Nvidia's Tokenization of the World: How Jensen Huang Perfected the 21st Century Oil Lamp Strategy

Billions for empty promises? The inconvenient truth about Nvidia's token factory

Nvidia CEO Jensen Huang has formulated a simple equation: those who don't calculate, lose. But behind the glittering facade of the AI ​​boom lies a ruthless business model reminiscent of the unscrupulous monopoly strategies of the 19th century. With an unprecedented hardware monopoly, the closed CUDA software ecosystem, and frontal assaults like the new RTX Spark chip, the tech giant is forcing the global economy into a dangerous dependency. Instead of measurable productivity, companies today are primarily buying one thing: the sheer consumption of "tokens." This is an in-depth analysis of how Nvidia is turning the rules of value creation on their head, why hyperscalers have to invest hundreds of billions – and why this spiral of profit and energy waste could cost us all dearly.

NVIDIA and the tokenization of the world: How Jensen Huang is dictating (and profiting from) a new economic order

The moment when teleshopping became a corporate strategy

In March 2026, Jensen Huang took to the stage at the Morgan Stanley Technology, Media & Telecom Conference in San Francisco and uttered a sentence that is unparalleled in its brevity and audacity: “Compute equals tokens, tokens equal intelligence, and intelligence equals economic output at every level, from companies to countries.” What reads like a fundamental physical equation is in reality one of the most ambitious marketing constructs in economic history: the reinterpretation of a data center as a printing press that produces profit—primarily for NVIDIA.

A few weeks earlier, at Computex 2026 in Taipei, Huang added to this picture with the RTX Spark, an ARM-based system-on-a-chip for Windows laptops and compact desktops. The narrative was already familiar: those who don't buy fall behind. Consumption itself is proof of economic activity. "The more you buy, the more you make"—a phrase that, in its beautiful simplicity, distills the entire logic of a business model based on the structural dependence of its customers.

To understand why this logic is so dangerous, it is worth taking a look back at the history of oil lamps.

The Oil Lamp Principle: How to Give the Gift of Dependence

Toward the end of the 19th century, John D. Rockefeller's Standard Oil Company spread a simple but revolutionary technology throughout American homes: the kerosene lamp. The lamp itself was cheap, sometimes even free. The oil it needed to run it was not—and without that oil, the lamp was worthless. By 1879, Standard Oil controlled roughly 90 percent of U.S. refining capacity, thus dictating the price of the only fuel that kept the lamps burning. The catch wasn't the lamp itself. The catch was the resulting system: once you switched to kerosene, there was no going back. You kept buying it—until the end of your days or until the Supreme Court ruled.

NVIDIA has carried this principle into the digital age, building on 17 years of patient work. Since 2007, the company has been developing its proprietary programming platform CUDA, now the de facto operating system of the global AI industry. With over 5 million registered developers, some 5,937 GitHub projects related to CUDA alone (compared to 187 for AMD's competing product ROCm), and virtually every relevant AI library—from cuDNN and TensorRT to the frameworks PyTorch and TensorFlow—NVIDIA has created a software chasm that cannot be bridged with capital alone. The lightbulb is called CUDA. The oil is called compute. And once you've entered the ecosystem, there's no way out.

This is clearly demonstrated by the story of the open-source project ZLUDA, which made it possible to run CUDA code unchanged on AMD hardware. When the threat became real, NVIDIA quietly and without consultation changed the terms of service for the CUDA platform: translation layers were prohibited via EULA. No court, no fair competition—just a contract clause that stifled a genuine alternative in its infancy.

The Token Factory: A New Paradigm of Value Creation

The term "AI Factory" is not a metaphor; it's a mission statement. At the GTC conference in March 2026, Jensen Huang explicitly defined what he meant by it: data centers are no longer passive infrastructure facilities, but active production plants whose output—measured in tokens per second—can be directly translated into company revenues and gross domestic product. The token is the new barrel unit of the digital commodity.

