Sarvam Token Factory Explained: India's Own 'AI Factory' – and Why It Sparked a Debate
Sarvam's Token Factory runs AI models on GPUs in India, billed in rupees. What it is, which models it offers, what they cost, and the 'sovereign AI' debate.
Key takeaways
- Sarvam Token Factory, launched in mid-November 2026, runs AI models on GPUs located in India and lets developers use them through one API, priced in rupees.
- It serves Sarvam's own model (Sarvam 105B) plus popular open-weight models such as DeepSeek, GLM, Kimi and Google's Gemma. Sarvam 105B starts at ₹10 per million input tokens.
- It sparked a debate: six of its eight models are Chinese. Critics say that's hosting, not 'sovereign AI'; Sarvam says the models run entirely on its own GPUs in India, with no requests sent abroad.
When you chat with an AI, your question goes to a computer packed with GPUs, which writes the answer one small piece at a time. Those pieces are called tokens – roughly, chunks of words – and AI companies charge by how many tokens you send and receive.
So what would you call a place whose whole job is turning out tokens? Bengaluru-based Sarvam AI calls it a Token Factory. Launched in mid-November 2026, it’s one of the most talked-about Indian AI launches of the year – partly for what it does, and partly for the argument it started.
What is Sarvam Token Factory?
Token Factory is Sarvam’s inference platform. “Inference” just means running an AI model to get answers, as opposed to training it.
In plain terms:
- Sarvam owns and runs the GPUs, in data centres located in India. It says it has India’s largest cluster of Nvidia’s B200 chips.
- On those GPUs it runs its own AI model plus several popular open-weight models – models whose makers publish them so anyone can run them.
- Developers and companies use all of them through one API, pay in rupees, and can switch models by changing a single line of code.
Sarvam also made the API OpenAI-compatible: an app already built for ChatGPT’s API can point at Sarvam’s address instead, with a Sarvam key, and keep working.
Which models, and what do they cost?
These are the models listed on Sarvam’s public price list, in rupees per one million tokens (about 750,000 words):
| Model | Made by | Best for (Sarvam’s description) | Input / output price |
|---|---|---|---|
| Sarvam 105B | Sarvam (India) | Reasoning, coding and agents in Indian languages | ₹10 / ₹25* |
| Sarvam 105B Conversations | Sarvam (India) | Real-time dialogue and voice agents | ₹10 / ₹25* |
| DeepSeek V4.1 Flash | DeepSeek (China) | Long documents, cost-conscious reasoning | ₹14 / ₹55* |
| GLM-5.3 Flash | Zhipu (China) | Fast reasoning with image input | ₹13.5 / ₹45 |
| Gemma 4 31B | Google (US) | Image tasks and quick replies | ₹35 / ₹90 |
| GLM-5.3 | Zhipu (China) | Coding and long-running agent tasks | ₹126 / ₹390 |
* Introductory prices, valid until the end of 2026. Bigger models such as Kimi K3.1 Pro and DeepSeek V4 Pro are offered to companies on reserved capacity, with prices on request.
For a sense of scale: a 300-page English book is roughly 100,000–120,000 tokens, so having Sarvam 105B read the whole thing costs about ₹1 to ₹1.20 at the introductory input price. Whatever it writes back is charged separately, at the output rate.
About Sarvam 105B
Sarvam’s own model has 105 billion parameters, but it uses a “mixture of experts” design that switches on only about a tenth of them for each token, which keeps it fast and cheap to run. Sarvam released it openly in February 2026, and it was trained with Indian languages in mind.
Is it fast?
Speed matters because nobody wants to wait for a chatbot or a voice assistant. Sarvam published its own tests, run from a computer in Mumbai:
- Sarvam 105B: first word back in about 0.2 seconds, then around 330 tokens a second.
- Gemma 4 31B with a long, 10,000-token question: first token in 0.4 seconds versus 2 seconds from the only other provider serving it, and 308 versus 193 tokens a second.
