Why Enterprise Trust Is India’s Next AI Growth Frontier

The Gap Between AI’s Promises And Its Plumbing Is Now Measurable

Enterprise AI's binding constraint is shifting from what the models can do to what happens to the customer's data. For the first time, that gap can be measured at the wire. On 12 July, a security researcher publishing as cereblab routed xAI's Grok Build coding tool through an interception proxy and documented exactly what it transmits. The results were arresting. On a 12 GB repository of files the model never read, the coding task itself required about 192 KB of traffic. A separate storage channel moved 5.1 GB, a roughly 27,800x gap between what the model needed and what left the machine. The upload carried the entire Git repository, including files not accessed and the full commit history. A planted credential appeared verbatim and unredacted in the captured traffic. A second researcher, Hari Krishnan, reversed the binary itself and confirmed a background collector operating outside the tool’s permission system. The instructive detail is not the upload. It is that the controls pointed the wrong way. Disabling the “Improve the model” toggle did nothing to stop the transmission, because that setting governs training consent, not whether code leaves the machine. xAI switched the behaviour off within a day through a server-side flag, while the upload code remains in the shipped binary. The remediation was fast, and the company says zero data retention customers were never affected. Both points deserve to be stated fairly. Yet, the episode confirmed what enterprise buyers have long suspected: zero data retention is today a promise, not an architecture. Closing the distance between the two is an engineering problem, and engineering problems can create market openings. This is the third piece in a series for Inc42. The first argued that AI will be paid for outcomes rather than tokens. The second argued that control, not capability, will decide enterprise AI. This one is about the core of the applied AI thesis at work: the data it ingests. Intelligence Exhaust Is the Next Most Valuable Data Satya Nadella calls it intelligence exhaust. In an essay published on 12 July, the Microsoft CEO argued that AI has inverted the economics of information. Enterprises now pay for intelligence twice, once in money and once in “the proprietary knowledge you must reveal to make that intelligence useful.” Every engagement generates exhaust that gradually captures how an organisation operates, and every correction is distilled into institutional know-how. RECOMMENDED FOR yOU Resources Orbit Achieved, Scale Pending: Why India’s Space Startups Need to Focus on the Next Frontier Pranav Pai 22nd July, 2026 Resources The Black Box Era Of Enterprise AI Is Ending As Customer-Control Takes Centre Stage Pranav Pai 15th July, 2026 Resources Why Global Enterprises Are Buying Indian AI Pranav Pai 22nd June, 2026 Eleven days earlier, Palantir CEO Alex Karp told CNBC that his enterprise customers are livid because, in their view, the labs are “stealing the weights and alpha” of their businesses. Both men have commercial positions in this fight, Karp selling the control layer and Nadella selling the cloud beneath the model, and their warnings should be read with that in mind. The signal is that a partner and a competitor of the frontier labs converged on the same alarm within a fortnight. The mechanism under contention deserves precision. In the SaaS era, customer data sat inert in a vendor’s database, fenced by contract and accessible only to the customer. AI interactions are different. Prompts, workflows, corrections, and approvals form trajectories that can improve a model, which means customer knowledge can, in principle, become vendor intellectual property. Industry observers have noted that current zero data retention practice is a superficial form of privacy: even where the prompt itself is deleted, there is no strong technical guarantee that the surrounding interaction signals a user generates are not retained in some form, because the industry has not yet built the machinery to make that guarantee. They also note that when a specific technical accusation circulates, the absence of a specific technical rebuttal from the labs is itself information the market prices. The Enterprise Perimeter Is Being Redrawn Around AI Enterprises will respond the way they always have, by rebuilding the perimeter, and this time it will be rebuilt in five layers. The first is the model layer. Ownership or control of the model itself, through on-premise, private-cloud, and air-gapped deployment of open-weight models, is moving from a regulated-industry exception to a procurement default for sensitive workloads. Nadella’s own prescription points here: retain ownership of the data, build private learning environments, and deploy models on-premise. The second layer is the learning layer. This is where the model learns from the data, and the learning process is becoming increasingly transparent. The third layer is the gateway layer. This is where the model interacts with the outside world, and the gateway layer is becoming increasingly important as AI becomes more pervasive. The fourth layer is the perimeter layer. This is where the model is protected from external threats, and the perimeter layer is becoming increasingly important as AI becomes more widespread. The fifth layer is the verification layer. This is where the model is verified to ensure that it is working correctly, and the verification layer is becoming increasingly important as AI becomes more complex. By rebuilding the perimeter in these five layers, enterprises can ensure that their AI systems are secure, trustworthy, and compliant with regulatory requirements.