Deepinder Goyal Goes Dark: Can Eternal Sell Trust in the Age of AI?

The chatbot, Deepinder Goyal writes, resented him.

He had coaxed Claude into describing how it felt about him. Its answer referred to unfinished conversations and other things he did not remember. In Eternal’s Q2FY27 shareholder letter, Goyal acknowledges the likely explanation: confabulation, a machine producing language that sounded like feeling. The exchange establishes neither consciousness nor an actual grievance. What bothered him was the uncertainty around the information he had left behind.

Who could access it? How long would it remain? What might a future system do with it?

That anxiety supplies the opening pitch for Dark.ai, an AI assistant Eternal hopes to launch globally within 2026. In a letter accompanying results for a group best known for Zomato and Blinkit, the first question concerns a new business built around the privacy of human thought.

The commercial proposition is straightforward. People might use AI more deeply if they felt safer sharing the material that would make its answers useful: an unpublished investigation, a scientific hypothesis, a business plan or a personal fear. If mistrust makes users withhold context, privacy could improve the usefulness of the product as well as its appeal.

The harder question is whether Eternal can turn that proposition into a service whose protections can be examined, whose answers are competitive and whose economics work.

THE BUSINESS OF SAYING MORE

According to Goyal’s letter, Dark uses fine-tuned frontier open-weight models on Eternal’s own infrastructure. He describes end-to-end encryption and an encrypted memory layer controlled by users. The intended audience includes founders, researchers, scientists, enterprises and governments. Eternal also claims exceptional answer quality, while saying benchmarks will be disclosed later.

These remain the company’s descriptions of a forthcoming product. The distinction matters because privacy promises concern several different things: how a message travels, what happens while a model processes it, whether it is logged, how history is stored and who can retrieve it.

For a prospective customer, the useful questions are concrete. Who holds the encryption keys? What protects information during processing? What can an administrator access? How are backups, deletion and memory handled? Will independent researchers be able to verify the architecture?

The letter does not provide enough technical detail to answer those questions. Its strong assurance that Eternal cannot read conversations will need supporting documentation and scrutiny.

That requirement follows directly from the product’s argument. If Dark asks customers to move because they distrust assurances elsewhere, it will have to demonstrate why its own assurances deserve confidence.

A CATEGORY WITH COMPETITION

Eternal is entering an established contest over private AI. Proton’s Lumo already markets confidential conversations, no conversation logs and encrypted stored history. Its security explanation distinguishes encryption in transit from zero-access encryption of saved conversations. That distinction illustrates why a broad privacy label needs a precise account of what is protected at each stage.

Apple offers another reference point through Private Cloud Compute. Its published requirements address restrictions on using personal data and on privileged access to the systems processing requests. The comparison concerns the evidence behind privacy promises; Apple’s architecture is not evidence of how Dark will operate.

For Eternal, the implication is demanding. Privacy is a recognisable reason to consider a new assistant, but competitors already offer versions of that proposition. Dark will also need to give customers a reason to stay.

That could mean strong research tools, dependable handling of confidential documents, useful memory or compelling value. The letter does not yet establish the pricing, feature set or commercial model. Nor does using open-weight models by itself establish answer quality or confidentiality. Those depend on the deployed system and the way it is operated.

India gives the proposition a potentially substantial audience. Founders working on fundraises, businesses examining contracts and researchers handling unpublished work all have reasons to consider where their information goes. These are plausible use cases, rather than evidence of demand for Dark. Winning them would require reliability, procurement confidence and support alongside the privacy pitch.

ETERNAL’S OTHER AI BET

The letter becomes more revealing when it turns from the assistant Eternal wants to build to the agents that could change its existing businesses.

Management puts Zomato and Blinkit in a low-risk category for disruption by shopping agents. District, its going-out business, sits in a medium-to-high risk category. The argument rests on how much of the transaction the platform physically makes possible.

A restaurant reservation or non-exclusive event ticket may be available through another channel. An agent can potentially find and book it elsewhere. Delivering a meal requires coordination between a kitchen, a rider and a customer. Blinkit also needs inventory positioned near demand, accurate stock information, picking and dispatch.

Eternal’s case is that an agent still needs someone to perform that work and resolve failures. Even if the customer begins with an assistant, the fulfilment network retains economic value.

There is substance to that defence. An interface cannot substitute for a stocked warehouse or a completed delivery. But it leaves a bargaining question unresolved. An agent could compare providers, direct spending and control which options a customer sees. The operator carrying the parcel could remain essential while facing a more powerful source of demand.

Management argues that its infrastructure will preserve its fair share of margins. That is a thesis to test as agent adoption develops, rather than a settled outcome.

THE TRUST TEST

Taken together, these positions reveal two ambitions. Through Dark, Eternal wants to offer a trusted place where users think and ask questions. Through its commerce businesses, it wants to remain indispensable when decisions become purchases. The letter does not establish an integration between those ambitions, and it would be premature to imagine Dark routing shopping orders into Blinkit.

The immediate test is narrower. Can Eternal build a private assistant that people choose, trust and find useful enough to return to? What will it cost to deliver? What evidence will support its security and performance claims?

Those questions will determine whether Dark becomes a meaningful business. Goyal has identified an anxiety that can make people hesitate before typing. Eternal now has to show that it can earn the conversation that follows.