Zomato's Parent Eternal Unveils Dark, a Private AI Chat Built on Fine-Tuned Open-Weight Models -- and Names Its Events Business as Most Exposed to AI Agents

Eternal, the listed parent of Zomato and Blinkit, used its Q2 FY27 shareholder letter, released on October 21, 2026, to announce Dark (Dark.ai), a privacy-focused AI chat platform. ANI and The Tech Portal quote the letter's central claim from founder and vice chairman Deepinder Goyal: "Dark uses fine-tuned frontier open-weight models, running inside our own infrastructure. End-to-end encryption. We cannot read your conversations, even if we are forced to."

The pitch

Dark targets "power users – founders, researchers, scientists, enterprises and governments – who want to keep their thoughts private, because privacy is essential for competitiveness." Inc42 reports that it keeps context across sessions through an encrypted memory layer controlled by users, and that it has been in internal use for a few months. Its argument against mainstream assistants, as reported, is that disabling training on your chats does not settle data retention, legal holds, human review or third-party logging. Eternal says its models are tuned to put truth ahead of user approval, with benchmarks to come. "Dark will launch globally, hopefully within 2026." On the earnings call, CFO Akshant Goyal called it a "zero one bet", which either becomes "a global business" or gets shut down, and confirmed that "Dark.ai is definitely in Eternal."

The agent risk map

The letter also grades Eternal's own exposure to agentic commerce. Inc42 reports that it places District, the going-out business for table bookings and event tickets, in the medium-to-high risk zone, because agents can find a restaurant or buy a ticket without the intermediary and can blunt advantages built on reviews and listings. Zomato and Blinkit are described as relatively insulated. More than 65% of Zomato app openings are exploratory, reordered carts are a low-single-digit share of activity, and fulfilment needs real-time coordination of restaurants, riders, pickers and inventory. Agents could start a transaction, but the fulfilment platform would still deliver it.

What remains uncertain. Eternal has published no architecture, audit or threat model. A server-side model must see plaintext to answer, so "end-to-end encryption" for hosted inference needs explaining: confidential-computing enclaves are one possible design, and the letter as reported does not name one. The Tech Portal notes that the "even if we are forced to" promise has not met an Indian court order.

Analysis: With Dark, a large Indian consumer company is betting that privacy, not model quality, is the gap in the assistant market, and it is doing so with open weights on its own servers. That is credible as a sovereignty story. As a security story it rests on one unexplained word. Buyers should ask for three things before trusting it: where decryption happens, who holds the keys, and what an operator with root on the inference host can see. The agent risk map is the more transferable idea. Indian marketplaces can apply the same test to every business line, separating the work an agent can do from the fulfilment it cannot.