The Geometry of Power
On a spring morning in 1993, three engineers met in a San Jose diner and decided to build a company around a hunch: that the future of computing would be visual, three-dimensional, and hungry for parallel processing power . Jensen Huang, Chris Malachowsky, and Curtis Priem founded Nvidia on 5 April of that year, naming it after the Latin word for envy — a curious choice for a firm that would, three decades later, become the object of Wall Street's most feverish desire .
By January 2025, that diner-table startup had become the first corporation in human history to cross both the four-trillion-dollar and five-trillion-dollar market capitalisation thresholds, its value now sitting at $5.01 trillion . To put that in perspective: Nvidia is worth more than the entire GDP of Japan, more than the combined value of every publicly traded company in France. It has become, in the bluntest financial terms, the most valuable enterprise ever assembled.
But valuations, however dizzying, are merely symptoms. The more interesting question is what Nvidia has become in substance — and the answer is both simpler and stranger than the trillion-dollar figure suggests. Nvidia has transformed itself from a graphics-chip maker into the physical substrate of artificial intelligence itself, the company that builds the shovels, the railways, and increasingly the entire supply chain for the new gold rush. And now, with its latest moves in robotics and open-source simulation, it is attempting to do for embodied AI what it did for machine learning: make itself indispensable.
The Long March from Phones to Planets
Nvidia's path to dominance was never linear. For years, it was simply the graphics card company, the name gamers knew, the chipmaker that made pixels dance. But the firm has always played a longer game, diversifying into adjacent markets with patient aggression.
By 2011, the company had begun its push into mobile computing, introducing the Tegra line of processors for smartphones and tablets. That year, Nvidia dominated the market for dual-core Android devices in the United States, powering both tablets and what the industry then called "superphones" . The Tegra 3, announced shortly after, was a quad-core marvel, the kind of chip that made early Android flagships feel, briefly, like the future .
Yet mobile never became Nvidia's kingdom. Qualcomm and Apple's custom silicon would see to that. What Tegra did, however, was teach Nvidia how to build system-on-chip architectures for constrained power envelopes — knowledge that would prove invaluable when the firm pivoted to AI inference at the edge. The Jetson family of embedded processors, now purpose-built for robotics and physical AI, descends directly from that Tegra lineage . The chips are optimised for real-time inference on leading open models, small enough to fit inside drones and robotic arms, powerful enough to process sensor data at the millisecond scale required for autonomous machines .
Nvidia also made early, prescient bets on cloud gaming and remote rendering. In the early 2010s, it introduced Nvidia Grid, a hardware-software platform designed to stream games from data centres to thin clients . The technology was ahead of its time — bandwidth was insufficient, latency too high — but the architectural thinking prefigured the company's later cloud AI offerings. Nvidia learned, in those years, how to build not just chips but entire ecosystems: hardware, drivers, APIs, and the software stack that ties them together.
Then came acquisitions, the kind that signal strategic intent. In 2019, Nvidia announced it would acquire Mellanox Technologies for approximately $7 billion in cash, a deal that brought high-speed networking into its portfolio — critical for the interconnects that link thousands of GPUs in modern AI supercomputers . That same period saw Nvidia acquire Cumulus Networks, an open-source networking firm specialising in data-centre optimisation, further tightening its grip on the infrastructure layer .
And then, in 2020, came the boldest move of all: Nvidia announced it would acquire Arm Limited, the British chip-design house whose architectures power virtually every mobile device on Earth, for $40 billion . The deal would have created, in the words of one analyst, "a computing juggernaut" . It would have given Nvidia control over the intellectual property that underpins Apple's M-series chips, Qualcomm's mobile processors, and the vast majority of embedded systems worldwide. Regulators in the United States, Europe, and China ultimately blocked the acquisition on antitrust grounds, but the attempt alone revealed Nvidia's ambition: not merely to dominate AI computing, but to own the entire stack, from mobile to cloud, from edge to exascale.
The Blackwell Doctrine
What Nvidia has built, in the end, is a monopoly not of markets but of capability. Its GPUs — graphics processing units repurposed for parallel computation — have become the de facto standard for training large AI models. Google, Meta, OpenAI, Anthropic: all run their models on Nvidia silicon. The company's CUDA programming environment, now two decades old, has created a moat so wide that competitors struggle even to be considered.
