Nvidia is worth more than Spain's economy

Today we talk about the figure that best sums up the mess of priorities in the contemporary global economy. An American company founded in 1993 by three engineers in a Denny's in San Jose is worth, as of 2025, around three and a half trillion dollars in market capitalization, with peaks above four trillion. Spain's annual GDP in 2024 was, according to INE data, roughly 1.6 trillion dollars. A single company is worth more than double the entire Spanish economy put together. How do you get there? My take is biased because I held Nvidia shares a decade ago, sold them early, and it still hurts. Form your own: the story is longer and less spectacular than the figure.

I've spent years explaining Nvidia's story at dinners with non-technical friends. The part that surprises them most is that for fifteen years it was a losing bet and little more. The current capitalization isn't the result of a decade, it's the result of three. Let's walk through the three in detail, because each one explains a bit of the figure.

The first twenty years, gaming and survival

Jensen Huang, Chris Malachowsky and Curtis Priem founded Nvidia in April 1993 in California. The story they like to tell is that the founding meeting was at a Denny's in East San Jose, sitting at a plastic table drinking American coffee. The exact date and the exact table matter little; the intention does. They wanted to make graphics chips for PCs. In those years, the market was dominated by 3Dfx, ATI, S3 and half a dozen smaller companies. Nvidia was one of them.

For the first five years, the company came close to bankruptcy several times. Its first product, the NV1, failed commercially. The NV2, commissioned by SEGA for a console, was canceled halfway through development. What saved them was the launch of the RIVA 128 in 1997 and, above all, the GeForce 256 in 1999, marketed under the category they invented themselves: the Graphics Processing Unit, or GPU. That nominal decision —calling GPU what the industry called a graphics accelerator— had more consequences than it seems. The word stuck.

Between 1999 and 2005, Nvidia consolidated as a respectable gaming company. Its main rival was ATI, later acquired by AMD in 2006. The market was stable but limited. The shares traded around 20 dollars and the capitalization hovered around 6 billion dollars. A decent company, not a spectacular one. In those years, if anyone had predicted Nvidia would come to be worth more than all of Spain put together, they'd have been called crazy.

The CUDA turn and the lost decade

In 2006, Nvidia made a decision the sector considered eccentric. It released CUDA —Compute Unified Device Architecture—, a software platform that let its GPUs be used not just for graphics but for any kind of parallel computation. The technical idea was simple. A GPU contains thousands of small cores designed to process pixels in parallel. Any mathematical operation that can be parallelized —linear algebra, physical simulation, signal processing— can benefit from that architecture.

The investment in CUDA was enormous. Nvidia partly redesigned its chips so that general-purpose compute would be efficient, published extensive documentation, trained developers, funded university courses and doctoral grants. The commercial return was nearly zero for seven or eight years. Some academic researchers began using CUDA for simulation problems, molecular dynamics, computational finance. None of those uses was massive enough to justify the investment.

Wall Street had spent years criticizing the bet. Analyst reports described Nvidia as a gaming company distracted by academic projects. The quarterly earnings calls included skeptical questions about when CUDA would generate revenue. Jensen Huang —at the helm since the founding— defended the investment as a long-term vision. The shares moved between 10 and 30 dollars for nearly a decade. The market wasn't buying the bet.

The year 2012 and the change of order

In September 2012, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton published ImageNet Classification with Deep Convolutional Neural Networks (NeurIPS, 2012), the paper where they introduced AlexNet. The convolutional neural network won the ImageNet contest by a vast margin over the runner-up, cutting the classification error from 25% to 15%. AlexNet had been trained on two Nvidia GTX 580 GPUs. The feat wouldn't have been possible without those GPUs, because training a network that size on CPU would have taken weeks instead of days.

That moment has been told a thousand times as the start of modern deep learning. What's told less, and worth holding on to, is what it meant for Nvidia. CUDA and Nvidia's GPUs became, overnight, the standard hardware for any serious deep-learning project. Every AI paper published since late 2012 begins, in its methods section, citing "trained on Nvidia GPUs." That citation is free, sustained advertising over an entire decade.

But the transformation wasn't immediate in the figures. Between 2012 and 2017, Nvidia remained, above all, a gaming company, with a growing but still minority business in data centers and AI. The shares began to rise steadily, going from 15 dollars in 2012 to 50 dollars in 2017. The capitalization was already around 30 billion dollars. Reasonable, not spectacular.

