The United States exerts significant control over the global AI supply chain through a combination of export controls, domestic investment incentives, and strategic alliances aimed at maintaining technological superiority, particularly against China.
The US government directed Anthropic to restrict access to its advanced AI models Claude Fable 5 and Mythos 5 (released around June 9, 2026) due to national security concerns, primarily over a reported jailbreak vulnerability tied to cybersecurity capabilities. On June 12, 2026 (just days after launch), the US government (via the Commerce Department) issued an export control directive citing national security authorities. It required Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign nationals, including those inside or outside the US — even Anthropic’s own foreign-national employees. Anthropic disabled the models for all customers worldwide to comply.
This appears to be the first such export control directive applied to LLM access. It has sparked debate about overreach, impacts on allies/innovation, and precedents for AI governance.
Export Control dominance stems from America’s leadership in advanced semiconductor design (e.g., via companies like Nvidia), critical software tools, and high-performance computing components essential for training frontier AI models. Through the CHIPS and Science Act, the U.S. has invested tens of billions to onshore and ally-shore manufacturing, advanced packaging, and R&D, reducing reliance on vulnerable foreign nodes like Taiwan while bolstering domestic production of leading-edge logic and high-bandwidth memory chips vital for AI.
According to Media reports, US officials learned of a method to bypass/”jailbreak” Fable 5. This involved prompting the model to read a specific codebase and fix software flaws — raising concerns about advanced cyber capabilities that could be misused.
Fable 5 is a safeguarded, publicly available version of a “Mythos-class” model. Mythos 5 is the more powerful/less restricted version (strongest cybersecurity capabilities claimed), initially for vetted partners like US government cyber defenders.
US Commerce Department’s Bureau of Industry and Security has layered increasingly stringent export controls on advanced AI chips, semiconductor manufacturing equipment (such as EUV lithography), and even model weights, using mechanisms like the Foreign Direct Product Rule to restrict access by adversaries and third countries. Recent actions, including directives on specific models like Anthropic’s Fable 5 and Mythos 5, demonstrate how the U.S. extends controls beyond hardware to high-capability AI systems themselves for national security reasons. While these measures create chokepoints that slow competitors, they also involve coordination with allies (e.g., Netherlands, Japan) and carry risks of supply chain fragmentation, innovation trade-offs, and incentives for rivals to pursue self-sufficiency. Overall, this multifaceted approach positions the U.S. as the central gatekeeper of the AI ecosystem, balancing security with the goal of sustained global leadership.
Developers, companies, and governments worldwide now face uncertainty: models can be pulled or restricted for national security reasons, affecting even domestic users through compliance burdens.
In contrast, leading Chinese open-weight models from labs like DeepSeek, Alibaba (Qwen series), Zhipu AI (GLM), Moonshot (Kimi), and others offer downloadable weights under permissive licenses. These can be run locally, fine-tuned, or deployed on private infrastructure without risk of sudden revocation. They have narrowed or closed performance gaps on many benchmarks while being dramatically cheaper (often 5-30x lower cost) and highly efficient, thanks to innovations driven by hardware constraints.
This dynamic accelerated after earlier U.S. chip export controls: Chinese labs optimized for limited compute (e.g., Mixture-of-Experts architectures), released capable open models, and captured growing global market share — reportedly around 30% of AI workloads in some metrics by 2025-2026. Recent events like the Anthropic restrictions have amplified adoption, as seen with immediate releases of competitive models (e.g., GLM-5.2 topping certain benchmarks right after the Fable pull). Chinese models come with their own trade-offs: potential compliance with China’s National Intelligence Law (raising data/security concerns for sensitive applications), built-in censorship on certain topics, and variable safety/guardrail robustness compared to some U.S. systems.
The U.S. strategy of hardware chokepoints and targeted model restrictions has inadvertently boosted a vibrant, diffuse Chinese open-source ecosystem that many now view as a reliable backup. This shifts the AI landscape toward greater pluralism — benefiting those prioritizing availability and independence, while complicating efforts to maintain unified Western technological leads. The trend favors diversification: smart players are hedging with a mix of U.S. closed models (where accessible), open Western alternatives (like Meta’s Llama), and Chinese open weights.
In India, Nandan Nilekani, the architect of Aadhaar and a key voice in India’s digital public infrastructure, has long argued that India should not prioritize building its own frontier AI models. He contends that the country should instead focus on becoming the “use-case capital of the world” by applying existing (mostly Western) models to solve real problems at population scale — leveraging high-quality local data, small/fine-tuned models, synthetic data, and India Stack for diffusion across sectors like healthcare, education, agriculture, and governance. In his view, the massive costs and diminishing returns of chasing ever-larger LLMs should be left to Silicon Valley “big boys,” while India excels at practical deployment and societal impact rather than model leadership.
While this application-first mindset has strengths — especially given India’s strengths in digital infrastructure and pragmatic scaling — it is increasingly shortsighted in light of geopolitical and strategic realities. Relying on foreign frontier models exposes India to sudden access restrictions, as seen in the U.S. government’s directive forcing Anthropic to suspend global availability of Claude Fable 5 and Mythos 5. Even if open-source alternatives exist, cutting-edge capabilities (especially in cybersecurity, defense, or sensitive domains) can be gated by export controls, citizenship checks, or geopolitical shifts. A nation of India’s size and ambitions cannot afford to be downstream on critical general intelligence tools for defense (e.g., 6th-generation fighters, autonomous systems), intelligence, or critical infrastructure.
Chinese open-weight models offer a hedge, but they come with their own compliance, censorship, and security concerns under China’s legal framework — hardly ideal for sovereign applications. Frontier models trained predominantly on English/Western data often fail on India’s linguistic diversity, dialects, cultural nuances, and local contexts (as demonstrated by poor outputs on Indian festivals or regional issues). Building or significantly advancing sovereign foundation models enables better fine-tuning, safety alignment, and relevance for India’s 1.4 billion people across 22+ languages. India has already made progress with efforts like Sarvam and Bhashini; dismissing frontier pursuit slows culturally attuned breakthroughs.
Advanced AI is dual-use. Without domestic frontier R&D, India remains vulnerable in areas like cyber defense, misinformation, and autonomous systems. Moreover, diffusion works best when you can influence or control the base models — something open collaboration and sovereign development enable. India can (and arguably should) pursue a hybrid strategy: aggressive diffusion today via existing tools, while investing in compute infrastructure, talent pipelines, chip design, and targeted frontier efforts for tomorrow.
Focusing only on applications risks locking India into a lower-value services mindset, similar to critiques of traditional IT outsourcing. Building frontier capabilities (or at least competitive foundation models) attracts top talent, spurs ecosystem growth, drives high-value exports, and positions India as an AI power rather than a consumer. DeepSeek’s efficient achievements show that with smart architectures and targeted investment, frontier progress is more accessible than before — weakening the “we can’t afford it” argument over time.
Framing frontier development as unnecessary cedes too much ground on sovereignty, security, and strategic autonomy. In a world of proliferating U.S. export controls and great-power AI competition, India needs both: world-class applications and the ability to build or meaningfully shape frontier models. A purely application-layer strategy risks perpetual dependence; balanced ambition better secures India’s future.
Galactic Views