conviction." ^[19] In May, BRICS foreign ministers failed three times to produce a joint communique, with Iran accusing the UAE of being "an active partner" in the attacks and the UAE reserving "all its sovereign, legal, diplomatic, and military rights." ^[20] The security dimension that underpins the sovereign-stack rationale — the fear of Western infrastructure being weaponized — is the very dimension on which BRICS cannot agree. Even the firms that might populate the sovereign stack are fleeing the binary. AI companies including OpenAI, Anthropic, and Meta's Superintelligence Labs are relocating to Singapore as neutral ground to bypass the US-China rivalry, while China imposed travel bans on the founders of Manus who left. ^[21] The choice is not between the Western stack and the sovereign one. There is a third node, and it is growing. US chips still account for more than 95 percent of global AI computing power, and a black market for smuggled Nvidia chips persists inside China. ^[22] The "irrelevance" of Western infrastructure is an aspiration driving the effort, not an achievement.

if they can’t buy from us, they’ll build their own — and then we’ll have a competitor we wouldn’t have otherwise created. — Jensen Huang

The competitor is being built. But it is not yet independent, and the bloc that would give it weight cannot agree on what it is for.