American semiconductor company and maker of the world's leading artificial-intelligence chips. Nvidia accounts for about 90% of commercially available AI chips; AMD is its only serious American rival. In July 2025 it became the first company ever to reach a market value of $4trn; by October it was the first to hit $5trn. Its high-end processors are produced by TSMC in Taiwan. Its software platform, CUDA, is by far the most widely used for programming AI chips; nearly all AI developers learn it, and it works only with Nvidia's hardware. None of Nvidia's 20 biggest suppliers is Chinese.
Before America's first export controls in October 2022, China accounted for about 22% of Nvidia's revenue. After those controls barred sales of its most powerful chips, Nvidia created the H800, a made-for-China model engineered to stay just under the limits. When America tightened the rules again in late 2023, banning any chip with too much computing power regardless of memory bandwidth, Nvidia responded with the H20, a chip that was inadvertently hobbled for training new AI models but well suited to inference—the process of running them. The Trump administration then, in effect, banned the H20 as well. Nvidia said the new rules would wipe $5.5bn from the value of its inventory.
In July 2025 Nvidia was again permitted to sell H20 AI processors in China, after a ban that may have cost the company around $10bn in forgone Chinese revenue. If Nvidia lost China entirely it would benefit Huawei, which already offers AI chips more powerful than the H20. In August 2025 Trump imposed a 15% surcharge on H20 proceeds—in effect an export levy that probably violates Article 1 of the constitution. China has pressed local firms to shun the H20.
China's share of Nvidia's revenue has fallen to 13%. Sales to Singapore have more than doubled and now make up nearly 18% of the total, making it Nvidia's second-largest market after America. The company says many clients invoice through Singapore but ship to permitted destinations; fewer than 2% of chips sold there are delivered locally. Banned Nvidia chips sell at a 30-50% mark-up through grey-market intermediaries.
In March 2026 federal prosecutors charged the co-founder of Supermicro with smuggling $2.5bn-worth of AI servers containing advanced Nvidia chips to Chinese customers—the largest such case to date. Nvidia insists that illicit diversion is a "losing proposition", since smuggled servers receive no service or support.
The Bureau of Industry and Security (BIS), the agency tasked with enforcing American export controls, has just one export-control officer responsible for all of South-East Asia and Australasia.
China's State Administration for Market Regulation (SAMR) is probing Nvidia alongside Alphabet. SAMR said in September 2025 that Nvidia had broken China's anti-monopoly laws in its 2020 acquisition of Mellanox, an Israeli supplier of computer networks—even though SAMR approved the deal at the time and did not specify which conditions Nvidia had breached. The antitrust agencies of the Trump administration are also investigating the company.
In December 2025 Trump authorised the sale of Nvidia's H200 chips to "approved customers" in China, in exchange for a 25% cut for Uncle Sam. The H200, though lagging behind Blackwell and the forthcoming Rubin, is more than six times as powerful as the H20. Chinese regulators have urged local companies to avoid Nvidia altogether, citing security concerns. David Sacks, Trump's AI adviser, argued against cutting China off entirely, a view that prevailed over that of China hawks in Trump's circle.
In late 2025 Alphabet, one of Nvidia's biggest customers, emerged as its fiercest competitor. Google launched Gemini 3, trained entirely on its own tensor-processing units (TPUs), and began selling them to others. Anthropic announced plans to use as many as 1m TPUs in a deal reportedly worth tens of billions of dollars. Bernstein estimates that Nvidia's GPUs account for over two-thirds of the cost of a typical AI server rack; Google's TPUs cost between a half and a tenth as much as an equivalent Nvidia chip. Mr Huang described Google as "a very special case" given that it began developing chips long before the current AI wave, and dismissed other efforts as "super adorable and simple". Other tech giants including Amazon, Meta and Microsoft have been developing custom processors, and OpenAI announced a collaboration with Broadcom to develop its own silicon. Nvidia's edge lies partly in CUDA, which AI developers have become accustomed to; Google's TPU software has been created with its own products in mind, whereas CUDA caters to a wide range of applications.
Mr Huang has pledged to help Trump reindustrialise America. In October 2025 TSMC produced a wafer for Blackwell at its Arizona fab, and Huang said Blackwell's processors, memory and packaging would all soon be made in America. He also described how Foxconn, using AI-created simulations ("digital twins") built on Nvidia technology, was building fully robotic factories in America.
According to PitchBook, Nvidia invested in 50 firms in 2025 alone. It owns around 7% of CoreWeave, a highly leveraged neocloud, and signed a $6bn agreement to guarantee its revenue. It has put $5bn into Intel, is providing some of the $20bn being raised by xAI, and pledged $15bn (alongside Microsoft) to Anthropic, which will for the first time train its models on Nvidia's chips.
