Palantir's CTO Shyam Sankar declared on the Q1 2026 earnings call that "tokens are the new coal" and "AIP is the train. It is a clever soundbite, without question. It also reveals the core thermodynamics of a business model built on governing AI consumption at an industrial scale. Sankar invoked the Jevons paradox (remember: the Victorian-era observation that more efficient steam engines did not reduce coal consumption but caused it to skyrocket) to explain why Palantir's clients are burning through record numbers of AI tokens as inference costs collapse to roughly one-thousandth of what they were 3 years ago (2023) at the peak of the “ChatGPT frenzy” and the beginning of inflated expectations.
I have to concede that his metaphor is brilliant because it is also uncomfortable. It accurately identifies a shift toward the commoditization of cognition. But it simultaneously conceals a more difficult question because nearly 200 years after the events, we know of the geopolitical consequences of burning so much coal: who owns the mine, who controls the railway, who breathes the smoke, and how on earth you got to so many businesses eager to burn your coal and pollute.
Palantir's Q1 2026 results reveal the sheer scale of the transformation. Revenue surged 85% year-over-year to $1.63 billion, with US government revenue rising 84% to $687 million and US commercial revenue climbing 133% to $595 million. US revenue as a whole doubled. On the same call, Sankar positioned Palantir as a "no slop zone," thus rejecting the industry trend of "tokenmaxxing" (meaning ‘maximizing token consumption for its own sake rather than driving measurable business outcomes’). "More tokens means more slop," Sankar said.
That is impressive engineering. It is also a very good business model. AIP governs every prompt, every response, every agent action: metered, attributed, and billable. The platform becomes the railway, the customer's operational data the cargo, the AI model the engine, and tokens the coal. Once that railway is embedded into a ministry of defense, a hospital system, or a battlefield command structure, the economic center of gravity shifts irreversibly.
But here lies the first tension: this represents a shift from capital expenditure on systems to operational expenditure on cognition. Public procurement frameworks were not designed for variable, usage-based AI spend where the meter runs on every query. When a Ministry of Defense analyst queries a classified dataset through AIP, or an NHS administrator runs a predictive workflow on patient pathways, tokens burn. And the taxpayer pays… not once, but continuously, with costs that are harder to audit and cap than those of traditional software licensing or on-premises solutions.
These are not chatbot use cases but operational systems. That is the profound commercial insight behind the coal metaphor, and the reason Palantir's revenue trajectory commands attention well beyond Silicon Valley.
Palantir's token consumption is not abstract. It flows from concrete, high-stakes use cases across multiple verticals and geographies.
Defense and intelligence. In the United States, Palantir's Maven AI system has become the primary AI operating system for the US military and is set to become an official Pentagon program of record, a designation that will streamline its adoption across all branches and provide stable, long-term funding. In the United Kingdom, the Ministry of Defense signed a £240.6 million contract with Palantir in December 2025, by direct award and without tender. In Spain, the Ministry of Defense holds at least 2 contracts with Palantir's Spanish subsidiary, including a €16.5 million "intelligence fusion" award, negotiated without public tender and protected by official secrecy.
Spain has often been singled out by President Trump for not raising defense spending to 5% from 2%. It has also been singled out by Mr Trump as a country "we should cut all our trade with". The US is Spain's 5th largest trading partner, with a positive balance for the US
Healthcare. NHS England awarded Palantir a £330 million contract to build the Federated Data Platform, a project that has drawn sustained criticism from the British Medical Association, patient groups, MPs who described the deal as "dreadful" and "shameful," and over 100 health workers who have called for the contract to be scrapped.
Financial regulation and commercial verticals. The UK Financial Conduct Authority has also contracted Palantir, while large European corporations (including BBVA, Mahou San Miguel, and Mutua Madrileña in Spain) reportedly use its platforms for credit risk, supply chain, and fraud detection.
To understand the gravity of the "railway" metaphor, one must look beyond just the Spanish market to the foundational engineering projects that define Palantir’s global footprint. We are not merely talking about vendor-client relationships but about deep-tissue structural integrations in which the software governs the physical and operational throughput of entire industries.
