China's AI Price War Reshapes Global Competition

China's AI industry has entered a new phase that fundamentally changes the global competitive landscape — not by building smarter models alone, but by making them dramatically cheaper. Chinese AI models now undercut Western equivalents by 5 to 30 times on API pricing while rapidly closing the quality gap, forcing a strategic reckoning for countries like India that depend on foreign AI infrastructure.

According to OpenRouter data, Chinese AI models processed 21.37 trillion tokens during the week ending June 21, 2026, compared with 5.76 trillion tokens for leading US models — a 3.7x volume advantage that signals enterprise adoption shifting decisively toward cost-efficient alternatives. DeepSeek's V4-Pro model costs as little as $0.0036 per million input tokens, compared with OpenAI's GPT-5.5 at roughly $0.50 per million cached input tokens — a staggering 97% discount.

Asian AI startups are already racing to fill the void left by restricted access to Western frontier models, and Chinese firms are leading this charge with aggressive pricing strategies.

Bernstein and Jefferies Sound the Alarm on India's AI Dependency

Two major investment research firms have independently flagged India's strategic vulnerability in foundational AI. Bernstein's report warns that India has "yet to experience its own 'DeepSeek moment'" and still lacks a globally competitive foundational large language model. The firm compares frontier AI models to "fighter jets" — arguing that nations are becoming less willing to freely share their most advanced capabilities.

Jefferies, in a separate report, describes the launch of GLM-5.2 by Hong Kong-listed Z.ai (formerly Zhipu AI) as another "DeepSeek moment" — enterprise-grade performance approaching Anthropic's Claude Opus 4.7 at roughly one-quarter the cost per token. Jefferies notes that as LLMs commoditize, enterprises will prioritize affordability, deployment flexibility, and data security over marginal benchmark improvements.

The stark conclusion from both reports: India could end up operating one or two generations behind global competitors, reducing the competitiveness of its $315 billion software industry despite having one of the world's largest pools of engineering talent.

India's IT industry already generates $10-12 billion in AI services revenue, but this revenue stream is built on applying foreign foundation models — not owning the platform layer itself.

The Structural Roots of India's AI Gap

The disparity between India and China in AI capabilities is not accidental. China spent the past decade building domestic internet giants — Baidu, Alibaba, Tencent, ByteDance — that generated vast proprietary datasets, deep AI research capabilities, and a deep bench of engineering talent. These structural advantages now power China's low-cost AI revolution.

India, by contrast, built its technology ecosystem around IT services and application development — a model that delivers reliable revenue but does not produce the proprietary datasets or research depth needed to train world-class foundation models. This structural deficit, combined with geopolitical constraints on access to cutting-edge US models, leaves India in a precarious position.

The $57 billion that Amazon, Microsoft, and Google have committed to India's AI infrastructure provides the hardware backbone, but without a homegrown foundation model, India remains an importer of AI intelligence rather than a producer.

India's Response: The IndiaAI Mission and Sarvam AI

India is not standing still. The government approved the IndiaAI Mission in 2024 with an outlay of ₹10,300 crore (approximately $1.2 billion), to be implemented over five years. The mission aims to strengthen the country's AI capabilities by supporting foundational models, research, innovation, infrastructure, and skilling.

In a critical first step, Sarvam AI was selected to build India's first sovereign foundational large language model. The Bengaluru-based startup has developed two models — Sarvam-30B and Sarvam-105B — trained on Indian soil using 4,096 NVIDIA H100 GPUs provided under the IndiaAI Mission's compute subsidy programme. Sarvam claims its 105B model outperforms DeepSeek R1 (a 600-billion-parameter Chinese model) on complex reasoning benchmarks while using six times less compute.

Union Minister for Electronics and Information Technology Ashwini Vaishnaw has stated that India's sovereign AI model strategy is "delivering tangible results." Reports suggest the government is considering taking a minority stake in Sarvam AI — potentially converting non-cash compute subsidies into equity — marking the first time the Indian government would hold a direct stake in a private AI enterprise.

The US government's decision to block foreign access to Anthropic's most advanced AI models shattered any remaining illusion that market forces alone would govern AI access. The message is clear: frontier AI is a strategic national asset, not a commodity.

The Road Ahead: Domain-Specific Models Over Replication

Bernstein explicitly recommends that India should not try to replicate OpenAI or DeepSeek head-on. Instead, the strategic play is to build domain-specific foundation models trained on proprietary Indian data in priority sectors: healthcare, manufacturing, financial services, and industrial automation.

This approach leverages India's existing strengths — vast datasets from the world's most populous nation, a thriving digital public infrastructure (UPI, Aadhaar, DigiLocker), and a fast-growing startup ecosystem — rather than attempting to compete on general-purpose intelligence where US and Chinese incumbents have an insurmountable lead.

Building sovereign AI capabilities in these specialised areas could reduce dependence on foreign platforms while creating globally competitive products tailored to Indian languages, culture, and business needs. Sarvam's models already support 22 Indian languages, addressing a market that no foreign AI company can serve effectively.

What's at Stake for India's Technology Sector

The stakes could not be higher. India's technology sector employs nearly 6 million people directly and is projected to reach $315 billion in revenue for FY2026. If India fails to develop sovereign foundation model capability, its $10-12 billion AI services revenue stream — built on applying foreign models to enterprise problems — could face margin compression as Chinese commoditization drives down the cost of AI inference globally.

Furthermore, the national security implications are profound. As AI becomes embedded in critical infrastructure — defence, finance, healthcare, governance — relying on foreign models controlled by hostile or geopolitically unpredictable powers becomes an unacceptable risk. The IndiaAI Mission's emphasis on sovereign AI is not just an economic imperative but a national security one.

FAQ

Q: What is China's AI price war?
A: Chinese AI companies like DeepSeek, Z.ai (Zhipu), and Alibaba are aggressively cutting API prices — DeepSeek V4-Pro costs 97% less than OpenAI's GPT-5.5 — while rapidly improving model quality to near-frontier levels. Chinese models now process 3.7x more tokens than US models on OpenRouter.

Q: Why does India need its own 'DeepSeek moment'?
A: Bernstein and Jefferies warn that without a homegrown foundational large language model, India will remain a consumer of foreign AI technology, operating 1-2 generations behind global competitors. Sovereign AI capability is increasingly treated as a strategic national asset alongside semiconductors and defence technology.

Q: How is India responding to this challenge?
A: The government launched the IndiaAI Mission (₹10,300 crore outlay), selected Sarvam AI to build India's first sovereign foundational model (Sarvam-30B and Sarvam-105B), and is considering a minority stake in the startup. Other firms like Soket AI, Gnani AI, and Gan AI are also developing foundational models.

Q: How does Sarvam's model compare to DeepSeek?
A: Sarvam claims its 105B-parameter model outperforms DeepSeek R1 (600B parameters) on complex reasoning benchmarks while using six times less compute. Both use mixture-of-experts (MoE) architectures. Independent verification is still pending.

Q: What sectors should India focus on for sovereign AI?
A: Bernstein recommends domain-specific models for healthcare, manufacturing, financial services, and industrial automation — areas where India has proprietary data and existing expertise, rather than trying to match US or Chinese general-purpose models.

Sources