What initially sounds like a plausible systematization, upon closer inspection, represents a fundamental shift in the attribution of value. Traditionally, economic value is measured by the result: Was a problem solved? Was a product built? Was revenue generated? In Huang's framework, value arises from the computation itself—regardless of whether the token contributes to solving a real problem or becomes expensive idle time. This calculation holds true for NVIDIA and the hyperscalers because they profit from every token created. For the end customer, the opposite is true.

Agentic AI, meaning systems that autonomously plan, research, and execute, can consume a million times more tokens than a standard prompt, according to Huang. This isn't a description of an efficiency revolution. It's a description of an exponentially growing operating expense. Those who deploy AI agents on a large scale aren't buying productivity—they're buying token consumption, the value of which has yet to be proven in real economic results.

Monopoly power: Numbers that silence

NVIDIA's position in the AI ​​hardware market is no longer market dominance in the traditional sense. It's a structural fact that even experienced capital market observers find cause for caution. In the fourth quarter of fiscal year 2026 (November 2025 to January 2026), NVIDIA achieved quarterly revenue of $68.1 billion, representing year-over-year growth of 73 percent. The data center business accounted for 91.5 percent of total revenue, and the adjusted operating margin climbed to 67.7 percent.

For comparison: Software companies, known for their high margins, rarely achieve values ​​above 40 percent. NVIDIA, formally a hardware company, generates margins that would be exceptional even for platform companies—an indication that its real competitive advantage lies in its software ecosystem, not in silicon. According to a Handelsblatt analysis, CUDA is the true operating system of the AI ​​industry, and NVIDIA's greatest competitive advantage lies in its code, not its chip.

In the discrete graphics card market, NVIDIA will hold a 94 percent market share as of the fourth quarter of 2025, according to data from Jon Peddie Research. AMD will have five percent, and Intel one percent. The share in the AI-specific GPU market is comparable. In the wafer production sector used for AI chips, NVIDIA is projected to secure a 77 percent share in 2025, according to a Morgan Stanley analysis—compared to 51 percent the previous year.

This concentration is not a law of nature, even if Huang likes to describe it as such. It is the result of a years-long strategy based on technological superiority, targeted market segmentation, and the building of an ecosystem in which switching costs for customers are so high that even massive price increases are accepted without complaint.

The flow of capital: Who pays the bill?

The true extent of NVIDIA's dependence isn't revealed in the company's own figures, but rather in the capital expenditure plans of its most important customers. The five largest American hyperscalers—Amazon, Alphabet, Microsoft, Meta, and Oracle—have announced combined capital expenditures of $660 billion to $690 billion for 2026, nearly double the previous year's figure. Of this, roughly 55 to 60 percent flows directly or indirectly to NVIDIA.

Amazon alone has announced investments of $200 billion for 2026—a sum exceeding Portugal's annual gross domestic product. Alphabet's capital expenditures are projected to increase from $91 billion to $180 billion, a rise of 98 percent. Microsoft is increasing its data center budget by 59 percent year-over-year. These expenditures are no longer financed solely from free cash flow. Amazon's free cash flow is expected to become negative by $17 billion to $28 billion in 2026, Meta's free cash flow is projected to fall by nearly 90 percent, and Oracle is expected to post a negative free cash flow by 2030.

Who ultimately pays? The hyperscalers pass on the costs through price increases. In January 2026, AWS raised prices for H200 GPU instances by 15 percent—a reversal of two decades of falling cloud computing prices. Enterprise customers who obtain AI services via the cloud are thus directly paying the price for NVIDIA's monopoly.

AllianceBernstein estimates that NVIDIA retains around 30 percent of total AI data center spending as profit. This means that for every euro a European company spends on cloud AI services, roughly 30 cents flow to an American company—without any requirement for a return on investment in the form of problem-solving, innovation, or societal benefit. The token is produced. That's enough.

