Saturday, 21 November 2026LatestVideosAboutContact
Latest
ISL 2026-27: Messi Bouli salvages draw for Punjab FC against Chennaiyin FCMissing for over a month, Hyderabad teenager found in Bidar'My effort is always to keep learning something new every day' - Manav Suthar reflects on brilliant start to his international careerOld City roads remain damaged despite GHMC’s temporary repairsBihar Gets Development Projects Worth ₹34,000 Crore; CM Samrat Choudhary Welcomes PM Modi, Highlights 'Nitish Setu' InaugurationMajor fire guts hotel and paint shop in Hyderabad’s HafeezpetPalghar ACB Arrests MSEDCL Assistant Engineer In Nalasopara For Allegedly Demanding ₹10,000 Bribe To Approve Electricity ConnectionsUttar Pradesh Police Mistakenly Give Guard Of Honour To AIMIM Chief Asaduddin Owaisi In Saharanpur, Probe Ordered | VideoScholarships This WeekHyderabad: Catering worker held for collecting debit cards in cyber fraudDr Durga Rao built extensive agent network for illegal egg harvesting racket: Malkajgiri policeInside BJP’s worst Zila Parishad loss in Rajasthan in three decades
StartupUpdated 3 hrs ago

Sarvam’s Token Factory Raises India’s AI Inference Stakes

Image: Inc42

Sarvam AI’s latest announcement of its Token Factory has brought a new layer of India’s artificial intelligence (AI) infrastructure ecosystem into focus. The Bengaluru-based startup on November 16 launched a commercial platform that allows developers and enterprises to access multiple AI models through its computing infrastructure.

The launch comes as Indian companies look to capitalise on the growing demand for AI inference, the process through which trained models generate responses to user prompts. While much of the country’s AI infrastructure investments have so far gone into setting up data centres and procuring GPUs, the growing adoption of AI applications is creating demand for infrastructure that can efficiently serve models at scale.

A token factory allows developers and enterprises to access multiple AI models through a common API, without having to procure GPUs, deploy models or manage the underlying infrastructure. It handles model hosting, computing resources and inference optimisation, charging customers for the input and output tokens they consume rather than for GPU capacity or server time.

Unlike GPU-as-a-service providers and neoclouds, which primarily offer the infrastructure on which customers run their own workloads, a token factory sits a layer above, using owned or leased GPU capacity to serve inference requests through APIs. This reduces the operational burden on customers and allows them to switch between models based on their requirements.

Sarvam is not alone in this pursuit. Larsen & Toubro (L&T), E2E Networks through its Jarvislabs business, and AI security startup Mirror Security are among the other players building offerings around token-based inference. The emerging category also has global precedents, with companies such as Together AI, Fireworks AI, DeepInfra, and Nebius operating similar platforms.

Sarvam’s Token Factory and The Sovereignty Debate

Sarvam’s Token Factory hosts the company’s proprietary Sarvam 105B model alongside open-weight models such as DeepSeek V4.1 Flash, Kimi K3.1 Pro, GLM 5.3 and Google’s Gemma 4 31B. 

The billing is in Indian rupee. Sarvam has priced its Token Factory on a pay-per-token basis, with separate rates for input, cached input, and output tokens. Its flagship Sarvam 105B model is priced at an introductory ₹10 per Mn input tokens and ₹25 per Mn output tokens. For other models, prices range from ₹14 (per million input tokens) to ₹950 (per Mn output tokens). 

The platform has been serving more than 450 Bn tokens daily through DeepSeek V4.1 Flash, highlighting the scale of demand for inference services, the company said in its blog. 

However, the launch of the token factory has also brought up two questions for Sarvam from the ecosystem: why a beneficiary of the government’s IndiaAI Mission, which is aimed at developing indigenous AI capabilities, is hosting Chinese-developed models, and whether GPUs subsidised under the mission are powering the service.

The company was selected under the mission to develop homegrown foundation models and received access to 4,096 Nvidia H100 GPUs. The six-month computing allocation was worth around ₹247 Cr, including a government subsidy of ₹98.68 Cr. 

Sarvam, however, clarified that its Token Factory does not use GPU capacity subsidised under the IndiaAI Mission. Instead, it runs on GPUs privately procured by the company. It also said that its access to the government-supported GPU capacity ended in March 2026. The company added that there were no commercial arrangements involving licensing fees, data sharing or royalties with the developers of the hosted models. These are open-weight models that can be downloaded and deployed on privately owned infrastructure.

The company has not responded to Inc42’s questionnaire around this. 

“Token Factory gives Sarvam a reason to keep its B200 cluster busy between training runs, and income that lets the company breathe while its trillion-parameter model is built,” said Sanchit Vir Gogia, chief analyst of Greyhound Research, told Inc42. “Sarvam’s product team has reason to feel confident and exposed at once. Its founders should show by the end of 2027 that customers who arrived for DeepSeek stayed for Sarvam.”

The Data Question And The Economic Model 

Running a token factory on its own GPUs also means Sarvam carries the cost of idle capacity. On the serverless tier, customers pay only for the tokens they use, while Sarvam pays for its B200 cluster whether it is busy or not. That makes utilisation, rather than headline volume, the number that decides whether the service makes money.

The company added that 95.7% of the prompt tokens it serves on DeepSeek V4.1 Flash come from cache, which it bills at ₹0.25 per Mn tokens against ₹14 per Mn for fresh input. At those rates, 450 Bn prompt tokens a day would bring in about ₹13.8 Cr a year, before output tokens are counted. Those rates are introductory and run until the end of 2026. Sarvam has not disclosed how many GPUs it runs, the utilisation or how much of the traffic is billed at full price. These finer points will decide whether this is a revenue play or a data play. 

