How the West, India, and China are building toward the same technology — with two fundamentally different architectures.
We are living through the agentic moment in AI.
Look at any boardroom deck in North America, Europe, or India today and the AI conversation lands in the same place: agents. Agents that schedule meetings. Agents that draft emails. Agents that route support tickets. Agents that close books at month-end. The shift from "ask the model a question" to "ask the model to do something on your behalf" is the most visible commercial breakthrough of 2024–2026.
The numbers reflect the pace. The global AI agent market crossed $7.6 billion in 2025 and is projected to exceed $50 billion by 2030, growing at a 46% CAGR. Roughly 85% of organisations have integrated AI agents into at least one workflow. PwC's May 2025 survey of senior executives found 79% already adopting agents, with 88% planning to expand AI budgets specifically because of them. McKinsey's State of AI 2025 reports a quarter of enterprises actively scaling agents and another two-fifths in pilot stage. NVIDIA's 2026 State of AI Report calls it the year experimentation hardened into deployment.
But beneath the volume, the texture of all this activity is remarkably consistent. The agentic stage of AI, as deployed across the West and most of India, has four defining characteristics:
- It is prompt-mediated. A human articulates a goal. The agent interprets and executes. Even the most autonomous deployments insist on a human-in-the-loop somewhere — 71% of users explicitly prefer it.
- It is enterprise-bounded. Agents live inside business software: Salesforce, Workday, ServiceNow, GitHub. They optimise the seam between a company and its customer, or between a company and its workforce.
- It is application-shaped. Each agent is a discrete product. You can buy it, deploy it, switch it off. It sits on top of existing systems.
- It is productivity-flavoured. The headline metric is throughput — time saved, tickets closed, documents drafted. ROI is measured against last quarter's labour cost.
This is AI as application. A powerful and increasingly indispensable category of application — but an application nonetheless. Invoked. Contained. Governed. Bounded.
Meanwhile, in China
In Hangzhou — the city that until 2016 ranked among the five most congested in China — something architecturally different is unfolding.
That year, the municipal government and Alibaba switched on the first version of "City Brain." By March 2025, Hangzhou had launched City Brain 3.0 — built on the open-source DeepSeek-R1 model — and become one of the first cities anywhere to integrate self-evolving AI directly into urban management. The system detects 92% of traffic incidents automatically, has raised average vehicle speeds by 15%, and coordinates everything from low-altitude drone flight paths to underground utility risk. Hangzhou's congestion ranking fell from 5th to 57th in China.
But City Brain isn't a single product. It is the substrate on which dozens of citizen-facing services now run. Jingxiao'ai, a 24/7 virtual police officer, walks residents through household registration. Hanghaomeng, China's first AI mental-health assistant, has served 1.8 million users for sleep and stress consultations. Yibao'er handles medical insurance inquiries. The same infrastructure now manages traffic, environmental monitoring, public safety, and welfare access — through one orchestrating intelligence.
Hangzhou is not an outlier. After DeepSeek-R1 was open-sourced in January 2025, Chinese cities adopted it for government services at unusual speed. Shenzhen's Longgang District integrated the model into its digital infrastructure within weeks, supporting over 20,000 civil servants. Shenzhen's Futian District went further: it deployed 70 AI "digital employees" handling 240 administrative scenarios — document accuracy above 95%, processing times reduced by up to 90%. Nanjing built an emergency-management platform that generates regulatory-compliant incident reports in around 300 seconds. Linyi, a tier-three city in Shandong, used DeepSeek to build corporate credit profiles that facilitated approximately ¥3.66 billion ($510 million) in loan approvals across more than 14,000 enterprises.
Healthcare is following the same pattern, but on a national scale. On 4 November 2025, China's National Health Commission announced a National AI Healthcare Strategy targeting universal AI-assisted diagnosis in primary care by 2030. Fifty hospitals and 500 township clinics will participate in pilots from April to December 2026. The plan envisions a unified national health database connecting facilities at every level — national, provincial, city, and county — with the explicit goal of bringing specialist-level diagnostic capability to the 600+ million rural Chinese who currently lack it. The five-year investment runs to roughly $2–3 billion. In parallel, DeepSeek-R1 has been deployed locally across tertiary hospitals — Shanghai's Huashan Hospital among the first — and Beijing's Tsinghua AI Agent Hospital opened its Phase II expansion in May 2025 with capacity for 10,000 outpatients daily.
By the time these systems are stitched together — city brain, public administration, hospital networks — what is deployed isn't a constellation of AI applications. It is an AI operating system for the daily lives of citizens.
The Quiet Divergence: App vs OS
The distinction matters more than the comparison.
In the agentic model, AI sits on top of existing systems and between a business and its customer. It is an interface improvement. The customer journey, the regulatory perimeter, the legal entity, the data ownership — all remain the same as before. The AI agent is the new front door, sometimes the new back office, but never the foundation.