Waste as a key performance indicator: The perverse logic of productivity

Jensen Huang has stated at events that he finds it deeply concerning if a well-paid software developer does not incur at least a quarter of a million US dollars in token costs per year. This statement is often cited in tech media as evidence of Huang's vision, but rarely examined for its economic substance.

A quarter of a million US dollars in token costs is not a productivity metric. It's a consumption metric. The crucial difference: Productivity measures output per input. Consumption measures only input. By elevating token consumption to a management metric, Huang breaks with one of the oldest principles of business administration: It's not the use of resources that creates value, but the result.

Practice, in a way, proves Huang right—but in a way that harms companies. Companies like Zapier already systematically track their employees' token consumption. Anyone who uses five times as many tokens as average is internally scrutinized for their usage patterns. What began as cost control threatens to become a new form of performance measurement mania, where employees learn to submit pointless prompts to avoid slipping in the internal rankings. Consumption becomes a demonstration of performance, waste a form of self-defense.

A recent Bitkom survey of 604 German companies reveals that a third of firms using AI have already been surprised by the costs involved. Bitkom President Ralf Wintergerst confirmed that many companies report AI agents require more support from traditional employees than initially anticipated. Brian Jabarian of the University of Chicago sums it up: "Everyone thought you simply deploy AI tokens, see a productivity increase, and that's it. But reality is more complicated."

The productivity lie and its methodological weaknesses

NVIDIA's core argument for the economic viability of its platform is the claim that AI triples productivity. This figure has a methodological limitation that is rarely discussed in the public debate: it is based almost exclusively on observations in the field of software development—precisely the professional group that benefits most from AI tools, possesses the technical expertise for optimal use, and already works extensively with digital tools.

The Institute for Employment Research (IAB) assumes that the overall impact of AI on the German labor market is real, but considerably more unevenly distributed than Huang's presentation suggests: Around 800,000 jobs could be lost due to AI, while at the same time around 800,000 new ones would be created—with an overall economic productivity increase of up to 0.8 percentage points per year. This figure is economically significant, but far from tripling.

The “European Growth Study 2026” by the strategy consultancy Simon-Kucher, based on 1,236 company interviews in 13 European countries, concludes that 73 percent of companies currently use AI in less than 30 percent of their processes—and expect noticeable productivity or employment effects only at a penetration rate of 30 to 50 percent. A labor market analysis by the Bertelsmann Foundation, based on approximately 60 million job postings, finds that the share of AI-related jobs has stagnated at an already low level since 2022 and even declined slightly in 2023 and 2024.

This does not mean that AI has no economic impact. It means that the impact is selective and unevenly distributed, and arrives far more slowly than propagated by industry—while the costs are incurred immediately.

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

• The Managed AI Solution - Industrial AI Services: The Key to Competitiveness in the Services, Industry and Mechanical Engineering Sectors

The RTX Spark maneuver: Destroy the market and sell the solution

One of the most interesting aspects of NVIDIA's current strategy is the introduction of the RTX Spark. Announced on May 31, 2026, at Computex in Taipei, the chip combines a 20-core ARM processor based on Grace architecture with a Blackwell GPU featuring 6,144 CUDA cores and up to 128 GB of shared LPDDR5X memory. It boasts up to one petaflop of AI computing power. Among the first devices to utilize it is Microsoft's Surface Laptop Ultra.

On the surface, this appears to be a reaction to Apple's M-series chips, which have dominated the premium laptop market with efficient ARM processors in recent years. However, a deeper look reveals something else: NVIDIA's massive demand for AI GPUs in data centers has significantly contributed to the scarcity and increased cost of memory chips, putting immense pressure on the traditional PC market. The entire GPU market (including integrated graphics solutions) shrank by 12 percent to 68.8 million units in the first quarter of 2025. And now NVIDIA is launching a premium ARM PC, effectively declaring the conventional desktop PC obsolete.

