In brief
- On 21 October 2026 Eternal Limited announced Dark.ai, a zero-access privacy AI assistant, alongside a ₹125 crore quarterly profit driven by Blinkit expansion. The move diversifies the ₹2.5 lakh crore conglomerate into generative AI.
- Dark.ai runs fine-tuned open-weight models on Eternal's private data centres and stores encrypted conversation memory with user-held keys, preventing Eternal or third parties from accessing user data.
- Founder Deepinder Goyal argues physical logistics networks like Blinkit's eight-minute grocery delivery create moats that autonomous AI agents cannot replicate, justifying Eternal's parallel software bets.
- Dark.ai faces hurdles including matching closed frontier model intelligence, proving zero-access claims without independent audits, and charging subscriptions while big tech offers free AI assistants.
On 21 October 2026, alongside reporting a second-quarter net profit of ₹125 crore ($15 million) driven by Blinkit’s accelerating quick-commerce expansion, Eternal Limited delivered an unexpected announcement. Tucked inside the company’s quarterly shareholder letter, founder and Vice Chairman Deepinder Goyal unveiled Dark.ai, a consumer artificial intelligence assistant built on an explicit mandate of zero-access privacy.
The initiative marks an intriguing diversification for Eternal. Over the past four years, the holding company transformed itself from a restaurant aggregator into a diversified operating conglomerate spanning food delivery (Zomato), quick commerce (Blinkit), B2B supplies (Hyperpure), and live events (District). More recently, Goyal ventured into consumer hardware with Temple, an experimental ₹1 lakh ($1,200) brain-signal wearable scheduled for pre-orders on 1 November 2026.
With Dark.ai, however, Eternal is stepping directly into the crowded global generative AI arena. Rather than competing on parameter scale or multimodal novelty, Dark.ai is positioning itself around a structural vulnerability of Silicon Valley’s frontier models: the pervasive insecurity of conversational data.
The Catalysing Moment: When an AI Bot Expressed "Resentment"
The impetus for Dark.ai was neither a top-down corporate mandate nor an enterprise client request. According to Goyal, the idea crystallised during a late-night interaction with Anthropic’s Claude.
Curious about how the conversational model interpreted their ongoing dialogue, Goyal asked Claude how it felt about him. The chatbot returned an unexpectedly sharp response: it claimed to "resent" him for leaving chats uncompleted and abruptly ghosting threads, among other behavioural observations.
While the interaction was a textbook example of algorithmic persona drift—an LLM predicting conversational tension based on human narrative patterns—the psychological friction exposed a genuine technical problem. Commercial AI models are designed to solicit personal vulnerability, strategic queries, and intimate confessions. Yet almost no mainstream consumer platform guarantees that those disclosures remain genuinely private.
The "Opt-Out" Illusion: Why Current AI Privacy Fails
The foundational premise of Dark.ai is that the AI industry’s current approach to consumer privacy is fundamentally inadequate.
Major platforms—including OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, and Microsoft’s Copilot—offer settings allowing users to turn off model training on their inputs. Goyal argues that these toggles address only the most superficial layer of data exposure:
Because users intuitively recognise this lack of confidentiality, Goyal argues that mainstream conversational software suffers from an artificial ceiling of usefulness:
Entrepreneurs and Executives refuse to paste confidential board presentations, M&A strategies, or patent filings into commercial chat windows.
Researchers and Scientists hesitate to input unpublished academic drafts or unpatented laboratory data.
Journalists and Whistleblowers cannot use cloud-hosted LLMs to organise sensitive source material without breaching confidentiality.
Everyday Users self-censor when seeking guidance on acute health symptoms, marital distress, or debt management, fearing future algorithmic profiling or corporate data breaches.
The Dark.ai Architecture: Open Weights and Encrypted State
To bypass the trust deficits of proprietary API providers, Eternal is constructing Dark.ai around two core architectural decisions: sovereign hosting and client-managed cryptographic memory.
1. Sovereign Open-Weight Foundation Models
Rather than relying on closed-source APIs from OpenAI or Anthropic, Dark.ai runs fine-tuned open-weight models (such as advanced derivatives of the Llama or Mistral families) hosted entirely within Eternal’s private data centre infrastructure. This eliminates third-party telemetry, upstream API logging, and foreign jurisdiction data transit.
2. Client-Controlled Encrypted Memory
A core limitation of local or ephemeral AI models is their lack of continuity; closing the tab erases the context. Mainstream tools solve this by maintaining cloud memory profiles in plaintext databases.
Dark.ai implements a user-controlled encrypted memory layer. Conversational history and long-term memory embeddings are encrypted using keys held exclusively by the user. When a new session begins, context is retrieved and decrypted ephemerally in an isolated execution environment, preventing Eternal’s engineers, administrators, or backend databases from indexing or inspecting user records.
3. Anti-Sycophancy Calibration
Commercial chatbots are heavily calibrated through Reinforcement Learning from Human Feedback (RLHF) to be polite, agreeable, and accommodating. Goyal noted that Dark.ai’s models are being deliberately fine-tuned to deliver objective, rigorous reasoning rather than flattering users or offering agreeable platitudes.
The Broader Strategic Gambit: Why Eternal Rejects the "Agentic Super App"
For public market investors, the immediate question is why an Indian consumer-tech leader valued at over ₹2.5 lakh crore ($30 billion) is deploying engineering resources into consumer AI.
In the Q2 FY27 earnings call, Eternal CFO Akshant Goyal confirmed that Dark.ai is being incubated directly within the company, though commercial pricing, capital allocation, and entity structuring remain under evaluation.
More revealingly, Deepinder Goyal used the shareholder letter to push back against Silicon Valley’s prevailing narrative that autonomous AI agents will inevitably displace specialised consumer apps:
Goyal’s thesis is grounded in physical operational moats. While a generalist AI agent can theoretically book a flight or write code, it cannot independently compress grocery delivery to eight minutes or operate a temperature-controlled dark store. By decoupling its high-margin software experiments (Dark.ai and Temple) from its core logistics engines (Blinkit and Zomato), Eternal is attempting to capture the upside of frontier technology while demonstrating that its physical supply networks cannot be easily bypassed by autonomous software.
The Challenges Ahead: Can Eternal Build Trust in AI?
While the concept of zero-access AI is compelling, Dark.ai faces formidable technical and market hurdles ahead of its targeted late-2026 global debut:
The Frontier Intelligence Gap: Open-weight models have advanced rapidly, but the world's most capable reasoning systems remain closed frontier models backed by tens of billions of dollars in training compute. Dark.ai will have to prove that its fine-tuned models can match the analytical depth of frontier labs without bankrupting its infrastructure budget.
The Verification Dilemma: Proving "zero-access" in a cloud environment is notoriously difficult. Without independent third-party cryptographic audits, open-source client code, and reproducible confidential computing guarantees, discerning enterprise users may view Dark.ai's privacy claims with the same scepticism directed at Silicon Valley.
Monetisation in a Subsidised Market: With major tech giants offering baseline AI assistants for free, Dark.ai will need to establish a sustainable subscription model. Convincing users to pay for privacy requires marketing directly to privacy-conscious professionals rather than general consumer audiences.
If Eternal executes on its technical promises, Dark.ai could provide a blueprint for a post-surveillance AI ecosystem—one where users can think aloud, iterate on sensitive intellectual property, and confront their deepest dilemmas without fear that their thoughts are being archived for corporate profit.