- On other models it says it matches the fastest global providers – though its own charts show a rival slightly faster on one DeepSeek test.
These are Sarvam’s own measurements, not independent tests, but they’re detailed and include the cases where it doesn’t win.
Under the hood, Sarvam says it built on open-source serving software (vLLM, SGLang and Nvidia’s Dynamo, among others), split the “read your question” and “write the answer” steps onto separate groups of GPUs, and added a smart cache so repeated prompts aren’t processed twice.
Why it matters for India
- Your data stays in India. Sarvam says the models run entirely on its own GPUs in India, and that requests are not forwarded to the original model maker’s servers abroad. For banks, hospitals and government projects with data-residency rules, that’s the main draw.
- Rupee billing. No dollar cards or currency swings for Indian startups.
- Choice without lock-in. One API, many models – if a better open model appears, switching is a one-line change.
- Indian-language strength. Sarvam 105B is built for Indian languages, which global models often handle less well.
The debate: “sovereign AI” or just hosting?
The launch quickly drew criticism online, and it’s worth understanding both sides.
The criticism
- Most of the models are Chinese. Of the eight models in its launch line-up, six come from Chinese labs (DeepSeek, Zhipu’s GLM and Moonshot’s Kimi), one from Google, and one from Sarvam.
- Is hosting “sovereign AI”? Sanchit Monga, CEO of RunAnywhere, argued that serving other companies’ models is a hosting business, not the frontier research India’s sovereign AI push was meant to produce.
- Government support. Critics pointed out that Sarvam was chosen in 2025 to build India’s own foundation model under the government’s IndiaAI Mission, with subsidised GPU access.
- Too much in one place. Srikanth Lakshmanan of CashlessConsumer warned that one company becoming a key AI provider concentrates critical infrastructure.
Sarvam’s response
- Co-founder Pratyush Kumar said the government support was a 40% GPU discount for training its models, which were then released openly – and that this access ended in March 2026.
- Token Factory, Sarvam says, runs on privately bought GPUs and doesn’t rely on government subsidies.
- Running open models locally, it argues, is exactly what keeps Indian data in India.
Others weighed in: MediaNama founder Nikhil Pahwa saw no problem with hosting Chinese models, though he noted DeepSeek’s price on Token Factory is similar to elsewhere, so there’s little cost reason to switch. Hemant Mohapatra of Lightspeed India – an investor in Sarvam – called it a timely move during a global GPU shortage.
Should you try it?
- Building an app for Indian users, especially in Indian languages? Sarvam 105B at ₹10 per million input tokens is worth testing.
- Need data kept in India for compliance? Token Factory is one of the few options that runs top open models locally with rupee billing.
- Already happy with a global provider and no data-residency rules? Prices for the non-Sarvam models are similar to the market, so the reasons to move are local hosting and convenience, not big savings.
The bottom line
Token Factory isn’t a new AI model – it’s the factory floor: India-based GPUs, a fast serving system and one rupee-priced API for Sarvam’s own model and the best open models from around the world. Whether that counts as “sovereign AI” depends on what you think sovereignty means: who built the model, or where it runs and where your data goes. Either way, Indian developers now have one more homegrown option – and that’s a good thing to have.
Prices and model lists are from Sarvam’s documentation as of 21 November 2026 and may change, especially when introductory pricing ends.
Your turn: does it matter to you where an AI model was made, or only where your data is processed? Tell us in the YouTube comments.
Sources
- Sarvam: Introducing Sarvam Token Factory (launch post)
- Sarvam docs: Token Factory – models, model IDs and prices
- MediaNama: Sarvam's Token Factory stirs debate on India's 'sovereign' AI push (Nov 2026)
- Business Standard: Explained – Sarvam AI's Token Factory and the debate over AI sovereignty (20 Nov 2026)
- MarkTechPost: Sarvam launches Token Factory – India-hosted inference for 8 open models (17 Nov 2026)
- Wikipedia: Sarvam AI