The primary architectural blueprint is Skywise, the aviation data platform co-developed with Airbus. By 2026, Skywise has matured into a ubiquitous digital services ecosystem for over 100 airlines, managing the technical telemetry and maintenance cycles of thousands of aircraft. When an airline makes use of these models to optimize fuel consumption or part replacement, it is engaging in a high-frequency inference process. This is the ineluctable outcome of Sankar’s railway: a world where aerospace logistics are inseparable from the proprietary logic of the platform.
In the energy sector, the integration with BP illustrates the "digital twin" concept at a planetary scale. By mapping the thermodynamics and flow rates of global extraction points, BP has reportedly captured an additional 30,000 barrels of oil per day through model-driven asset allocation. In this context, the tokens consumed are the invisible catalysts in a feedback loop that transforms raw sensor data into actionable yield. It represents the commoditization of industrial oversight, where the "railway" manages the very extraction of the energy that once powered the original coal engines.
The most significant strategic milestone, however, is the US Army’s Maven Smart System (MSS). In March 2026, the Pentagon designated Maven as an official Program of Record, a formal engineering transition from experimental prototype to a permanent component of national defense infrastructure. Maven utilizes AIP for complex sensor fusion and computer vision across the battlefield. This move ensures that the "tokens" consumed for situational awareness are now a permanent, multi-year line item in the defense budget. This permanence signifies that the economic center of gravity has shifted: the military no longer just buys hardware, it subscribes to the vectorized intelligence required to operate it.
These global examples clarify the paradox for public procurement and corporate strategy alike. When we build our future on a proprietary "railway" where every decision is a metered event, we are trading the autonomy of Capital Expenditure for a perpetual, variable Operational Expenditure on cognition. The risk for sovereign entities is not just the cost of the coal, but the reality that they no longer own the tracks, the engine, or the engineering standards that define the journey.
Each query, each workflow, each agent action consumes tokens. When public budgets fund this consumption, the question is not merely technical: Is this the most efficient, transparent, and sovereign way to harness AI for the public good?
The term sovereign AI has gained currency across Europe, generally defined as a nation's ability to develop, host, and govern AI systems using its own infrastructure, data, and workforce. By this definition, what Palantir delivers is something closer to captive AI: a system in which the client owns the data but not the means of its interpretation, where switching costs become prohibitive, and where the vendor's ontology (the very framework for decision-making) becomes an infrastructural lock-in.
If sovereignty were automatic, so many contractual safeguards would not need to be spelt out. The NHS contract is a case in point. NHS England insists that patient data is protected, retained in the UK, fully audited, and that Palantir cannot commercialize NHS data or use it to train AI models. These safeguards are important. They also prove the point: true sovereignty does not require this density of defensive contractual language.
The Spanish researcher Alejandro Pozo has warned that "one of the pillars of Spain's sovereignty is already being handled by a United States company." As Genís Roca, a technology consultant, has noted, Palantir is placed "at the heart of defense" and, once embedded, becomes "critical infrastructure"... extremely difficult to displace.
Europe is already drawing conclusions from these risks. Switzerland formally rejected Palantir after an internal Swiss Armed Forces report concluded that sensitive military data could potentially be accessed by US government intelligence agencies. Swiss agencies rejected Palantir at least nine times over seven years. Denmark has embarked on building its own alternatives to Palantir's surveillance and data-analysis platforms, including Maven Smart System and Foundry, reflecting a growing European assessment that Palantir represents a security risk.
When a government contracts Palantir to "modernize defense" or operate the NHS data spine, is that sovereign AI? Or is it captive AI, that is, a structural dependency in which the vendor's operating grammar becomes the grammar of the state?
Captive AI arises when the workflow, the ontology, the institutional memory, the audit trail, the approval logic, the cost meter, and the habit of decision-making all accumulate within a single operational platform. At that point, replacing the vendor is no longer an IT project. It becomes a cognitive reorganization of the institution.
This is the distinction Europe must make. Sovereign AI is not a flag on a data center. It is the ability to continue operating, auditing, adapting, and exiting without institutional paralysis.