In the operating-system model, AI sits beneath services and between citizens and the institutions that serve them. It is an infrastructure replacement. Traffic signals, hospital appointments, building inspections, environmental compliance, welfare access, emergency response — these are no longer discrete products coordinated by humans with software helpers. They are services running atop a continuously inferring layer.
The framing is increasingly visible to Western observers too, though usually only at the enterprise stack layer. Researchers at Cambridge and elsewhere have begun describing AI as "a foundational layer of social, economic, and cognitive infrastructure." Vendors from Red Hat to SAP to VAST Data are racing to define what an "AI Operating System" looks like for enterprise inference. The Astronomer engineering team puts it neatly: just as no modern application runs directly on hardware without an OS, no AI system can run directly on raw data.
What the West has only recently begun articulating at the enterprise level, China has been quietly executing at the civic level for nearly a decade. The architectural gap is not technological — both stacks share the same model families, the same chips, the same conceptual primitives. The gap is philosophical.
Western and Indian AI deployment optimises for commercial intermediation: the AI sits between firm and customer, mediates a transaction or interaction, and earns its keep there. Chinese AI deployment optimises for civic orchestration: the AI sits between citizen and state, coordinates the delivery of a public service, and is funded by public budget.
Both are powerful. Both are valid responses to the political economies they inhabit. But they are not the same thing — and the longer the divergence runs, the harder it becomes to translate between them.
The Data Sovereignty Thread
There is one structural reason why the AI-as-OS model is easier to build in China than in India or most Western democracies. It is also why the divergence is unlikely to close on its own.
OS-level AI cannot exist without OS-level data.
For Hangzhou's City Brain to coordinate traffic, hospitals, environmental monitoring, and welfare, it must access data flows that span agencies, jurisdictions, and individual interactions. For the National AI Healthcare Strategy to deliver AI diagnostics to 600 million rural patients, it requires a unified clinical record across roughly 25,000 township and village clinics. For Shenzhen Futian's digital employees to operate across 240 administrative scenarios, they need a shared substrate of citizen, business, and regulatory data.
China's Personal Information Protection Law (PIPL), in force since 2021, regulates personal data — but within a framework that explicitly supports state-coordinated data aggregation and forbids domestic biodata from being exported abroad. The resulting regime is designed to enable national-scale data pooling under state oversight.
India is building toward something fundamentally different. Under the Digital Personal Data Protection Act, 2023, and the DPDP Rules notified on 13 November 2025, every personal data flow must be tied to a specific purpose, consented to by the individual, retained only as long as necessary, and erasable on request. The Consent Manager framework — activating on 13 November 2026 — institutionalises a market for managing individual permissions. The full enforcement regime arrives on 13 May 2027.
This is not an inferior architecture. It is a different architecture. India is choosing to make data sovereignty a property of the individual citizen, not the state. The trade-off is built in: it becomes harder to pool data across agencies, harder to train monolithic civic AI systems, harder to deliver AI-coordinated public services in the seamless way Hangzhou demonstrates.
What this means for Indian institutions — whether wealth managers, hospitals, transport authorities, or city governments — is that the AI-as-OS architecture, if it emerges here at all, will look fundamentally different. It will require:
- Federated inference rather than centralised data lakes
- Consent-as-infrastructure: granular, programmable, withdrawable, auditable, and embedded at the OS layer rather than bolted on at the app layer
- Purpose-bound model training, not purpose-flexible scraping
- Data fiduciary obligations baked into every layer of the stack — discovery, retention, processing, sharing
- Cross-agency interoperability achieved through open standards and consent flows, not common databases
In other words: if China is building an AI operating system on a unified data backbone, India will need to build the same conceptual stack on a consent-mediated data fabric. That is the real engineering challenge of the next decade.
The Observable Truth
It is tempting to read the divergence as a contest. It isn't. It is two civilisations building infrastructure that reflects their politics.
China is optimising AI for state-level coordination, public service delivery, and economic productivity at population scale. India is optimising AI for individual agency, consent, and data principal rights. The West is currently doing neither cleanly — it has the consent framework (GDPR), but is deploying AI mostly inside private enterprise rather than as civic infrastructure.
Whether any one model is more humane, more productive, more sustainable, or more democratic is a question every society will answer differently — and probably differently again a decade from now. What is no longer in question is the framing: the era of AI-as-application is ending. The era of AI-as-operating-system is being built. The only open questions are who builds it, beneath whose data, and on whose terms.
For Indian regulators, technology firms, and large data fiduciaries, the practical implication is that the choices made over the next eighteen months — about consent infrastructure, about data discovery and minimisation, about federated architectures, about the contractual perimeter around processors — won't just determine DPDPA compliance. They will determine what kind of AI operating system India ends up with.
A society that makes data the property of the individual will, in time, build a different operating system than one that makes data the property of the state. Both will exist. Both will work. Neither will be the same.