Every major government is making a choice about AI right now. The choices look nothing alike, and the differences matter more than the benchmarks.
The coverage tends to frame this as a race: who has the biggest model, who spends the most, who is ahead on some index. That framing misses the more durable question: what does each government's approach mean for businesses and consumers operating under it? The answer depends less on capability than on strategy, and the strategies are genuinely divergent.
China: Open Everything
China's bet is the most strategically coherent of the four. Most Chinese labs publish model weights and charge far less than Western competitors. A March 2026 report from the US-China Economic and Security Review Commission describes this as deliberate: open collaboration where frontier labs refine each other's base models, enterprises adapt them for niche applications, and the resulting deployment data feeds back into capability improvements.
The government has positioned China as the global provider of open AI development, designed to capture developer mindshare, establish Chinese models as industry standards, and create dependency relationships over time. Z.ai's Beijing laboratory released its GLM-5.2 model under an MIT licence in June 2026, making it fully free, modifiable, and accessible to any developer anywhere.
On infrastructure, provinces including Gansu, Guizhou, and Inner Mongolia have slashed data centre power bills by up to 50%. The National Development and Reform Commission is drafting a five-year AI data-centre plan reported at approximately $295 billion, with a mandate that 80% of chips come from domestic suppliers.
What this predicts: China's open-weight models are already in production globally. DeepSeek and Qwen are self-hostable by any developer without permission from any US entity. As those models improve, developers outside China will increasingly build on Chinese foundations, not for ideological reasons but because capable, free models are hard to argue against. The dependency relationships China is building are real and compounding.
What it costs: China's open-source strategy is partly a workaround for US export controls on advanced chips. Open models reduce the compute required for effective deployment, which matters when you cannot easily access the most advanced training hardware. The strategy is adaptive, but the hardware constraint is not fully solved by it.
Effect on other countries: Every time a US lab gates a new frontier model, it nudges global developers one step closer to open-weight alternatives that China already dominates. Washington can restrict what US companies publish. It cannot unpublish what Chinese labs have already released.
Effect on consumers: Access to capable AI tools at near-zero cost is increasing globally. A small business in any country can now run a competitive AI workflow on Chinese open-weight models hosted locally, with no ongoing API costs. The practical floor for AI capability is rising fast, and China is setting it.
Source: US-China Economic and Security Review Commission, "Two Loops: How China's Open AI Strategy Reinforces Its Industrial Dominance," March 2026. uscc.gov. Also: "China Expands Its Open-Source AI Strategy," cointribune.com, June 2026.
India: Build the Infrastructure
India's approach is the most interventionist outside of China. The IndiaAI Mission, a $1.25 billion programme scaling through 2026, covers subsidised GPU compute, indigenous foundation model development, a national datasets platform, startup funding, and AI skilling. Pilot compute infrastructure nodes went live in Bengaluru, Hyderabad, and Pune in early 2026. The government has procured more than 38,000 subsidised GPUs.
The education signal is the most telling one for the long term. India is rolling out an AI curriculum in all schools starting from Grade 3, beginning with the 2026-27 academic year. The stated goal: build the largest AI-literate workforce in the world. The governance approach is deliberately light-touch. India published AI Ethics and Risk Mitigation Guidelines in February 2026, influenced by the EU AI Act but designed not to slow startup growth.
What this predicts: India is building a workforce advantage that will take a decade to fully materialise but is already underway. A generation of workers trained in AI from primary school will enter the labour market in the early 2030s. Combined with India's existing technology sector and cost structure, that is a significant competitive position for any industry that can be delivered digitally.
What it costs: The infrastructure programme is government-funded and government-directed. Indigenous foundation model development may produce capable models, or it may produce models that require further subsidy to remain competitive with what private labs in the US and China produce without government mandates.
Effect on other countries: India's AI workforce strategy puts pressure on countries leaving AI education to individual institutions. The skills gap being created by institutional incoherence in Western universities is the same gap India is systematically filling.
Effect on consumers: Indian consumers and businesses will have subsidised access to compute and datasets that most other countries do not. For Indian startups building AI products, the barrier to entry is lower in 2026 than it was in 2024.
Source: IndiaAI Mission, Ministry of Electronics and Information Technology. indiaai.gov.in. Also: "India's AI Policy 2026: GPU Procurement, Data Sovereignty, and Startup Support," valueaddvc.com, June 2026.