The real acceleration began around 2018-2019 with the consolidation of deep learning in commercial production. The big tech companies —Google, Microsoft, Facebook, Amazon, Apple— began buying Nvidia GPUs en masse to train and serve models. Prices rose. Margins rose. Nvidia's market share in the data-center segment went from 50% to over 80%.

The post-ChatGPT takeoff

ChatGPT came out on November 30, 2022. Nvidia's stock closed that day around 169 dollars. Fifteen months later, in February 2024, it had topped 800 dollars. In June 2024 the company did a 10-for-1 split, which adjusted the nominal price but not the capitalization. By the end of 2024, the capitalization had topped 3 trillion dollars. In spring 2025, at some isolated peaks, it hovered around 4 trillion.

The specific trigger was the demand for H100 GPUs and, later, H200 and B200 to train frontier models. Every large model trained between 2023 and 2025 —GPT-4, Claude 3.5, Gemini, Llama 3, the Chinese models from Alibaba and DeepSeek— consumed tens of thousands of Nvidia GPUs over weeks or months. The unit price of an H100 runs around 25,000 to 35,000 dollars depending on the integrator. Nvidia's gross margin on each chip exceeds, according to consistent estimates from SemiAnalysis and its own financial reports, 70%.

The combined multiplier —massive demand, limited supply, high margin— produced the highest quarterly profits in the history of the tech sector. The data-center segment's revenue went from 15 billion dollars in fiscal year 2023 to over 100 billion in fiscal year 2025. The stock followed that trajectory. The capitalization followed the stock.

The absurd figure in perspective

For the capitalization figure to make sense, it helps to compare it with references the reader can picture.

Nvidia's capitalization at its 2025 peak: around 3.5 to 4 trillion dollars. Spain's annual GDP in 2024 according to the INE: 1.6 trillion dollars. A company is worth, in accounting terms, double the annual sum of all Spanish economic activity. The comparison has nuances —capitalization is a projection of future profits, not a direct equivalent of annual GDP— but the asymmetry is hard to play down.

Other useful comparisons. Nvidia's capitalization in 2025 exceeds the GDP of Italy (2.3 trillion), that of Brazil (2.1 trillion) and the combined GDP of the Netherlands, Belgium and Switzerland. It's higher than the combined annual military spending of the United States, China, Russia, the United Kingdom, France and Germany. When it approaches the four-trillion figure, it even exceeds Germany's GDP (4.5 trillion), though not quite reaching it.

What that magnitude expresses isn't only Nvidia's value. It's the brutal concentration of capital the AI sector is producing around a single hardware provider. The seven most valuable American companies —Apple, Microsoft, Nvidia, Alphabet, Amazon, Meta and Tesla— together add up to more than 20 trillion dollars in capitalization, more than the European Union's combined GDP. That concentration isn't new in absolute terms —there have always been large companies— but it is in relative terms to other eras: never before had they added up to so much.

The least-told piece, the underlying fragility

Here comes the observation the headline doesn't capture. Nvidia's capitalization rests, ultimately, on three structural dependencies the company itself doesn't control.

The first dependency is TSMC. As we saw in 0071, Nvidia designs the chips but doesn't make them. Its ability to serve demand depends absolutely on TSMC's capacity in Taiwan. A serious interruption at TSMC turns Nvidia's inventory of designs into wet paper for months or years.

The second dependency is ASML, via TSMC. If ASML doesn't deliver EUV machines, TSMC can't produce the advanced nodes Nvidia's GPUs require. That dependency is indirect but structural.

The third dependency is the AI investment cycle. Nvidia's record capitalization rests on the market's implicit assumption that demand for AI GPUs will keep growing at high rates for several years. If that demand moderates —through market saturation, through insufficient ROI as McKinsey's State of AI 2025 shows, through algorithmic innovation that reduces hardware dependence as DeepSeek suggested—, the capitalization would adjust downward with the same speed it rose.

The DeepSeek case of January 2025 was a first test of that fragility. In a single trading session —January 27, 2025—, Nvidia lost 589 billion dollars in capitalization, the largest single-day loss in the history of the U.S. market. The trigger was the perception that the technical frontier could move with algorithmic efficiency, not just with hardware. The recovery that followed was fast, but the warning stood.

The political question

This is personal opinion, but I hold it on the very nature of the bottleneck. The concentration of value in Nvidia isn't sustainable indefinitely and isn't healthy for the sector or the global economy. It's the product of three combined factors: a visionary technological bet (CUDA), a temporary execution advantage, and an extraordinary demand badly calibrated as to its real returns. All three factors have an expiry date.