The pattern is familiar: An established market is destabilized by external factors. Then a vendor appears with a solution to the problem they themselves helped create—naturally, at premium prices. The RTX Spark is explicitly aimed at the high-end market. Exact prices haven't been announced, but industry observers expect significant markups compared to similar devices with Intel or AMD processors. Those who enter this new ecosystem abandon the x86 standard and thus become dependent on ARM, a dependency further reinforced by the proprietary CUDA ecosystem. In the future, users will be able to generate their own tokens—on Huang's hardware, with Huang's software, according to Huang's rules.

Machines producing for machines: The circular argument of an economy

In the most radical form of his vision, Huang describes a world in which AI agents provide services for other AI agents, which in turn depend on AI infrastructure that is monitored by further agents. Economic activity is self-sufficient—it no longer needs a human end-use to be measurable, as long as the tokens keep flowing.

This circular reasoning has an elegant internal logic for NVIDIA, but a worrying one for the rest of the economy. If tokens are considered a proxy for economic activity, then every token created justifies further investment in infrastructure that generates more tokens. The result is a spiral in which compute investments are legitimized by token output, the actual economic benefit of which remains unclear. For the technology sector, this is a flywheel. For the wider economy, it could prove to be a new version of the crowding-out effect: capital flowing into token factories is unavailable for productive investments in manufacturing, infrastructure, education, or healthcare.

The hyperscaler figures make it clear: Amazon's free cash flow is expected to turn negative in 2026, and Meta's will fall to almost zero. This capital commitment is not a sign of sound economic judgment—it's the result of an arms race in which no one can opt out without losing market share. Those who don't buy fall behind. Those who do buy subsidize NVIDIA's margins.

The environmental dimension: The invisible third party in the equation

An economic analysis of the token economy that ignores the environmental costs would be incomplete. Global electricity consumption by AI data centers will increase from 50 billion kilowatt-hours in 2023 to approximately 550 billion kilowatt-hours in 2030—an elevenfold increase. This will be accompanied by a rise in greenhouse gas emissions from data centers from 212 to 355 million tons of CO₂ equivalent, despite the parallel expansion of renewable energy sources.

In a report commissioned by Greenpeace Germany, the Öko-Institut (Institute for Applied Ecology) concludes that data centers will continue to rely heavily on fossil fuels in the coming years because local power grids are reaching their capacity limits. The IMF quantifies the combined share of AI data centers and cryptocurrencies in global electricity consumption at two percent for 2023, with a projected increase to 3.5 percent by 2027. A ChatGPT query consumes three to ten times more electricity than a conventional Google search.

These costs don't appear on any NVIDIA balance sheet. Nor do they appear in the pricing of a token. They are externalized costs—borne by energy consumers, climate systems, and future generations. In economic terms, these are significant negative externalities that systematically subsidize the token economy's business model without any transparency.

CUDA as a standard oil: The analogy and its limits

The historical comparison between Rockefeller's Standard Oil and NVIDIA's CUDA platform has a real analytical basis, but it also goes beyond it. Standard Oil controlled pipelines and refineries—physical infrastructure that could, in principle, be duplicated, albeit with enormous capital expenditure. Its breakup in 1911 was possible because the facilities already existed and could be divided among 34 successor companies.

CUDA is more difficult to divide. It's not a pipe you can simply cut open. It's an ecosystem of millions of lines of code, libraries, documentation, developer expertise, and network effects, built over 17 years. A CUDA translation layer that makes code executable on AMD hardware is contractually prohibited. Open-source alternatives like ROCm or OpenCL lag behind with a fraction of the reach and market maturity. The $12.9 billion R&D budget that NVIDIA is pouring back into its own ecosystem in fiscal year 2025 buys every new performance advantage before a competitor can catch up.

At the same time, NVIDIA's strategy with open-weight models is particularly subtle: The company is investing $26 billion over five years in the development of open AI models—models that anyone can use for free. But NVIDIA's Nemotron models are trained in NVIDIA's proprietary 4-bit NVFP4 format and only unleash their full performance advantage on Blackwell hardware. It's like giving away the oil lamp, but supplying the oil from only one refinery.