The EU: Write the Rules First
The EU's bet is on governance as a competitive advantage. The argument is that clear rules create trust, and trust creates adoption. The EU AI Act, the most comprehensive AI regulatory framework in the world, is being phased in through 2027. The EU is not building national compute infrastructure or open-sourcing models. AI capability development in Europe is happening at the private sector level, with companies like Mistral in France and Aleph Alpha in Germany.
What this predicts: The EU has made this bet before with data privacy, and it worked. GDPR became the effective global standard for data regulation because companies serving EU customers had to comply, and compliance at that scale became the path of least resistance for companies serving everyone else too. If the EU AI Act follows the same path, the countries that wrote the rules will have shaped the global framework. Article 50's transparency obligations, in force from August 2, are already being adopted voluntarily by major platforms worldwide.
What it costs: Compliance overhead is real. A small business in the EU faces disclosure requirements, documentation obligations, and enforcement risk that a comparable business in the US does not. Whether that overhead is justified by the trust dividend it produces is a question that will take several years to answer.
Effect on other countries: EU regulations routinely become the effective global standard for companies that want to operate across borders. A US business with European customers is already subject to Article 50. A US platform with European users is building compliance into its product regardless of what Washington requires.
Effect on consumers: EU consumers are getting clearer disclosure about when they are interacting with AI, which content is AI-generated, and what their rights are. Whether they use those disclosures to make meaningfully different choices is the open question.
Source: EU AI Act, Regulation (EU) 2024/1689. artificialintelligenceact.eu.
The US: Get Out of the Way
The US approach since January 2025 has been to remove constraints. President Trump revoked President Biden's 2023 AI safety order on his first day back in office. The December 2025 executive order sought to preempt conflicting state AI laws. The March 2026 National Policy Framework urged Congress to adopt a light-touch regulatory approach. The June 2026 executive order focused on cybersecurity and a voluntary framework for frontier model developers.
The US position is that private sector investment and market competition produce better outcomes than government programmes. That position has a reasonable argument behind it: US private sector AI investment dwarfs every other country, and the frontier model labs are all American.
What this predicts: The US approach works as long as the frontier matters more than the floor. If the most capable models continue to be American and closed, the US maintains a strategic advantage. If open-weight Chinese models continue to close the capability gap, the advantage erodes. The US cannot gate open weights that China has already published. The June 2026 executive order gated two Anthropic models from export. China responded within weeks with additional open-weight releases.
What it costs: The absence of a national AI infrastructure programme means US small businesses and startups compete without the subsidised compute that Indian businesses have access to and without the open-weight models that Chinese developers freely use. The US market is competitive enough to compensate for that, but the asymmetry is real.
Effect on other countries: US export controls on chips are the most consequential policy instrument deployed. They have accelerated China's domestic chip development and pushed global developers toward open-weight alternatives. The intended effect, slowing China's AI capability development, is real but partial. The unintended effect, accelerating open-source adoption globally, is also real.
Effect on consumers: US consumers have access to the most capable closed models and the same open-weight models as everyone else. Regulatory protection is minimal compared to EU consumers. The tradeoff is more choice and faster innovation, with less disclosure and fewer rights.
Source: Executive Order 14365, December 11, 2025. Executive Order 14409, June 2, 2026. "New Executive Order Signals Evolving Federal Approach to AI," Lathrop GPM, June 2026. Also: "Does US AI Gatekeeping Hand China the Open-Source Edge?" digitalapplied.com, June 2026.
The Bottom Line
Four governments, four different bets. China is betting on open access and infrastructure. India is betting on workforce. The EU is betting on governance. The US is betting on private sector competition.
None of them is obviously right. All of them are making tradeoffs that will take years to fully assess. The businesses that will navigate this best are the ones paying attention to all four, not because they operate in all four jurisdictions, but because tools, talent, regulations, and competitive dynamics cross borders regardless of where a company is incorporated.
Where Agent Micho Fits
For a small business, the government strategy question shows up in three practical places.
First, the tools you use. Chinese open-weight models are now a legitimate option for workflows that do not require the most capable closed models. Document intake, data categorisation, content drafting: a self-hosted open-weight model may cost you nothing beyond compute after the initial setup. We help clients evaluate which workflows benefit from closed frontier models and which run just as well on open alternatives.
Second, the compliance environment. If any of your customers are in the EU, Article 50 is already your problem. Building disclosure and labelling into your AI workflows now rather than retrofitting them later is the lower-cost path. We build that into every client-facing automation we deploy.
Third, the talent gap. Every country is producing graduates with different AI competencies. The businesses that outrun this gap are the ones that do not depend entirely on individual employees to figure out AI on their own. A well-designed workflow standardises good practices and makes them repeatable regardless of who is running them.