I want to be clear: Palantir's engineering deserves serious admiration. As a former NHS lead acknowledged, the company is "uniquely suited to the messy NHS data problems that have been accumulating over the last 25 years." The platform saw earlier than most that the enterprise AI problem was not the model alone, but the operating layer around it (permissions, data structures, provenance, cost control, and human authorization). In that sense, Palantir is closer to an industrial AI company than to a conventional software vendor.
Yet engineering excellence does not confer moral neutrality, and the ethical questions surrounding Palantir are neither marginal nor manufactured.
The company has been repeatedly criticized for its work with US Immigration and Customs Enforcement (ICE). In April 2025, ICE awarded Palantir a $30 million contract for Immigration OS, a system that tracks "self-deportations," identifies priority deportation cases, and supports the "Catch and Revoke" initiative. Amnesty International has documented how these tools enable "constant mass monitoring, surveillance, and assessments of people" and has condemned Palantir's "track record of flagrantly disregarding international law."
In April 2026, the unease deepened when Palantir published a 22-point corporate manifesto. CEO Alex Karp criticized the belief that all cultures are equal, called for universal national service (including compulsory military enlistment) described the post-WWII disarmament of Germany and Japan as an "overcorrection," and backed AI-powered weapons systems. The manifesto prompted Prof. Shannon Vallor, chair of ethics of data and AI at Edinburgh University, to warn that "every alarm bell for democracy must ring."
When a company that builds the data infrastructure for public health and national security advances a worldview of civilizational hierarchy and permanent military readiness, it is reasonable to ask whether its platform is merely a neutral tool or an instrument that encodes a particular vision of power.
The alternative is not to reject operational AI. It is to build it on different foundations.
Gartner predicts that by 2027, organizations will use small, task-specific AI models with a usage volume at least three times greater than general-purpose large language models, driven by the need for contextualized, reliable, and cost-effective solutions. This is the real industrial revolution underway. We are moving (or should be moving) away from the "coal" era of burning massive amounts of generic compute toward a "precision engine" era. Custom SLMs do not require the proprietary furnaces of captive AI vendors; they require high-quality, aligned data, secure deployment environments, and institutions that retain control over their own intelligence.
Tokens may be the new coal. But Europe should not merely decide whose coal it burns. It should decide which engines it builds, which data it trusts, which models it adapts, which languages it protects, which public values it encodes, and which institutions remain capable of governing their own intelligence.
This is precisely where Pangeanic's mission occupies a fundamentally different space.
Where Palantir thrives on proprietary enclosure, Pangeanic builds for sovereignty and ethical AI from the ground up. Our mission is to provide the sovereign infrastructure that makes the world's intelligence truly multilingual and accessible. Every deployment of its Deep Adaptive Neural Machine Translation and the ECO Intelligence Platform ensures that clients, whether media broadcasters, public administrations, or regulated enterprises, retain full ownership over their data, models, and decision-making frameworks, without ever surrendering privacy, auditability, or control to a third party.
Pangeanic's approach rests on four operational pillars: Transparency, Fairness, Privacy, and Accountability. Its proprietary technology—combining custom Small Language Models, secure data foundries, Retrieval-Augmented Generation, cross-lingual search, AI-powered anonymization, and rigorous human-in-the-loop workflows—is designed to meet the exacting demands of regulated sectors without creating captive dependencies.
The evidence that this alternative is real, not rhetorical, is already visible. Pangeanic's collaboration with the Barcelona Supercomputing Center (providing data annotation, RLHF, and training data support for the Salamandra and ALIA language models) demonstrates sovereign AI as a public, multilingual, technically grounded effort, not a branding exercise. Its ECO Platform supports deployment in private cloud, on-premise, or air-gapped environments where data sovereignty and infrastructure control are non-negotiable. It operates across more than 200 languages.
The age of "tokens as coal" will produce many powerful locomotives. But the question that ultimately matters is not how fast the train goes. It is who controls the tracks, who owns the fuel, and whether the destination is chosen by the passenger or the driver. In that decisive respect, Pangeanic and Palantir do not merely follow different business models, they represent two irreconcilable philosophies of what AI should become in the life of a democratic society.
Pangeanic — Our Mission | ECO Intelligence Platform