Counterforces and structural limits of dominance

It would be analytically dishonest to portray NVIDIA's position as immutable. Real counterforces exist, though their strength is often overestimated. Google's TPUs, Amazon's Trainium, Meta's MTIA, and Microsoft's Maia are serious internal alternatives that have reduced NVIDIA's share of hyperscaler capex from around 70 percent in 2023 to an estimated 55 to 60 percent in 2026. AMD's MI300 and MI400 series is gaining market share, particularly in certain inference workloads.

But this decline from 70 to 55 percent is occurring amidst a massive overall market growth. In absolute terms, NVIDIA's revenue continues to rise. Hyperscalers are building their own chips because they know and fear their dependence on NVIDIA—but they can only diversify the market to the extent that CUDA-compatible alternatives are mature enough to handle production workloads. That point is still a long way off.

DeepSeek from China demonstrated in early 2025 that significant efficiency gains are possible by achieving comparable model quality with a fraction of the computational effort. The Hasso Plattner Institute cites DeepSeek as achieving the same training accuracy with one-hundredth of the energy expenditure of conventional methods. Should this efficiency logic prevail, the demand for raw compute volume would structurally decline—putting pressure on NVIDIA's token volume model. Huang has recognized this threat and is positioning efficiency—measured in tokens per watt—as the new CEO-level decision parameter. Again, the message is clear: Buy more, but buy more efficient machines—from NVIDIA.

Regulation: Is antitrust law coming too late?

The question of whether NVIDIA's market position warrants antitrust action is being increasingly debated in Brussels and Washington. The comparison with Standard Oil is more than just rhetoric: Back then, Rockefeller controlled the American oil industry with a 90 percent market share before the court ruling in May 1911 led to its breakup into 34 successor companies. EU competition authorities have at least established a regulatory framework with the Digital Markets Act and the AI ​​Act. However, direct intervention against NVIDIA's CUDA ecosystem is still pending.

The conceptual problem is well-known: unlike physical networks such as pipelines or railway lines, a software ecosystem cannot be easily opened up through regulatory intervention. Interoperability requirements, i.e., the obligation to grant CUDA alternatives the same hardware access, would theoretically be feasible—but in practice, costly and technically complex. Moreover, any regulatory measure would have to be implemented quickly enough to alter a market structure that is becoming increasingly entrenched daily through new model generations, new hardware architectures, and new vendor lock-in effects.

Until then, the following applies: Anyone who invests in data centers, uses cloud AI services, or trains their developers on CUDA-based frameworks is paying—directly or indirectly—NVIDIA's monopoly profits. This is not a conspiracy theory. This is the structure of a market in which a single vendor controls 94 percent of the discrete graphics card segment, 77 percent of AI chip wafer production, and virtually all relevant software libraries for AI development.

When consumption becomes an end in itself

Jensen Huang's formula—compute is revenue, tokens are profit—is one of the most honest corporate strategy statements of recent years. Honest not in the sense that it's formulated for the benefit of customers, but in the sense that it states what many others keep silent about: The business model is not based on the value generated at the end of a computational process, but on the process itself.

This is a fundamental reversal of the value creation logic. In every other industry, price is defined by the result: a built bridge, a developed drug, a sold car. In the token economy, price is defined by the input: the computing hours consumed, the electricity flowing, the processed data packets. NVIDIA earns money before anyone can assess whether the investment is worthwhile.

This isn't a law of nature. It's a business model. And like any business model, it has limits, weaknesses, and—with a little patience—alternatives. The question is whether companies, regulators, and the public will recognize and promote these alternatives quickly enough before the dependency becomes as deeply ingrained as the oil lamp once was in the American home. It took Rockefeller's Standard Oil from 1870 to 1911 to be broken up. This time, the flywheel is spinning faster.

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

 

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here wolfenstein@xpert.digital:or simply call me at +49 7348 4088 965. My email address is

I'm looking forward to our joint project.

 

 

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business