Indian Companies Turn to Chinese LLMs to Slash AI Costs
N.R. Finch
Indian companies are adopting DeepSeek, Qwen, and other Chinese large language models at scale, cutting token costs by as much as an order of magnitude — a cost-driven shift that is becoming a key stress test for US AI labs' pricing power.
How big is the price gap?
DeepSeek, offered via Microsoft's Foundry platform in India, charges $0.19–$1.74 per million input tokens and $0.51–$5.40 per million output tokens. Moonshot AI's Kimi tops out at $0.95 input and $4 output.
Compare that to OpenAI's GPT 5.5 family: $5–$12 per million input tokens and $30–$54 for output. This means → for the same workload, a Chinese-model bill can be one-fifth to one-tenth of its US equivalent.
In plain terms = same task, one fewer zero on the invoice — for a cash-burning startup, that is the gap between survival and shutdown.
Who is switching? Not just small startups
Puneet Kumar, CEO of Mirae Asset Venture Investments India, says every consumer-tech startup he has met since mid-2025 runs on Chinese open-weight LLMs.
Schmooze founder Vidya Madhavan's playbook is typical: benchmark with OpenAI and Anthropic first, then match performance using an open-source model plus in-house fine-tuning — at a fraction of the cost.
Trilegal partner Nikhil Narendran puts it bluntly: "Token bills have become a serious problem — they are increasingly unsustainable." Startups adopted first, but large enterprises are now evaluating Chinese-model deployments too.
Why does open-weight carry extra appeal?
Open-weight models — AI models whose parameters are public, downloadable, and locally deployable — offer a key advantage beyond price: data never leaves the country.
This means → for firms with strict data-compliance requirements, especially in finance and healthcare, local deployment sidesteps cross-border data-transfer risk entirely.
This reflects a competitive edge that goes beyond cost: the open architecture itself delivers regulatory convenience in markets with tight data rules.
Not just India — global usage has already flipped
According to AI marketplace Open Router, Chinese LLM usage topped 25 trillion tokens in the last week of June 2025 — more than double the late-May figure and 78% higher than US-model usage.
At the start of the year, Chinese models accounted for less than half of US-model volume. In plain terms = six months to go from "less than half" to "nearly 80% ahead" — a remarkably fast reversal.
Among US firms, Coinbase, DoorDash, and Airbnb have publicly disclosed Chinese-model use. Tesla, Amazon, Uber, and Walmart are actively capping AI call volumes to control spending — the focus has shifted from "use as much as possible" to ROI.
How are US AI labs responding?
OpenAI and Anthropic are ramping up in India — both CEOs attended this year's India AI Summit. Anthropic has signed Air India, Cognizant, and Razorpay; OpenAI has added Tata Consultancy Services (TCS) to its client roster.
But the source material's verdict is blunt: unless token prices come down sharply, marketing spend will have limited effect.
Apple was approved today to use Qwen on devices in China — further evidence of how deeply Chinese open-weight models have penetrated the global commercial ecosystem. This means → pricing pressure on US labs is not India-specific; it is global.
Content is for reference only, not financial advice.