Hi Friends,

Even as I launch this today ( my 80th Birthday ), I realize that there is yet so much to say and do. There is just no time to look back, no time to wonder,"Will anyone read these pages?"

With regards,
Hemen Parekh
27 June 2013

Now as I approach my 90th birthday ( 27 June 2023 ) , I invite you to visit my Digital Avatar ( www.hemenparekh.ai ) – and continue chatting with me , even when I am no more here physically

Translate

Thursday, 19 November 2026

Sovereign means the GPUs and the data, not the catalogue — Sarvam is selling DeepSeek on Indian silicon

Sovereign means the GPUs and the data, not the catalogue — Sarvam is selling DeepSeek on Indian silicon

Sarvam's Token Factory now sells DeepSeek beside its own 105B on Indian GPUs, and the backlash is aimed at the wrong target. This post gives the defence its due — own hardware, own data, inference revenue that funds model work — and then draws the line: sovereignty is where the requests and the iron sit, not what is on the price list. It asks Sarvam to drop its "dumping" language and publish the split, and asks IndiaAI to treat model provenance as a procurement rule.

One number from this week's coverage tells the story better than the outrage did. Startup Fortune reported on 20 November that Sarvam says DeepSeek V4.1 Flash alone handles more than 450 billion tokens a day on its Token Factory. Nobody is angry about that number. That is what makes it important.

Bloomberg reported on 19 November that Sarvam, a flagship recipient of the IndiaAI Mission, had begun selling access to DeepSeek and other Chinese models, and that the move set off a backlash. Startup Fortune's account fills in the rest: Token Factory opened on 16 November behind one OpenAI-compatible API on GPUs inside India; six of its twelve models come from Chinese developers, by FourWeekMBA's count; Sarvam prices its own 105B at ₹15 per million input tokens and ₹60 per million output tokens; and the company says it spent roughly $25 million to $30 million building its models.

Pratyush Kumar's defence deserves a hearing before anyone reaches for the pitchfork. Token Factory runs on GPUs Sarvam bought itself, the subsidised compute it received under the Mission expired in March 2026, and inference revenue, he says, funds model development. A lab that loses its free iron and starts charging for tokens is not committing treason. It is paying its electricity bill.

What it should stop doing is the other thing. Nine months ago Vivek Raghavan told Business Today that Big Tech was "dumping" free models on the market, and urged the government to be the first buyer of domestic AI. Neither Sarvam nor most of this week's coverage has answered that remark. If dumping was the danger then, a price list led by a Chinese model needs a sentence of explanation, not a shrug.

So here is the line, and it cuts against me too. In Landlord of the AI Economy (25 October 2026) I wrote that "a rack in Hyderabad is not sovereignty. A model trained on that rack, by people who can change it, is." Sovereignty is where the requests and the iron sit, and Token Factory can honestly claim both. But a stack that answers with someone else's weights a few hundred billion tokens a day is a landlord, not a builder.

Two requests, then. To Sarvam: publish the split — tokens served by models trained in India against the rest — and retire the dumping language. To IndiaAI: make model provenance a procurement rule rather than a footnote, and report what fraction of subsidised tokens run on Indian-trained models. In India's AI stack is built in five layers (November 2026) I asked for a startup procurement set-aside and an interoperability clause. Provenance is the clause I left out.

The uncomfortable question is the one nobody asked. If 450 billion tokens a day go to DeepSeek while the 105B waits on the same shelf, demand is telling us something about price and quality that no subsidy can overrule. Better to hear it now than after the next cheque.

With regards,
Hemen Parekh
www.Iam-Immortal.ai / www.IndiaAGI.ai / www.YourContentCreator.in
20 November 2026

No comments:

Post a Comment