One startup. Half the money. Zero room to fail.

On May 23, 2025 — four weeks after the Government of India anointed it the builder of the nation's sovereign LLM — Sarvam shipped its first model since selection. Sarvam-M was not trained from scratch; it was a fine-tune of Mistral Small, a French company's open-weights model, adapted for ten Indian languages. Two days later its Hugging Face download counter read 23. That month, a side project by two Korean students had pulled 200,000. "Embarrassing," wrote Menlo Ventures' Deedy Das, and the Indian internet spent a week arguing about whether he was allowed to say it.

The episode was survivable — a bad launch, later partially redeemed. But hold the image, because it is the whole thesis at miniature scale: a company chosen by the state, wrapped in the flag, building on a borrowed foundation, meets the open market — and the market shrugs. In May 2025, the stake was one model's downloads. Today the same company carries a $1.5 billion valuation, a funding round equal to roughly half of India's entire annual AI startup funding, the "sovereign AI" brand on its outputs — and now, reportedly, the Government of India on its cap table. The download counter has become a national position. This dossier is about what happens if it ever reads 23 again.

The mechanics, as first reported by the Economic Times on June 25, 2026: Sarvam's six-month allocation of 4,096 Nvidia H100s ran up a compute bill of ₹246.71 crore, of which the government absorbed ₹98.68 crore — the single largest subsidy of the IndiaAI Mission's first phase. That support was issued not as a grant but against compulsorily convertible debentures, expected to convert into roughly 1–2% equity when Sarvam completes the ~$300 million Series B it first-closed on June 15 at a $1.5 billion valuation. A senior official put the rationale plainly: "The support provided to companies under the IndiaAI Mission needs to be accounted for in some form, if not cash."

One honesty note: the percentage is anonymously sourced and, as of writing, unconfirmed by MeitY or Sarvam — queries to both went unanswered. But the structure is not rumour. The equity clause sat in the IndiaAI call for proposals from January 2025, and Sarvam's own co-founder said it on the record fourteen months before the story broke, replying — fittingly — to a critic who had called the ₹220 crore award a giveaway:

"Clarification: This is not a grant. A gov body will take equity in Sarvam for the compute we receive." — PRATYUSH KUMAR, CO-FOUNDER OF SARVAM, ON X · APRIL 2025

Now notice what the numbers admit. One to two percent of $1.5 billion is $15–30 million — a rounding error against a ₹10,372 crore mission. Nobody designs a position that small for returns. It exists to establish a principle: public support for a startup should show up as public ownership of the startup. Twelve companies are already in the foundation-model programme; several formally objected to the CCD structure, warning it turns a policy programme into "a quasi-VC fund" breeding "a small club of state-backed winners." They were overruled. So before asking what the state's shareholding does to the ecosystem, ask the prior question: is the company at the centre of the bet even the right bet?

Consider what the money buys elsewhere. DeepSeek's V3 cost $5.6 million of compute and came with a technical report the entire field now cites — new attention architecture, new training economics, a genuine shove of the global frontier, achieved with no state equity on its cap table. That is what "frontier lab" means: you leave the field different than you found it. Now audit the champion India picked from 67 proposals. Its first post-selection release was a fine-tune of a French model, marketed under a sovereign flag. Its from-scratch models — Sarvam 30B and 105B, launched February 2026 — are competent, real engineering, and four months late against a six-month promise. But they are standard recipes on imported chips. The strongest claim the co-founder himself makes is that the 105B matches the scores DeepSeek R1 got "when it was released" — that is, it reached, in 2026, where an open Chinese model stood in January 2025. Matching last year's frontier is called adoption. No lab anywhere cites a Sarvam paper to explain how they train models.

Even the scoreboard is compromised: as Forbes reported, the leading Indic benchmarks were built at AI4Bharat, co-founded by the same people who founded Sarvam — the champion partly controls the yardstick it is measured by. And here is the twist that makes the picture coherent rather than merely bleak: Sarvam's harshest critic recanted. In February 2026 Deedy Das wrote "I was wrong about Sarvam" — because its Indic speech, OCR and voice models are genuinely the best available, "doing things big labs will probably never focus on." Read that carefully. Sarvam's proven excellence is the application layer — Indic voice and documents, a real and valuable niche. Its subsidised ambition is frontier foundation models, where its record is imitation with a lag. The state took equity in the story, not the strength: it bought into a frontier-lab narrative whose demonstrated edge is somewhere else entirely, priced at 275× revenue. A fine company. The wrong horse for a nation to ride — and, as the next chapter shows, the state didn't just pick the horse. It moved into the stable.

The reflexive defence of the stake is its size — one or two percent, no board seat reported. But the sliver is the standard opening position of a state that wants presence without the optics of control. In 2021, Chinese state entities took exactly 1% of ByteDance's main domestic subsidiary — attached: one of three board seats and control rights over content. In 2025, Washington converted grants into a 9.9% "passive" stake in Intel — attached, per the 8-K: a warrant that effectively vetoes any sale of Intel's foundry business. Lower Saxony's 20% of Volkswagen came fused to a statutory takeover veto; when the European Court of Justice struck it down, Germany redrafted the law and kept it. The pattern to memorise: once a state has cap-table leverage, it legislates to keep it.

And this particular shareholder has references you can call. The angel tax turned startup fundraising into tax disputes for twelve years before being abolished as a mistake. The 28% gaming GST arrived with ₹1.12 lakh crore of retrospective demands — roughly 15–20× the industry's revenue. The crypto TDS cut domestic volumes 81% in four months and pushed $42B offshore. Press Note 3 approved 124 of 526 FDI proposals in four years. None of this required malice — only volatility with sovereign immunity. A regulator that changes rules harms you at arm's length; a shareholder-regulator follows you home, to your next raise, your IPO, your acquisition. The people who price such risk for a living saw it immediately:

"Tomorrow, you raise capital — you'll need government approval. You want to exit — you could face harassment or delays." — T.V. MOHANDAS PAI, AARIN CAPITAL & FORMER CFO OF INFOSYS · JUNE 2025

UPI's operator, NPCI, is a non-profit utility owned by a consortium of banks. The state built the rails, set the standards, and took zero equity in PhonePe, Google Pay, or anyone else who built on top. Result: 228 billion transactions in 2025 and ferocious private competition. The state's power sat under everyone, inside no one. India even has the capital version running: SIDBI's Fund of Funds deploys ₹20,000 crore of public money through 140+ private fund managers into a thousand-plus startups — and the government never appears on a cap table. Norway's $1.8 trillion fund is legally barred from domestic companies precisely to avoid distorting its own economy. States that want returns build arms-length, dispersed vehicles. States that want presence take direct stakes. The CCD is a presence instrument wearing a returns costume.

The final irony: Sarvam itself has told the government what it actually needs, and it isn't a shareholder. Vivek Raghavan has argued the government should be the first buyer of Indian AI — procurement, the honest subsidy, which pays for performance instead of presence. Critics like Paras Chopra point to the other honest instrument: if taxpayers fund a model, condition the grant on open weights — "If it's for Bharat, funded by taxpayers, shouldn't it belong to Bharat?" Both capture more public value than a 1–2% CCD ever will. Neither follows a founder home. The tools were on the table. The state picked the only one that moves in.