The CUDA bet will sooner or later be partly replicated by alternatives —AMD ROCm, Intel oneAPI, open frameworks like OpenAI's Triton. The execution advantage will sooner or later erode as competitors build similar capacity. The extraordinary demand will sooner or later adjust when the ROI of the investments turns out tangible or false. The three corrections combined could produce a serious market correction at some point in the next five years.

This isn't a prediction of bankruptcy. Nvidia will almost certainly remain a very important company in the tech sector. What probably won't be sustainable is that it's worth double Spain's GDP. That particular figure is a symptom of the current euphoria, not of long-term structural value.

For the European reader, this figure carries an additional, uncomfortable reading. While the United States has produced in Nvidia a company worth double Spain, Europe has, in its combined capitalization, no company comparable in the AI sector. The most valuable European company by capitalization in 2025 —ASML— is worth around a tenth of Nvidia. The most valuable Spanish company —Inditex— is worth around a twentieth of Nvidia. That asymmetry isn't accidental; it's the product of industrial and regulatory decisions accumulated over decades.

The concrete figure that closes it. According to Nvidia's quarterly 10-K reports filed with the SEC in February 2025, fourth-quarter fiscal 2025 revenue was 39.3 billion dollars, with a 61% operating margin. That magnitude, equivalent to 156 billion dollars annualized, already exceeds the Spanish State's general budget approved for 2024 (a little over 200 billion euros, but funding an entire country of forty-eight million people and its healthcare, education, military and pension systems). A single American company bills, on a quarterly basis, a significant fraction of the annual budget of a mid-sized European state. That's the real dimension of the figure.

Definitions

Market capitalization: the total market value of a listed company, calculated as price per share multiplied by the number of shares outstanding. It measures the valuation the market gives the company at any given moment.

CUDA (Compute Unified Device Architecture): Nvidia's proprietary software platform for programming general-purpose compute on GPUs. Launched in 2006, it's the foundation of Nvidia's current dominance in AI.

GPU (Graphics Processing Unit): a chip specialized in parallel computation, originally designed for graphics. Modern GPUs are the standard tool for training and running AI models.

Operating margin: the percentage of revenue left as profit before interest and taxes. Nvidia's in its data-center segment regularly exceeds 60%, an extraordinarily high figure for the tech sector.

References

Stephen Witt, The Thinking Machine: Jensen Huang, Nvidia, and the World's Most Coveted Microchip (Viking, 2025). A recent corporate biography of Jensen Huang and Nvidia's trajectory.

Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton, ImageNet Classification with Deep Convolutional Neural Networks (NeurIPS, 2012). The foundational paper of modern deep learning, trained on Nvidia GPUs.

Nvidia Corporation, Annual Reports 10-K (SEC filings, 1999-2025). Official documentation on the evolution of revenue, margins and business segments.

Chris Miller, Chip War (Scribner, 2022). Chapters on Nvidia in the context of the semiconductor chain.

Clayton M. Christensen, The Innovator's Dilemma (Harvard Business Review Press, 1997). Theoretical framework for understanding why established companies lose their advantage and others like Nvidia gain it.

Bloomberg, coverage of Nvidia's stock price (2022-2025). Historical price and capitalization data available at bloomberg.com.

Instituto Nacional de Estadística (INE), Annual National Accounts of Spain (2024). The Spanish GDP figures used in the comparison.

Going deeper

SemiAnalysis (semianalysis.com). Recurring technical analysis of Nvidia's business, margins and competitive position.

Andrew McAfee & Erik Brynjolfsson, Machine, Platform, Crowd (W. W. Norton, 2017). An economic framework on how value concentrates in platform companies.

Carlota Pérez, Technological Revolutions and Financial Capital (Edward Elgar, 2002). A historical framework for placing Nvidia's current valuation within the classic pattern of tech bubbles.

Cade Metz, Genius Makers (Dutton, 2021). A narrative history of deep learning with sections on the transformation ImageNet 2012 produced at Nvidia.

You may also like

• Why OpenAI wouldn't exist without a factory in Taiwan
• ASML, the Dutch company ChatGPT, Claude and the rest depend on
• Why the United States banned selling chips to China
• The real cost of training GPT-5

Elsewhere

• Europe needs a strategy to close the artificial intelligence compute gap (Bruegel)
• Semiconductors — Center for Security and Emerging Technology (Georgetown)