In 2024, Nvidia's then-CEO Jensen Huang stood on stage and declared: "We created a processor for the generative AI era" . He was referring to the Blackwell platform, Nvidia's latest architecture, designed to run real-time generative AI on models with trillions of parameters — at up to 25 times less cost and energy consumption than its predecessor . The Blackwell platform is not merely faster; it is efficient enough to make certain classes of AI application economically viable for the first time. Training a trillion-parameter model previously required the electrical output of a small city; Blackwell makes it possible with the power budget of a large office building.
This is the alchemy Nvidia has mastered: not just raw performance, but performance-per-watt, performance-per-dollar, performance at scale. The Blackwell chips, like their predecessors, will ship not as standalone components but as part of integrated systems — racks of servers, cooling solutions, networking fabric, all optimised to work together. Nvidia, in other words, has become a systems integrator, selling not components but entire AI factories.
The company filed its annual report for the fiscal year ending 26 January 2025, a document that, while dense with the usual corporate prose, reveals revenues and margins that would make an oil cartel blush . Nvidia's gross margins on AI chips hover near 80 per cent; the firm has pricing power that borders on the absolute. When demand exceeds supply by orders of magnitude, you can name your price.
The Open Question of Openness
Yet Nvidia's dominance has always been shadowed by a certain opacity. The company has historically been reluctant to open its hardware documentation, a posture that has frustrated the open-source community for decades. In 2006, Fedora decided not to upgrade to X.org 7.1 because proprietary module vendors, Nvidia among them, had not provided adequate support . The open-source Nouveau driver, built by reverse-engineering Nvidia's hardware, has never achieved feature parity with the company's proprietary drivers.
In recent years, however, Nvidia has shifted — partially, strategically — towards openness, at least in certain domains. It has committed to releasing more documentation over time, offering guidance in specific areas as resources allow . The DIB-R machine-learning framework, which generates 3D models from single 2D images, was released with open licensing, a gesture towards the academic community that relies on Nvidia's tools . More recently, the company released Isaac Sim, a comprehensive open-source platform for training, simulating, and validating AI-based robots in physically accurate virtual environments .
Isaac Sim is more than a simulator; it is a bet on the future of embodied AI. Training a robot in the real world is expensive, slow, and dangerous. Training it in simulation — across thousands of virtual environments, with physics engines that model friction, inertia, sensor noise — allows for rapid iteration at near-zero marginal cost. Nvidia has built Isaac Sim atop its Omniverse platform, a suite of tools originally designed for film and architectural visualisation, now repurposed for the robotics age.
The firm has also released software to accelerate the development of self-driving vehicles, leveraging AI reasoning techniques to improve transparency and decision-making in autonomous systems . The Alpamayo-R1 model, unveiled recently, aims to make self-driving logic more interpretable — a critical requirement as regulators begin to scrutinise the black-box nature of neural-network-based autonomy .
The Age of Generalist Robotics
And then, in early 2025, came GROOT — not the tree-like Marvel character, but Nvidia's open-source foundation model for humanoid robots. The GROOT N1 model, as it is formally known, represents Nvidia's most direct entry into embodied AI . It is a general-purpose model designed to accelerate the development of robots that can navigate unstructured environments, manipulate objects, and learn tasks through demonstration rather than hard-coded programming.
"The age of generalist robotics is here," Nvidia declared when unveiling GROOT, and the claim, for once, may not be hyperbole .
For years, robotics has been stymied by the specificity problem: every robot was a bespoke system, trained for a narrow set of tasks in controlled environments. Factory robots weld car frames with inhuman precision but cannot open a door. Warehouse robots shuffle boxes but cannot fold a shirt. Humanoid robots, the holy grail of the field, have remained largely experimental, too slow and too stupid to be commercially viable.
GROOT aims to change that by providing a pre-trained foundation model — analogous to GPT for language or DALL-E for images — that can be fine-tuned for specific robotic platforms. The model is trained on vast datasets of human motion, object manipulation, and spatial reasoning, then deployed on Nvidia's Jetson hardware for real-time inference . The result is a robot that can generalise: taught to pick up a cup, it can pick up a bottle; taught to navigate one room, it can navigate another.
The decision to release GROOT as open source is a calculated one. Nvidia does not sell robots; it sells the chips and the software that make robots possible. By open-sourcing the model, the company ensures that every robotics startup, every research lab, every ambitious hardware manufacturer will build on Nvidia's stack. The more robots that run GROOT, the more Jetson modules Nvidia sells. The more researchers train models in Isaac Sim, the more GPUs Nvidia leases in the cloud. It is the razor-and-blades model, applied to the infrastructure of intelligence itself.
