The first fight in enterprise AI is over per-token pricing. The next is over ownership of data, weights, and competitive advantage, and regulators are already picking a side. The winners of the next phase will deliver outcomes on infrastructure the customer controls, and India is well placed to build them.

The next phase of enterprise AI will be decided by control, not just capability. Alex Karp made that case in the most theatrical way possible on CNBC last week. The Palantir CEO said the chief executives he speaks with privately are livid with the leading model developers, arguing that enterprises pay steep fees for AI tools while the providers collect proprietary data that can be used to sharpen their own models, an arrangement he likened to a wealth tax on business.

Enterprises, in his telling, are paying for tokens that create little measurable value while handing over the very data and process knowledge that constitutes their competitive edge. Palantir’s shares jumped about 9% the day the interview aired, which suggests the market heard something more than showmanship.

Karp is not a neutral witness. The interview was built around a partnership announcement, Palantir sells precisely the alternative he was describing, and he conceded on air that he too profits from the practices he was attacking.

The right response to a self-interested argument is to test it, not to dismiss it. Stripped of the theatre, his claim is based on emerging consensus, and the evidence assembled below supports it independently of the messenger.

In a previous piece for Inc42, I argued that AI billing should move billing from effort to outcomes, because enterprises will not indefinitely fund a meter that charges for activity rather than results.

That was the first half of the problem. The second half is harder. Even a fairly priced token is a bad trade if the transaction also transfers your data, your workflows, and your accumulated advantage to the counterparty.

Every serious AI procurement conversation is converging on the same set of questions.

•       Who owns the data?
•       Where is it cached?
•       Who controls the weights?
•       How can we ensure you aren’t training your model with our data?

Karp framed these as the questions enterprises are now asking, and described the demand from technical customers as control over their compute, their models, their data stack and their alpha. Within a year, these questions will sit in standard procurement templates the way security certifications and audit logs do today.

The conflict beneath them is emerging. A customer’s proprietary data and process knowledge is exactly the raw material that improves a foundation model from here. The vendor’s incentive and the customer’s interest point in opposite directions, and no contractual assurance fully resolves that tension when the system doing the work is a black box.

The consequence follows directly. Any product whose design requires the customer to give up visibility into what happens to its data carries a ceiling on the market it can win. The technology may work perfectly. The trust never forms at the scale that large, multi-year contracts require, and large, multi-year contracts are where enterprise AI revenue lives.

India’s own regulator has already moved, and earlier than most. The RBI’s FREE-AI framework signals a clear preference for indigenous, sector-specific AI models over generic third-party LLMs, naming data sovereignty and vendor concentration as the risks, and anticipates coordinated standards across RBI, SEBI, and IRDAI over time.

Boards and senior management remain ultimately accountable for third-party models, which institutions are expected to validate as rigorously as their own. A bank’s board cannot validate what it cannot inspect. The regulatory logic and the black-box architecture are therefore incompatible by construction.

The direction is global. Certain categories of data legally cannot leave infrastructure the enterprise controls, including protected health information, specific classes of financial data, and defence work. Zero-retention enterprise tiers reduce the exposure without eliminating it, and the EU AI Act becomes fully applicable on 2 August 2026, with European data residency rules already driving a measurable move toward self-hosting.

Where such rules do not yet exist in banking, insurance, and energy, they are coming, because every serious regulator will follow the same systemic-risk reasoning. Each new notification shrinks the addressable market for opaque AI in the sectors that spend the most on technology.

 The next runway exists, and it has matured faster than most buyers realise. For classification, extraction, summarisation, retrieval-based question answering, and most coding, a good open-weight model in 2026 simply does the job, with the closed-model lead still existing in long-horizon agentic reasoning. Most regulated enterprise workloads do not live in that corner.

Open-weight systems now post coding scores within a few points of the best closed models at a tenth to a thirtieth of the cost per token, and a single open family crossed one billion downloads in January 2026.

The market has voted with its workloads. Open-weight providers now account for over 45% of traffic on OpenRouter, the largest model aggregator, up from under 2% a year ago. The savings at stake are not marginal. Prosus, one of the largest technology investors in the world, has reported inference cost reductions of up to 26x from running open models in place of proprietary ones.

The deeper point is what open weights actually purchase. Three rights:

• the right to run the model on infrastructure you control,
•  the right to fine-tune it on your own data,
• and freedom from a vendor’s pricing and lifecycle decisions.

Those three rights are the direct answers to the procurement questions. That symmetry is why the shift is durable rather than temporary.

India has been deploying open-source models, as well as developing its own. That flexibile design posture is exactly what regulated buyers need, because it permits inspection, fine-tuning, and collaborative deployment inside their own perimeter.

The larger prize sits one layer up. Karp calls it the application layer. India should read it as a services-shaped opportunity of the kind it has won before.

India built a technology industry of $315 Bn (₹30 Lakh Cr) in annual FY26 revenue, per NASSCOM’s Annual Strategic Review 2026, by implementing the world’s enterprise software on client terms. The AI version of that industry is larger per engagement, because implementation now spans model selection, fine-tuning on private data, evaluation harnesses, pipeline integrations, and the compliance evidence that frameworks like FREE-AI will demand. The delivery base already exists: 2,117 GCCs generating $98.4 Bn (₹9.4 Lakh crore) and employing 2.4 Mn professionals.

Indian builders are already proving the pattern at both ends. H2LooP, headquartered in Bengaluru, builds hardware-aware vertical AI models for safety-critical embedded systems across automotive, aerospace, and semiconductor industries, positioned deliberately around IP sovereignty and on-edge reliability, the exact stack a general-purpose black box cannot serve.

Smallest.ai, also built from Bengaluru, runs foundational voice models handling over a million inference calls for global enterprise customers including RingCentral, Paytm, and Truecaller, demonstrating that Indian model companies can win global-scale enterprise workloads on these terms.

The prescription for founders follows sequentially.

• Build model-agnostic deployment stacks for regulated Indian industries, so the client keeps the weights and the audit trail.
•  Treat compliance artefacts under emerging regulations like FREE-AI and its successors as a product, not a cost centre.
• Price on outcomes or directed inference, delivered on infrastructure the client controls, which joins this argument to the previous one.

My earlier piece argued that enterprises will pay for outcomes rather than effort. This one adds the condition attached to the payment. The outcome must arrive without the customer’s advantage leaving the building.

The winning AI business of the next decade sells a result the customer can measure, on a model the customer can inspect, running on infrastructure the customer controls.

India has the open models, the engineering depth, and now the emerging regulatory clarity to build that business for its own critical industries first. India for the World begins at home, inside the perimeter.

The post The Black Box Era Of Enterprise AI Is Ending, As Customer-Controlled AI Takes Spotlight appeared first on Inc42 Media.