Beyond the sovereignty debate, Sarvam offers zero data retention (ZDR) on Token Factory only on request.

ZDR, in simple terms, means not storing any data intentionally after it has served the immediate purpose. Enterprise users of the Sarvam token factory will have to explicitly request ZDR. “Under an agreed ZDR arrangement, prompts and responses are not retained. Operational logs are scoped separately in the terms for your deployment,” the company said in its blog. 

Sarvam will by default use user data to train or fine-tune its models, unless the feature is disabled by the end user. This could be a concern for businesses processing sensitive or proprietary information through the platform.

For Sarvam, the case for Token Factory rests on what global providers cannot easily offer. 

Prompts run on GPUs in India and are not forwarded to overseas servers, which matters to enterprises and government bodies with data-residency requirements. 

Plus, billing is in INR, the platform carries ISO 27001 and SOC 2 Type II security certifications, and Sarvam promises 99.95% uptime under contract for enterprise deployments. Token Factory also shares credits and API keys with Sarvam’s speech and translation services, and puts its own models in front of developers alongside the ones they already use.

However, the risks are just as visible. The open-weight models Sarvam hosts are available from other Indian providers too. E2E Networks’ Blaze serves DeepSeek, Qwen, GLM and Gemma, as well as Sarvam’s own models, and Jarvislabs offers a similar catalogue, which leaves price and speed as the main ways to compete. 

The India-hosting advantage is also contested, with OpenAI moving to bring inference to India. And the only usage figure Sarvam has shared is for DeepSeek V4.1 Flash, which will test whether customers who came for DeepSeek stay for Sarvam’s own models.

India’s Impending Token Factory Boom 

It’s not just Sarvam that has looked to fill the gap. Competition is growing rapidly with anyone who had a hand in the data centre business now jumping on to the inference bandwagon. 

“This is a positive development for the entire ecosystem, indicating to enterprises that operating the models in India is feasible. There is scope for several participants in this market, as the demand for tokens will outgrow any single entity’s capability,” Manoj Dhanda, founder and CEO of Utho Cloud, another player looking to enter the domestic inference space by early 2027. 

Infrastructure giant L&T is building out its AI infrastructure business through Vyoma.AI, with its AI Factory offering spanning data centres, GPU infrastructure, cloud services and model deployment.

Its AI-centric data centres includes a Token Factory designed to provide high-throughput inference, alongside a model garden, AI security and agentic AI services. The company is also offering GPU-as-a-service, managed inference and model customisation.

Similarly, E2E Networks has expanded its AI infrastructure capabilities through the acquisition of identified assets of Jarvislabs, including its intellectual property, hardware, expertise and customer base in August 2025.

Tamil Nadu-based Jarvislabs already offers a Token Factory through which developers can access open-weight models through a common API and pay based on token consumption. Its catalogue includes models from DeepSeek, Gemma, GLM, Nemotron, and Qwen.

“It’s a 2-month-old product, and we have served close to 50B tokens in November and are growing fast. We are seeing customers using it for coding, transcription services, and conversations,” said Vishnu Subramaniam, founder at Jarvislabs. 

To be sure, the company was already serving models through serverless products for more than a year; Subramaniam called the token factory an evolution in that direction. 

Further, more recently, E2E launched Blaze, a pay-per-token inference service that provides access to open models, including Sarvam, DeepSeek, Qwen, GLM and Gemma, through a single OpenAI-compatible API. The service runs on E2E’s Nvidia GPU infrastructure in India, with the company managing model hosting, scaling and optimisation.

Noida-based Utho Cloud plans to deploy 10,000 GPUs between now and 2028. 

“Our infrastructure runs in Tier III and Tier IV data centres inside India, and it is ISO certified,” the company’s CEO said, adding that Utho’s upcoming inference platform will also enable enterprises to comply with norms regarding data residency and transparency on how data is processed, where it travels, and how much it costs per request.

Mirror Security, a cybersecurity firm focused on homomorphic encryption, on the other hand, is approaching the category from a data security standpoint. 

“We provide an encrypted token factory that secures tokens across workflows. We are seeing growing interest among companies in owning their intelligence, giving them greater control over their workloads, token consumption and costs,” said Pankaj Thapa, CEO and cofounder. 

“Our key differentiation is end-to-end encryption. For instance, when you use Claude directly, your prompts are sent in plain text. With our token factory, prompts are sent in an end-to-end encrypted format, while delivering the same responses from the underlying model. There is a marginal latency overhead of around 1–3%, but customers get the benefit of end-to-end encrypted data,” Thapa added.

India As An Inference Hub

For India, the opportunity extends beyond catering to domestic AI adoption. The country’s growing GPU infrastructure, data centre investments and access to a skilled developer ecosystem could help it emerge as a destination for inference workloads serving global customers.

With GPU availability and costs remaining key constraints for AI developers globally, infrastructure providers are looking to improve utilisation by serving multiple models and customers on the same computing infrastructure.

The shift towards token factories could also change the economics of India’s AI infrastructure businesses. Instead of simply renting out GPU capacity, providers can monetise the computing power through managed inference services, while building recurring revenue streams tied to actual usage.

“India can become a place where the world’s tokens are made long before it becomes the place where the leading models are designed, and that is a business worth winning. Three Indian token factories opened between September and November 2026, and six foreign neoclouds have been exploring Indian capacity for inference,” said Gogia.

In the case of Sarvam and other players looking to ride the inference wave, the economics will depend on GPU utilisation, model performance, token pricing, electricity costs and the ability to optimise workloads. 

The post Sarvam’s Token Factory Raises India’s AI Inference Stakes appeared first on Inc42 Media.

This content is published by Inc42 and has not been created or edited by Hello Headlines. View original

More in Startup

See all