Redistributed

General

Nvidia's $3 Trillion Question

Published March 11, 2026

the-wireinfrastructurehardwaremonopolyvalueredistribution

What happened

In June 2024, Nvidia briefly surpassed Microsoft and Apple to become the most valuable company on Earth. Its market capitalization crossed $3 trillion — a figure that would have seemed hallucinatory five years earlier, when the company was primarily known for making graphics cards for gamers. By the end of fiscal year 2024, Nvidia's data center revenue had surged to over $47.5 billion, up from roughly $15 billion the prior year. The company now commands an estimated 80-90% market share in AI training and inference chips.

That number bears repeating. One company supplies the silicon underneath nearly every large language model, every image generator, every AI system your employer is rushing to deploy. The H100, H200, B100, and B200 GPUs are not merely popular products. They are the substrate on which the AI industry is being built.

Jensen Huang, Nvidia's CEO, has become the most consequential figure in AI who does not build AI. He is the infrastructure provider of the AI race — profiting enormously regardless of which AI company wins, because all of them need his chips. The comparison to picks-and-shovels sellers during the Gold Rush understates it: Nvidia does not merely sell tools. It sells the only tools that work.

Why it matters

The redistribution dynamics here are structural, not incidental. Every dollar of value that AI creates passes through Nvidia's toll booth first. When OpenAI trains a new model, it pays Nvidia. When Google scales Gemini, it pays Nvidia. When a hospital deploys diagnostic AI, somewhere in the supply chain, Nvidia gets paid. The company has positioned itself as the essential intermediary — the bottleneck through which the entire AI economy must flow.

This is not simply a story about a successful company. It is a story about infrastructure monopoly — the Hardware Bottleneck made concrete — a pattern with deep historical parallels. Standard Oil controlled the refineries through which all petroleum flowed. AT&T owned the wires through which all calls traveled. The railroad barons of the nineteenth century extracted rents from every bushel of grain and ton of steel that moved across the continent. As Langdon Winner argued in his foundational work "Do Artifacts Have Politics?", technical systems embed power relations — and Nvidia's architecture is no exception. The company has not merely built better chips. It has built a system of dependency.

The real moat is not hardware. It is CUDA — Nvidia's proprietary software platform for parallel computing. CUDA is to AI development what Windows was to personal computing in the 1990s: the ecosystem that developers build on, the toolkit that graduate students learn, the platform around which an entire industry's workflows are organized. Switching costs are immense. A research lab that has spent years writing CUDA-optimized code cannot simply swap in an AMD chip. The software lock-in is more durable than any transistor advantage. This is Pricing Power enforced through code, not contracts.

Kate Crawford's Atlas of AI documents how the material infrastructure of artificial intelligence — from lithium mines to data centers — concentrates power in ways that remain invisible to most users. Nvidia sits at the apex of this concentration. The company's supply chain stretches from TSMC's fabrication plants in Taiwan to server farms consuming megawatts of power across the American Southwest. Every link in this chain represents a dependency, and dependencies create leverage.

The competitive landscape tells a story of how difficult it is to challenge an infrastructure monopoly once entrenched. AMD's MI300X chips have gained traction, particularly for inference workloads, but AMD's software ecosystem remains years behind CUDA's maturity. Intel's Gaudi accelerators have struggled to gain market share. Google's Tensor Processing Units (TPUs) are powerful but available only within Google's cloud — a walled garden, not an open alternative. Amazon's Trainium and Microsoft's Maia chips represent attempts by hyperscalers to reduce their Nvidia dependency, but these are defensive moves by companies that collectively spend billions on Nvidia hardware each quarter.

Then came the DeepSeek shock. In January 2025, Chinese AI lab DeepSeek released models that rivaled frontier Western systems — built, reportedly, with far fewer and less powerful chips than conventional wisdom deemed necessary. Nvidia's stock briefly cratered, shedding hundreds of billions in market value in a single day. The market's fear was existential: what if the AI industry does not need as many GPUs as everyone assumed?

The panic receded. Nvidia's stock recovered. But the DeepSeek episode revealed something important about the redistribution question. The assumption that AI progress requires ever-increasing compute — what some researchers call the "scaling hypothesis" — directly serves Nvidia's interests. If efficiency gains can substitute for raw compute, the chokepoint loosens. DeepSeek demonstrated that algorithmic innovation can, at least partially, route around hardware monopoly. It also highlighted the geopolitical dimension: US export controls restricting chip sales to China have forced Chinese labs to innovate under constraint, producing Nvidia-specific variants for the Chinese market while simultaneously motivating research into doing more with less.

The redistribution arithmetic is stark. Nvidia's fiscal year 2024 net income exceeded $29 billion. That profit was extracted from every organization building, deploying, or using AI. It flowed overwhelmingly to Nvidia's shareholders — a relatively small group of investors who now capture an outsized share of the value that AI generates across the entire economy. The compute-rich get richer. Organizations that cannot afford top-tier Nvidia hardware — universities in the Global South, nonprofits, public institutions, startups without venture backing — are locked out of the frontier. Access to AI capability is, in practice, access to Nvidia GPUs. This is the Digital Divide expressed in silicon.

What to watch

The durability of Nvidia's position depends on three dynamics:

  • The CUDA moat's resilience. Open-source alternatives like AMD's ROCm and the Triton compiler are improving. If the software ecosystem diversifies — if developers can write code that runs efficiently across multiple hardware platforms — Nvidia's lock-in weakens. This is the most consequential long-term threat to the monopoly and the most promising avenue for redistribution.

  • Public compute infrastructure. The US National AI Research Resource (NAIRR) and similar initiatives in the EU aim to provide publicly funded compute access for researchers and institutions priced out of the commercial GPU market. Whether these programs receive adequate funding — and whether they reduce Nvidia dependency or simply become another revenue channel for the company — will shape who can participate in AI development.

  • Efficiency versus scale. If the DeepSeek trajectory continues — if algorithmic advances reduce the raw compute required for capable AI systems — then Nvidia's pricing power erodes. But if the scaling hypothesis holds and frontier models continue to demand exponentially more compute, Nvidia's position only strengthens. The answer to this technical question has profound redistributive consequences.

Nvidia did not set out to become the tollbooth operator of the AI economy. But infrastructure monopolies rarely emerge by grand design. They emerge because someone builds the thing everything else depends on, and then the dependencies compound. The $3 trillion question is not whether Nvidia's chips are good — they are excellent. The question is whether a single company should control the essential infrastructure for a technology that its boosters claim will reshape civilization. History suggests that when one entity controls the chokepoint, the value flows up and the costs flow down. Every railroad baron knew this. The question is whether we remember it — and whether the alternatives being built (open hardware initiatives, public compute infrastructure, algorithmic efficiency gains) can grow fast enough to matter.

Where this fits

Nvidia's dominance is the hardware layer of the Frontier Lab Oligopoly. It shapes the Economics of Training a Frontier Model, determines who can participate in the valuation race, and gives material force to the Compute Is New Oil metaphor. The DeepSeek episode is the most significant challenge to date, suggesting that Small Language Models and efficient architectures may loosen the chokepoint. The AI Chip Export Controls story is inseparable from Nvidia's story — the company designed export-compliant chip variants specifically to maintain access to the Chinese market. And the Environmental Cost of AI is, in large part, the environmental cost of running Nvidia hardware at scale.

Sources

← Back to all articles
On this page