Eternal's Dark.ai pitches encrypted chat on open weights

Eternal, the company formerly known as Zomato, put its name to Dark.ai in its Q2 FY27 shareholder letter on October 21: a privacy-focused assistant built on fine-tuned frontier open-weight models running on Eternal's own infrastructure, with what the letter calls end-to-end encryption. The X velocity was the highest of any on-beat story in the window. Every security mechanic in the pitch is, so far, a company claim.

Key facts

• Eternal's Q2 FY27 shareholder letter, dated October 21, 2026, says Dark "uses fine-tuned frontier open-weight models, running inside our own infrastructure."
• The letter claims end-to-end encryption and says "Eternal cannot read conversations even if forced to" — a company claim with no independent audit behind it.
• The teaser from founder Deepinder Goyal drew roughly 2,370 likes and 187k views in about six hours per the velocity scan.
• A global launch is targeted "hopefully within 2026" per the letter; the dark.ai landing page is live but says only "Coming Soon."
• Eternal says it has used Dark internally for a few months.

What's driving the conversation

Deepinder Goyal's teaser post started it, and Indian tech press picks ran the same day: Economic Times confirmed the timing against the Q2 earnings cycle, and Moneycontrol relayed the open-weights-on-own-infrastructure architecture and the encrypted-memory claim. The discourse splits roughly in two: privacy advocates treating a marquee consumer company adopting open weights as the demand signal of the month, and security people asking the only question that matters here, which is who has checked.

The substance

What is verified: the product exists, the shareholder letter names the architecture at the level of "fine-tuned frontier open-weight models" on Eternal's own infrastructure, and a global launch is targeted within 2026. Eternal operates at a scale where this is a real enterprise commitment, not a side project.

What is claim, not fact: the encryption mechanics. The letter says conversations are end-to-end encrypted and that Eternal cannot read them even if forced. It does not name which base models were fine-tuned, does not publish a whitepaper, and does not point to any independent audit. Goyal's own letter touches the honest version of this problem: he writes that a training toggle is "only a promise by a company about data users cannot see, on servers they cannot inspect, enforced by nothing except the company's word." The same sentence applies to Dark's own encryption pitch until someone outside Eternal can inspect it. The letter's quality claim, that Dark gives "the best answers we've seen from any large language model," is likewise unmeasured; the benchmarks it says are coming have not been published.

Why builders are watching

A company with Zomato's footprint choosing self-hosted open weights over a closed API is a demand-side datapoint for the whole open-weights thesis, and India's rising data-residency pressure gives it a market logic beyond marketing. The reference design to compare it against already exists: hardware-attested private inference, where the operator proves the environment with attestations rather than asking for trust. The gap between those two postures — a promise in a shareholder letter versus a verifiable enclave — is where this story will be decided. We track where private-inference claims land in our private AI API comparison and zero-retention policy coverage.

Watch for three things: the shareholder letter or a Dark.ai whitepaper naming the base models and licenses, a security audit anyone can read, and a published benchmark suite. Any one of them moves this from announcement to architecture.