Rajandran R Creator of OpenAlgo Sketchmaker, RagZ. Building GenAI Applications. Telecom Engineer turned Full-time Derivative Trader. Mostly Trading Nifty, Banknifty, High Liquid Stock Derivatives. Trading the Markets Since 2006 onwards. Using Market Profile and Orderflow for more than a decade. Designed and published 100+ open source trading systems on various trading tools. Now the recent curiosity on building Trading AI Agents.

Sarvam Token Factory Explained: Token Factory vs Neocloud vs Inference Provider

3 min read

Watercolour of an old Indian mill turned into an AI factory with server racks and a conveyor of glowing slips

A token factory is a company that owns AI computers (GPUs), runs AI models on them with its own fast software, and charges you for the words the AI writes, not for the time you use the computer. Sarvam launched its Sarvam Token Factory on 16 November 2026, It runs Sarvam’s own model and popular free-to-use models such as DeepSeek, GLM, Kimi and Gemma on computers in India. Many called it “just another neocloud”. It is not, and here is why.

Key takeaways

  • Neoclouds rent you AI computers. You set up the AI model yourself.
  • Aggregators like OpenRouter own no computers. They give you one login to use many AI companies.
  • A token factory owns the computers and the software that runs the AI, and charges per word.
  • The same AI model can be over 15 times faster on one company than on another.
  • Sarvam’s big selling point: your data stays in India. Banks, hospitals and government care about that.

The AI stack in one picture

The AI compute stack from chip makers and hyperscalers to neoclouds, aggregators and token factories

Think of AI like electricity. Someone makes the generators, someone runs the power plants, and someone sells you units on a meter. In AI, the units are called tokens. A token is a small piece of a word. About 100 tokens make 75 English words.

  • Chip makers like Nvidia make the AI chips, called GPUs.
  • Hyperscalers like Amazon, Google and Microsoft buy huge numbers of these chips and rent them out.
  • Neoclouds are newer, smaller companies that only rent AI computers.
  • Aggregators and token factories are the easy layer: you send a question and get an answer back.

What is a neocloud?

A neocloud is a company that rents out AI computers. A recent podcast on the Sarvam launch explained the two common types:

AI Bootcamp 2.0 for traders and investors

AI FOR TRADERS

AI Bootcamp 2.0 for traders and investors

Use ChatGPT, Claude and AI agents for research, coding and automation in your trading workflow.

  • Prompting and agent workflows for markets
  • Build tools without being a full-time coder
  • 6 courses in one bootcamp
See the bootcamp →by Marketcalls
  • Rent by the hour (RunPod, CoreWeave, Lambda): like renting a car. If the chip costs them $3 an hour, they may charge you $3.5. You do the driving.
  • Pay per second (Modal): like a taxi meter. The computer switches on only while your code runs, and you pay only for those seconds.

In both cases, you must set up and run the AI model yourself. The company does not know how many words your AI writes, so it charges only for time.

What is an aggregator?

An aggregator like OpenRouter owns no computers at all. It is like a food delivery app: one app, many restaurants, one bill. It does not cook anything. Sarvam Token Factory is the restaurant, not the delivery app.

What is a token factory and an inference provider?

Comparison of GPU rental, serverless, aggregator and token factory: who owns GPUs, inference stack and per token pricing

An inference provider is any company that runs an AI model for you and sends back the answers. (Inference simply means the AI giving an answer.) A token factory is the hardest kind: it owns the computers and writes its own software to run models faster, then charges per million tokens.

Why does this matter? The faster its software, the more words each computer can produce, and the more money it makes. Sarvam lists some of its tricks:

  • Memory (KV cache): not re-reading the same question twice. Sarvam says 95.7% of question text on one model is served from this memory.
  • Split the work: one group of chips reads your question, another writes the answer, so neither waits.
  • Tuned code: hand-written chip code and smaller number sizes that run faster with almost the same quality.
  • Guess ahead: guessing the next few words and checking them all at once.

Sarvam says its own model starts answering in 0.2 seconds and writes 330 tokens a second. Its launch price for one DeepSeek model is ₹14 for every million tokens you send and ₹55 for every million tokens you get back. These are Sarvam’s own numbers, not independent tests.

Inference providers compared: speed and cost

Time per task for gpt-oss-120B across inference providers from Artificial Analysis, Cerebras fastest at 0.9 minutes

Here is why the software matters. Artificial Analysis, an independent testing site, ran the same free model, OpenAI’s gpt-oss-120B, on many companies and timed the same task. Cerebras finished in 0.9 minutes. Together AI took 15.7 minutes.

Cost per task versus time per task for gpt-oss-120B inference providers, Artificial Analysis data
ProviderTime per taskCost per taskFirst answer token
Cerebras0.9 min$0.762.1 s
SambaNova2.9 min$0.495.7 s
Groq3.1 min$0.326.9 s
Crusoe4.9 min$0.1210.5 s
Nebius5.9 min$0.357.7 s
CoreWeave11.2 min$0.0718.9 s

Faster usually means costlier. Cerebras is fastest but most expensive. CoreWeave is cheapest but slow. Crusoe and Groq sit in a good middle. If an AI tool must answer you live, pay for speed. If the work can run overnight, choose cheap. Our guide to LLM metrics explains how to read these scorecards.

Why the Sarvam Token Factory matters for India

Until now, an Indian company using a model like DeepSeek usually sent its data to computers in the US. Banks, hospitals and government offices often cannot do that. Sarvam runs these models on computers in India and says it has India’s largest cluster of Nvidia’s newest B200 chips. It can also install the same setup inside a company’s own office.

AI chips are also very expensive, as our Anthropic IPO analysis shows. Selling tokens around the clock keeps them busy and funds the company.

The fair worry, raised in the same podcast, is focus. Many people backed Sarvam to build India’s own AI models for Indian languages and accents. A token factory is a hard and useful business. The real test is whether it pays for that bigger goal or pulls Sarvam away from it.

What this means for traders and developers

  • If you build AI trading tools, you can now use strong free models hosted in India, using the same code you use for OpenAI, with no computers to manage.
  • Choose by job: fast and costlier for live trading tools, slow and cheap for end-of-day analysis.
Python Elite membership

MEMBERSHIP

Python Elite membership

15 Python, AI and algo trading programs in one plan, including QuantFlow and AI Bootcamp live mentorship.

  • Python from fundamentals to trading apps
  • Backtesting, broker APIs and OpenAlgo
  • Live QuantFlow and AI Bootcamp sessions
View Python Elite →by Marketcalls

Learn by watching: free videos every week

Market Profile, order flow, Python, AI and OpenAlgo tutorials on YouTube.

Rajandran R Creator of OpenAlgo Sketchmaker, RagZ. Building GenAI Applications. Telecom Engineer turned Full-time Derivative Trader. Mostly Trading Nifty, Banknifty, High Liquid Stock Derivatives. Trading the Markets Since 2006 onwards. Using Market Profile and Orderflow for more than a decade. Designed and published 100+ open source trading systems on various trading tools. Now the recent curiosity on building Trading AI Agents.

Leave a Reply

Get Notifications, Alerts on Market Updates, Trading Tools, Automation & More