FT Edit subscription offering curated daily articles

Nvidia's Ascent to the Center of the AI Economy

Nvidia has become one of the most consequential companies in the technology industry, with its market capitalization repeatedly brushing against and surpassing the $3 trillion mark — a threshold previously reached only by Apple and Microsoft. The company's rise is inseparable from the explosion in demand for artificial intelligence infrastructure. Its H100 and more recently H200 graphics processing units have become the essential raw material of the AI era, powering everything from the large language models built by OpenAI and Anthropic to the internal AI projects of Google, Meta, and Amazon. The waiting lists for Nvidia chips stretched to months at peak demand, and the scarcity drove prices for a single H100 server cluster into the tens of millions of dollars.

What makes Nvidia's dominance unusual is how thoroughly it is built on software as much as silicon. The CUDA programming platform, developed over nearly two decades, created an ecosystem so entrenched that rivals — including AMD and Intel — have struggled to dislodge it even when offering competitive hardware. Developers trained on CUDA are reluctant to retool their workflows, and the libraries, frameworks, and optimizations built atop it represent an enormous accumulated advantage that no chip specification sheet can easily replicate.

Jensen Huang and the Architecture of Monopoly-Like Power

Nvidia's chief executive Jensen Huang has become one of the most closely watched figures in global business. His decisions about chip allocation, pricing, and product roadmaps directly affect the competitive positions of some of the world's largest companies. When Nvidia announced its Blackwell architecture — the successor generation to Hopper — hyperscalers and AI startups scrambled to secure supply, with some committing to billions of dollars in forward purchases before full product specifications were even finalized.

Huang has been unusually candid about Nvidia's strategic position, arguing that accelerated computing is not a temporary trend but a permanent restructuring of how computation works. Traditional central processing units, in his framing, are reaching the limits of what they can do efficiently, while the parallel processing architecture of GPUs becomes more, not less, relevant as models grow larger. This argument has proven commercially compelling: Nvidia's data center revenue, which was a secondary business line just five years ago, now dwarfs its original gaming GPU segment.

The company's gross margins — hovering around 70 to 75 percent for its data center products at peak demand cycles — have drawn comparisons to software businesses rather than traditional semiconductor manufacturers. That profitability has in turn funded aggressive investment in next-generation architectures, creating a compounding advantage that competitors find structurally difficult to close.

Challengers, Geopolitics, and the Limits of Dominance

Nvidia's position is not without serious vulnerabilities. The United States government has imposed increasingly strict export controls on advanced chips destined for China, including successive rounds of restrictions that have effectively cut Nvidia off from what was once a substantial and fast-growing market. China represented a meaningful portion of Nvidia's data center revenue before the controls tightened, and each successive regulatory tightening — covering the A100, then the H800, then further derivatives — has forced Nvidia to design progressively neutered versions of its products for the Chinese market, products that Chinese customers have shown diminishing interest in purchasing.

The export controls have had a secondary effect of accelerating China's domestic chip ambitions. Huawei's Ascend series of AI accelerators has gained traction among Chinese technology companies that have no legal alternative, and the Chinese government has backed semiconductor self-sufficiency with substantial state investment. Whether domestic Chinese chips can match Nvidia's performance and — crucially — the software ecosystem surrounding it remains genuinely uncertain, but the pressure is real and growing.

On the competitive front, AMD has made measurable progress with its MI300X accelerator, winning meaningful deployments at Microsoft and elsewhere. Custom silicon projects at the major cloud providers present a longer-term structural challenge: Google's TPUs, Amazon's Trainium and Inferentia chips, and Meta's MTIA project all represent deliberate attempts by the largest Nvidia customers to reduce their dependence on outside suppliers. None of these has yet displaced Nvidia at scale for training frontier models, but they represent a sustained, well-funded effort to do so.

What the Nvidia Story Reveals About the AI Investment Cycle

The broader significance of Nvidia's position extends beyond any single company's fortunes. The tens of billions of dollars flowing through Nvidia's order books represent the physical infrastructure layer of the AI buildout — the concrete and steel equivalent of the internet era's fiber optic cables. Whether that investment ultimately generates proportionate economic returns across the industry is among the most important open questions in technology today.

There is a lively debate among investors and analysts about whether the current pace of AI infrastructure spending is sustainable or whether it represents a capital allocation cycle that will eventually correct. The hyperscalers — Alphabet, Microsoft, Amazon, and Meta — have each committed to capital expenditure plans running into the hundreds of billions of dollars over the next several years, with GPU procurement representing a large share of those budgets. The bull case holds that AI will generate sufficient new revenue streams, in the form of productivity tools, autonomous agents, and new application categories, to justify the investment. The bear case holds that monetization is lagging infrastructure build-out in ways reminiscent of previous technology bubbles.

For Nvidia itself, the near-term picture has remained stronger than skeptics predicted. Demand from sovereign AI initiatives — governments in the Gulf states, Europe, and Asia building nationally controlled AI computing capacity — has added a new customer category that did not exist in meaningful form two years ago. Countries including Saudi Arabia, the UAE, and France have announced large GPU procurement programs, partly as industrial policy and partly as genuine infrastructure investment, extending Nvidia's addressable market in ways that partially offset the China losses. The company's ability to find new demand pools each time an old one closes has been one of the more striking features of its recent history.