AI Pricing as Hidden Wealth Tax: How Token Billing Makes Swiss Enterprises Dependent

The New Reality: Token Billing as Uncontrollable Cost Driver

In a widely noted CNBC interview on July 1, 2026, Alex Karp, CEO of Palantir Technologies, chose unusually blunt language: the pricing model of leading AI providers resembles a wealth tax on enterprises. His criticism targeted a business model where companies pay for tokens that create no measurable value while simultaneously feeding proprietary data into providers training pipelines.

The numbers validate Karps warning: Uber consumed its entire AI budget for 2026 by April just four months in. With 5,000 engineers and spending between 500 and 2,000 dollars per power user monthly, costs escalated faster than financial planning anticipated. Andrew Macdonald, Ubers COO, admitted in the Rapid Response podcast: It is very hard to draw a line between one of those stats and Okay now we are actually producing like 25 percent more useful consumer features.

Gartner estimates global spending on AI agent software at 207 billion dollars in 2026 a 139 per cent increase from 86.4 billion in 2025. Yet fewer than one per cent of organisations achieve significant ROI exceeding 20 per cent. For Swiss decision-makers, this means the pricing power of frontier labs becomes a strategic risk when no control exists over spending and data sovereignty.

Why Token Billing Breaks Budgets

Token-based billing models appear fair at first: pay only for what you use. Reality is more complex. Agentic AI workflows systems that autonomously execute tasks, make decisions and communicate with other systems consume five to thirty times more tokens than simple chatbot interactions.

• Uncontrolled scaling: Every employee can independently trigger API calls without central budget accountability.
• Opaque pricing structures: Model upgrades, context window expansions or new features change consumption patterns without warning.
• Missing governance: Only 43 per cent of organisations have formal AI governance policies, just 21 per cent have mature agentic governance frameworks.
• Vendor lock-in: Once implemented, workflows are difficult to migrate to alternative providers, reducing price negotiation leverage.

Sam Altman, CEO of OpenAI, articulated the provider vision clearly in March 2026: We see a future where intelligence is a utility, like electricity or water, and people buy it from us on a meter. For enterprises, this means control over costs and data no longer resides internally.

The Swiss Context: Data Sovereignty as Compliance Risk

For Swiss enterprises, the problem intensifies through regulatory requirements. The revised Data Protection Act (revDPA) and FINMA directives for financial institutions demand comprehensive control over data processing and storage. Frontier AI services operating via public APIs create structural conflicts:

1. Data outflow: Training data, even when not used for training is promised, passes through external systems.
2. Jurisdictional risk: Server locations, encryption standards and access protocols lie outside direct control.
3. Audit gaps: No complete traceability of which data was processed when and where.
4. Vendor dependency: Price changes, service discontinuations or geopolitical developments (US export controls) become operational risks.

Karps criticism addresses precisely this point: Enterprises are livid over AI models that steal their business value. The data used to train models comes from enterprise interactions yet the value flows back to providers, not to data contributors.

Sovereign Alternatives: The Palantir-Nvidia Model

On June 29, 2026, Palantir and Nvidia announced a Sovereign AI Operating System an air-gapped solution based on Nvidia Blackwell Ultra GPUs and Nemotron models. The system runs entirely within enterprise infrastructure, without hosted API calls, without data outflow, without token billing.

This model addresses three central requirements of Swiss decision-makers:

• Cost control: Fixed infrastructure costs instead of variable, uncontrollable token bills.
• Data sovereignty: Data, models and weights remain in controlled data centres.
• Compliance: Complete audit trails, revDPA conformity, FINMA compatibility.

For Swiss banks, insurers and infrastructure operators, this is not a technical novelty but strategic necessity. The question is not whether to deploy AI, but how to deploy it with control.

Rethinking the Make-vs-Buy Calculus

The classic make-vs-buy decision was simple: develop internally or purchase externally? With AI, this question shifts: do we build our own models, rent tokens, or outsource AI capacity as an external division?

AI outsourcing does not mean surrendering control. It means operating AI capacity with defined SLAs, transparent costs and sovereign infrastructure without the pricing power of frontier labs. The article Forward-Deployed Engineering: When AI Outsourcing Becomes the New Normal describes this approach in detail.

Three Scenarios for Swiss Enterprises

Scenario one: Continue as before. Token billing runs uncontrolled, budgets are regularly exceeded, compliance risks rise. This scenario ends with emergency caps like Ubers or project cancellations.

Scenario two: Develop proprietary models. High initial investments, long development timelines, risk of technological obsolescence. Economically viable for very few Swiss enterprises.

Scenario three: Controlled outsourcing. AI is operated as an external division sovereign, cost-transparent, regulatory compliant. Models run in Swiss or EU data centres, data never leaves control, costs are fixed and calculable.

Most Swiss decision-makers will choose scenario three once they experience the cost reality of scenario one. The article From Pilot Trap to ROI: How Swiss SMEs Successfully Scale AI Agents demonstrates how the transition succeeds.

Action Recommendations for Swiss Decision-Makers

1. Create cost visibility: Implement central budget controls for all AI spending. 62 per cent cannot forecast monthly costs you should not be among them.
2. Establish governance: Formal policies for AI usage, agentic workflows and API access. Who may do what, with which budget, under which conditions?
3. Verify data sovereignty: Where is data processed, stored, potentially trained? RevDPA and FINMA demand answers.
4. Evaluate alternatives: Air-gapped systems, self-hosted models, controlled outsourcing. The technology is available; the question is strategic.
5. Measure ROI before scaling: Token billing scales automatically. ROI does not. Measure before you expand.
6. Reassess make-vs-buy: The question is no longer buy-or-build but control-or-depend. Choose control.

Enterprises are livid over AI models that steal their business value. They are paying for tokens that create no value while their data improves the providers models.

Conclusion: Pricing Power Is Strategic Risk

Karps wealth tax metaphor captures the essence: token billing is not a neutral accounting method but a business model that systematically transfers value from enterprises to providers. For Swiss decision-makers, this means regaining control over costs, data and strategic dependencies.

The technology for sovereign alternatives exists. The question is whether you act before the budget runs out or compliance requirements force you to. AI outsourcing as a controlled, transparent external division is not a future vision but available reality. The decision is yours.