The Memory That Wasn't
Not every Nvidia initiative has been seamless. In 2015, the company faced a public-relations debacle over the GeForce GTX 970 graphics card, a popular gaming GPU marketed as having 4 gigabytes of video memory . Enthusiasts soon discovered that only 3.5 gigabytes of that memory was readily accessible; the final 512 megabytes sat on a slower partition of the memory bus, causing performance to plummet when applications tried to use the full 4 gigabytes .
The controversy was not merely technical but ethical: had Nvidia intentionally misled consumers, or was it a communications failure? The company's response was muddled. It clarified that the GTX 970 could, in theory, utilise all 4 gigabytes — but only under "highly stressful conditions," and with a performance penalty . Nvidia also stated, definitively, that there would be no driver update to improve memory allocation or performance specifically for the GTX 970 . The firm offered to help retailers process returns, a tacit acknowledgement that the product had not met consumer expectations .
The incident was a rare stumble for a company that has otherwise managed its reputation with precision. It also underscored a tension that persists in Nvidia's relationship with its customers: the firm's products are so dominant that users have little choice but to accept their quirks, their closed documentation, their proprietary lock-in. The GTX 970 flap faded; sales barely dipped. Nvidia's market position, even then, was already too strong to be seriously threatened by a single product misstep.
The Infrastructure of Thought
What Nvidia has built, ultimately, is not a product line but a platform — the infrastructure upon which the AI revolution is being constructed. Every breakthrough in generative AI, every new robotics capability, every autonomous vehicle that inches closer to viability: all of it runs, at some point in the pipeline, on Nvidia hardware.
The company's market capitalisation, now $5.01 trillion, is not a bubble in the traditional sense . It is a reflection of something more fundamental: the recognition that AI is not a single technology but a new substrate for computing, and that Nvidia controls the means of production. Just as Standard Oil once dominated energy, just as AT&T once dominated communications, Nvidia now dominates the infrastructure of machine intelligence.
But monopolies, however efficient, are fragile. They invite regulation, competition, and the slow entropy of technological disruption. Google is designing its own TPUs; Amazon its own Trainium chips; even Apple is rumoured to be working on custom AI silicon. Open-source alternatives to CUDA are proliferating, albeit slowly. The fortress Nvidia has built is formidable, but it is not eternal.
The GROOT initiative, and the broader push into robotics, represents Nvidia's attempt to secure the next frontier before competitors arrive. If AI was the first wave — software intelligence, running in data centres — then robotics is the second: physical intelligence, running in the world. Nvidia learned from its mobile missteps that it cannot afford to cede new markets to rivals. This time, it is moving early, investing heavily, and giving away just enough — through open-source models and accessible simulation tools — to ensure that when the robotics revolution arrives, it will be built on Nvidia's terms.
The Diner, Revisited
Thirty-two years after that meeting in San Jose, Nvidia is no longer a startup with a hunch. It is the central bank of the AI economy, the institution that sets the terms of trade, the entity without which the entire edifice wobbles. Its chips are allocated like a scarce commodity; CEOs of Fortune 500 companies personally call Jensen Huang to beg for early access to the latest hardware. Governments treat Nvidia as a strategic asset, regulating its exports as if they were weapons-grade plutonium.
And yet, for all its dominance, Nvidia remains, in some fundamental sense, a bet: the bet that AI will continue to scale, that demand for computation will continue to outstrip supply, that the future will require ever-more silicon dedicated to ever-larger models. The robotics push is a hedge against that bet's failure — a way to diversify into embodied AI, into edge computing, into markets where the scaling laws of transformer models do not apply.
Whether GROOT and Isaac Sim will prove as transformative as CUDA and the GeForce line remains to be seen. Robotics is harder than software; physics is less forgiving than probability. But Nvidia has spent three decades learning how to build platforms, not just products — how to create ecosystems that lock in developers, researchers, and customers alike.
The company that three engineers founded over breakfast is now worth more than the market capitalisation of entire nations. It has become, in a very literal sense, too big to fail — at least for now, at least while the AI boom lasts, at least until the next paradigm shift arrives and some other diner-table conversation begins to sketch the future.
For the moment, though, the future runs on Nvidia's chips, simulates in Nvidia's software, and increasingly, will walk on Nvidia's robotic legs. Five trillion dollars is not the end of the story. It is merely the price the market has placed on the infrastructure of thought itself.