01 Trigger Event

On July 13, Nikkei Asia, via a relay by 36Kr, reported that Indian companies are increasingly relying on large language models developed by Chinese firms such as DeepSeek, Alibaba, and Moonshot. The key signal came from Puneet Kumar, CEO of Future Capital Venture (India): since mid-2025, several tech startups he has worked with have adopted these Chinese open-weight models and reduced their AI costs by an order of magnitude.

This is not as simple as “some companies are trying Chinese models.”

An order of magnitude is not procurement optimization. It means the budget structure itself is being rewritten.

I have not seen the full invoices, token mix, or latency constraints of those companies, so I cannot mechanically extrapolate that figure to all Indian enterprises. But even if we cut it in half, it is still a strong supply-side signal: once model quality enters the “good enough” zone, what determines adoption is no longer benchmark rankings, but cost per million tokens, deployment freedom, whether private deployment is possible, and whether teams can avoid the pricing power of closed APIs.

Since mid-2025, several tech startups have been using such Chinese open-weight large language models, helping them reduce costs by an order of magnitude.

The most valuable part of that sentence is not “Chinese.” It is “open weights + cost curve.”

02 What This Really Means

What is really happening is that Chinese model vendors are turning their advantage from domestic market share into the global clearing price of AI.

The issue is not whether Indian companies “like” Chinese models. It is that many AI builders simply cannot afford the best closed-model option. Once supply from models like DeepSeek, Qwen, and Moonshot pushes quality above the usable threshold and drives prices low enough, the moat of closed models no longer holds automatically—especially in tasks such as customer service, content processing, code completion, and document extraction.

This is a variant of aggregation theory in AI: if upstream models lack distribution, they will be aggregated by platforms and gateways; but if a class of open-weight models keeps defining the minimum acceptable cost, it is effectively aggregating demand in reverse. Developers do not have to love it. If it is cheap, deployable, and fine-tunable, demand will be pulled toward it.

I may be underestimating the stickiness of brand, security and compliance, and enterprise procurement processes. But for startups, switching cost is often lower than people imagine—especially once the application layer already has model routing, prompt abstraction, and fallback policy in place. At that point, the underlying model becomes a replaceable component.

So this is not a single news item. It is a signal of pricing power moving elsewhere.

Closed-model vendors sell the “strongest model.” Chinese open-weight players sell “system capability that is strong enough and meaningfully cheaper.”

The latter may not win every high-end scenario, but it is more than enough to rewrite the broader market.

03 Historical Analogy / Structural Comparison

This looks more like AWS after 2014 than the iPhone in 2007.

The iPhone was an experience inflection point. AWS was a cost and supply inflection point.

Many companies did not move to the cloud because they “loved cloud.” They did it because cloud made businesses that were previously too expensive, too hard to scale, or too slow to pay back suddenly viable. The logic behind Indian startups adopting Chinese models today is highly similar: not an ideological shift, but unit economics forcing them to redesign their architecture.

An even closer analogy is the wave of API dependence after ChatGPT in 2022. The difference is that back then, developers around the world became dependent on U.S. closed APIs because capability was clearly ahead and substitutes were scarce. Now, India’s growing dependence on Chinese models is driven not by absolute leadership, but by a cost-performance ratio strong enough to punch through budget constraints.

That means the layer of competition is changing.

The first phase was about “who is smartest.”

The second phase is about “who can turn intelligence into a commodity faster.”

The third phase is about “who controls distribution, hosting, compliance, and developer workflow.”

I have not run Indian enterprise procurement cycles internally, so I may be understating the buffering role of localized service providers. But historically, once foundational capability becomes commoditized, the profit pool does not usually remain with the earliest inventor of that capability. It tends to flow to whoever controls channels, hosting, operations, and enterprise relationships.

In other words, open weights are not the endgame.

They are the accelerator pushing the moat away from model IQ and toward infra and distribution.

04 What This Means for AI Builders

For AI builders, what needs to change this week and this month is not sentiment, but default assumptions.

First, revisit base model selection. If your product still defaults to “closed flagship model = the only viable option,” that assumption is probably outdated. Open-weight options should be pulled back into the evaluation set, at minimum using real workloads to compare quality, latency, context cost, KV cache hit rate, fine-tuning difficulty, and hosting cost.

Second, make the application layer model-agnostic as quickly as possible. Prompting, tool calling, RAG, safety policy, and the evaluation pipeline should not be locked to a single vendor. Today the arbitrage may come from DeepSeek or Qwen; tomorrow it may come from someone else. What will actually be priced is switching capability, not loyalty.

Third, the value of API gateways and routing is rising. As the supply side fragments into three layers—top-tier closed models, self-hosted open-weight models, and regional compliance models—builders no longer need “one strongest model.” They need a control plane that can trade off price, performance, geography, and compliance in real time. MCP, agent runtime, observability, and cost governance will matter more than any single benchmark.

Fourth, beware the false savings of “cheaper models causing larger token consumption.” A cost drop of an order of magnitude does not automatically mean the total bill improves. Many teams loosen context limits, increase agent loops, and stack more tool calls on cheaper models, only to see token usage expand instead. I have not seen the usage curve in the reporting, so this point has to remain provisional.

Fifth, if you serve India, the Middle East, or Southeast Asia, you should assume Chinese models are already in the candidate set, not a fringe option. This is not a political judgment. It is procurement reality.

05 Counterarguments / Risks

The strongest counterargument is that this news may overstate the trend and understate the friction.

First, the report cites investor observation, not a large-scale market audit. “Several startups” is still far from “Indian companies broadly shifting over.” I may be reading an early signal as structural migration when it may simply be a short-term cost-cutting move among seed- to Series A-stage companies.

Second, cheap does not automatically mean durable. Once a business moves into large enterprise, finance, healthcare, or government procurement, data sovereignty, auditability, supply-chain security, and geopolitics may all raise switching cost. At that point, the brand, support system, and compliance packaging of closed models become expensive again—but also rational again.

Third, the global expansion of Chinese open weights may not ultimately be monetized by the model vendors themselves. Most of the profit may instead be captured by cloud vendors, hosting platforms, regional channel partners, AI gateways, and local SIs. If so, “model leadership” will quickly slide into “model intermediation.” I would not say that has already happened, but the direction is worth watching.

Fourth, closed-model labs will not sit still. If OpenAI, Anthropic, or Google push further on batch pricing, prompt caching, long context, distillation rights, or private deployment, the cost advantage visible today could compress quickly. The question is not whether Chinese models are strong enough. It is how long their edge can last.

So the safest conclusion is not “India is pivoting toward Chinese models,” but the colder sentence underneath it: global AI competition is shifting from “who builds the smartest model first” to “who defines the world market price of good-enough intelligence first.”

Once someone else defines the price first, latecomers can only recover profit through distribution, compliance, and workflow.

That is the part of this news that should really sting.