It also announced a proposed investment of up to $100bn in OpenAI, starting in the second half of 2026, to help the maker of ChatGPT buy 4m-5m of Nvidia's AI chips. The investment would increase in $10bn increments for every gigawatt of Nvidia-supported data-centre capacity that OpenAI builds, up to 10GW. An unspoken benefit is that the deal would make OpenAI more reliant on Nvidia's chips, reducing the incentive to build its own. Critics worry about the "circular dynamics" of Nvidia investing in companies it supplies with GPUs.
In 2025 Nvidia said it would unveil a fresh AI chip every year rather than every couple of years. Mr Huang remarked that "when Blackwell starts shipping in volume, you couldn't give Hoppers away." This rapid obsolescence raises questions about how the big cloud companies depreciate their servers; most have stretched assumed server lifetimes from four or five years to six. Jim Chanos, a veteran short-seller, has argued that if the true economic lifespan of Meta's AI chips is two to three years, then "most of its 'profits' are materially overstated." A Barclays analysis estimated that higher depreciation costs would shave 5-10% from the earnings per share of Alphabet, Amazon and Meta. Nvidia's top three customers—widely believed to be Alphabet, Amazon and Microsoft—each spent some $15bn on its chips in 2024.
Nvidia's latest Blackwell graphics-processing units (GPUs) and GB-series AI superchips (which combine two Blackwells with a general-purpose processor) account for almost 60% of revenue. A single Blackwell chip needs 1kW of power, three times more than its Hopper predecessor. Nvidia sells modules packed with 36 GB superchips (72 Blackwells and 36 general-purpose chips) designed to operate at 132kW; a secondary cooling system can add 160kW per rack. If half the chips sold end up in America and are operated at capacity, they would raise American power demand by 25 gigawatts. Analysts estimate a potential American power shortfall of 17-62GW by 2028-30. Nvidia plans to launch next-generation Rubin chips in 2026.
Jensen Huang, Nvidia's boss, has warned that America risks being left behind if American firms do not compete in China as it builds a "rich ecosystem" of AI applications. If they do not, Chinese technology and leadership "will diffuse all around the world", he said.
In late 2023 Mr Huang began advocating that every country should have its own AI system—"sovereign AI." Governments are a lucrative new revenue source: Jefferies estimates sovereign initiatives could generate $200bn in cumulative revenue; Nvidia reckons spending could reach $1trn. Revenue from sovereign AI tripled in the last fiscal year to more than $30bn, about 15% of Nvidia's AI sales.
A paper by Nvidia Research argues that "small, rather than large, language models are the future of agentic AI." Small language models (SLMs), with 40bn parameters or fewer, are sufficiently powerful for agentic tasks and more economical—a 7bn-parameter model can be ten to 30 times cheaper to run than a model up to 25 times bigger. The paper does not represent Nvidia's strategic thinking; the company maintains that business customers want models "of all shapes and sizes."
Mr Huang has more than 50 direct reports—an unusually wide "span of control" in management parlance. He does not hold one-to-one meetings.
In 2022 Nvidia published the first results from FourCastNet, an AI weather program trained on decades of weather data. The company claimed it accurately predicted hurricanes and rainfall a week in advance with just two seconds of computing time—thousands of times less than a numerical weather model needs. See AI weather forecasting.
Nvidia has a generative-AI platform for drug discovery, and is signing deals to offer molecule-design services to pharmaceutical companies. In October 2025 it teamed up with Eli Lilly, the world's most valuable drugmaker, to build the pharma industry's most powerful supercomputer. Mr Huang has declared that "the ChatGPT moment for robotics is here." See industrial robots.
Nvidia is carving out a role as an essential supplier for self-driving vehicles. It sells AI chips both to Waymo (to run simulations and sit inside its cars) and to carmakers developing autonomous systems of their own, including Mercedes-Benz and Stellantis. Tesla has spent billions training its AI systems on 100,000 Nvidia GPUs.
Nvidia plans to become a "foundational company" on which the AI economy rests—selling different types of chips, bundling products into complete "AI factory" systems and embedding its technology more deeply into industries. In the year to January 2026 the firm generated $216bn in revenue, eight times what it made three years earlier. Its GPUs account for over two-thirds of the total processing power available on the world's AI chips.
In December 2025 Nvidia paid $20bn to license technology and hire engineers from Groq, a startup specialising in inference chips, and is pushing into CPUs using designs from Arm. Nvidia has released several families of open-source AI models aimed at specific industries, including Alpamayo for self-driving cars, GR00T for robotics and BioNeMo for biomedical research.
Bernstein says local Chinese suppliers such as Huawei, Cambricon and MetaX could grow from less than a fifth of China's AI-chip market in 2023 to more than nine-tenths by 2027.
Nvidia got its start in gaming but has been eclipsed by its AI business; data-centre sales now account for more than 90% of revenue. Short supply means gaming products are frequently sold well above the recommended price. Power cables in some high-end cards have been melting during use, and Nvidia admitted that some cards had been sold with vital components missing.
The light at the end of the tunnel may be an oncoming dragon.