Sarvam AI’s Token Factory serves Chinese-developed AI models from Indian GPU infrastructure, a combination that has drawn criticism from named founders and technology commentators. Their objection concerns the gap between operating foreign models locally and developing the indigenous AI capabilities associated with Sarvam’s government-backed role.
Sarvam’s launch post, dated November 16, 2026, describes a service offering its own Sarvam 105B alongside models from DeepSeek, Moonshot AI, Z.ai and Google. By November 18, YourStory had documented criticism from Runanywhere AI CEO Sanchit Monga and technology commentator Shakthi, as well as support from investors and other industry figures.
Sarvam’s response draws an important distinction: it says Token Factory runs on privately procured GPUs, while its earlier subsidized IndiaAI GPU access ended in March 2026. Its technical pitch centers on keeping inference in India.
The dispute turns on what independence should mean: control over where AI runs, ownership of the models themselves, or both.
What Token Factory Actually Serves
Token Factory is an inference platform. It supplies the computing infrastructure that runs trained models and generates responses, allowing developers to use them without buying GPUs or managing the serving software themselves.
Sarvam’s launch post lists the following models behind one OpenAI-compatible API:
| Model developer | Models listed in the launch post |
|---|---|
| Sarvam | Sarvam 105B |
| DeepSeek | DeepSeek V4.1 Flash 1017; DeepSeek V4 Pro |
| Z.ai | GLM 5.3; GLM 5.3 Flash |
| Moonshot AI | Kimi K3.1 Pro |
| Gemma 4 31B |
The catalog matters because Token Factory is neither exclusively a Chinese-model service nor solely a distribution channel for Sarvam’s own research. It combines both businesses: developing models and operating infrastructure for models developed elsewhere.
An OpenAI-compatible API provides familiar request conventions for developers. It does not mean OpenAI operates the service or supplies the underlying models.
Open-weight models also change the deployment arrangement. Their learned parameters can be made available for deployment by another operator, subject to the applicable terms. A provider can therefore run a Chinese-developed model on Indian hardware without sending each inference request to the model developer’s API. The Indian Express explains this distinction as a central part of Token Factory’s commercial rationale.
Training and inference remain separate achievements. Running DeepSeek efficiently demonstrates serving capability; it does not turn DeepSeek into an Indian-developed foundation model. Conversely, offering DeepSeek does not, by itself, show that Sarvam has stopped developing its own models.
Who Objected, and What They Objected To
Sanchit Monga’s criticism focused on the expectations attached to Sarvam’s funding and access to subsidized compute. As reported by YourStory, the Runanywhere AI chief executive wrote:
“I wanted Sarvam to be the lab that ships frontier research out of India.”
His objection was about the company’s research ambition and priorities, not simply whether the platform works.
Shakthi Vadakkepat made a related argument about sovereignty. ETEnterpriseAI reported that he questioned whether hosting predominantly Chinese models matched expectations of a company backed to advance indigenous AI. He also challenged the service’s differentiation from existing inference providers and called for scrutiny of how public funding and GPU subsidies support domestic capability.
Those are identifiable criticisms from particular industry figures. They should not be inflated into a consensus across India’s AI industry.
Sarvam’s own public language helps explain the reaction. YourStory reported that, at the Bengaluru Tech Summit on November 17, co-founder Vivek Raghavan warned about dependence on foreign AI, saying:
“We are exporting data and importing intelligence.”
He was not directly addressing Token Factory. Still, the distinction is consequential: local inference can address the data-export part of that warning while continuing to use intelligence developed abroad. That tension gives the criticism substance, even though it does not establish that hosting foreign models conflicts with Sarvam’s research commitments.
Sarvam’s Response Separates Private GPUs From Public Support
Sarvam told Moneycontrol that Token Factory operates on privately procured GPUs, that its subsidized IndiaAI GPU access ended in March 2026, and that building its own models remains its core mission.
That is the company’s account. The supplied evidence does not independently verify Token Factory’s hardware procurement or funding, and it provides no basis for asserting that the service currently runs on government-subsidized GPUs.
The distinction matters because two different questions have become entangled:
- Current service funding: What hardware and financing support Token Factory?
- Public-policy results: What indigenous capabilities has Sarvam delivered in return for earlier support?
YourStory reports that Sarvam was selected for indigenous model development in 2025 and received subsidized access to 4,096 Nvidia H100 GPUs. That history explains scrutiny of its priorities. It does not demonstrate that every subsequent commercial product uses that subsidy.
My view is that these questions need separate accounting. Private procurement would answer a specific allegation about the service’s present hardware. Evaluating the earlier public support requires evidence about the program’s commitments and delivered research, beyond Token Factory’s catalog.
Supporters also see a business case. Hemant Mohapatra, a Lightspeed India partner and Sarvam investor, argued that international inference providers already earn revenue hosting foreign models in India. According to YourStory, he presented Token Factory as a way to capture that demand amid constraints on computing capacity.
His investment interest is relevant context. The argument itself is straightforward: an Indian company can earn revenue operating models that customers already want to use. Whether that revenue materially supports future indigenous research remains a prospective benefit.
Local Hosting Changes the Data Path, Not the Model’s Origin
Sarvam says Token Factory serves inference from GPU clusters located in India. That is a meaningful deployment choice for customers concerned about where their prompts and responses are processed.
A Chinese-developed model does not automatically require a connection to Chinese-operated inference servers. With locally hosted weights, the model can generate responses on the serving provider’s infrastructure. As The Indian Express notes, this can avoid the need to send data to the original model developer’s servers.
Data residency, however, answers a narrower question than overall security.
Local hosting alone does not settle questions about model behavior, software dependencies or access controls. Nor does a statement about inference location provide a complete account of retention, logging or administrative access. Those require their own documentation and assessment.
For an enterprise buyer, I’d treat model origin and hosting location as separate procurement fields. Both matter, but they describe different risks. A locally hosted foreign model may satisfy a location requirement while still needing task-specific checks for reliability and security.
The Performance Dispute Needs Comparable Evidence
Sarvam also positions Token Factory as an infrastructure engineering product, claiming competitive speed and cost. Its launch post includes detailed vendor measurements, which are more useful than an unsupported claim of leadership but still require attention to conditions.
Frequently Asked Questions
3 questions
1Does Token Factory Serve Only Chinese Models?
No. Sarvam’s launch post lists its own Sarvam 105B and Google’s Gemma 4 31B alongside models from DeepSeek, Moonshot AI and Z.ai.