NVIDIA

Nvidia Corporation
Type	Public
Traded as	Nasdaq: NVDA; DJIA component; S&P 500 component
Industry	Semiconductors; Artificial intelligence; Computer hardware
Founded	5 April 1993
Founder	Jensen Huang; Chris Malachowsky; Curtis Priem
Headquarters	Santa Clara, California, United States
Key people	Jensen Huang (President and chief executive officer); Colette Kress (Chief financial officer)
Products	GPUs; Systems on chips; Data centre accelerators; Networking hardware; CUDA and related software
Employees	Approximately 36,000 (2026)
Website	nvidia.com

Nvidia Corporation is an American multinational technology company headquartered in Santa Clara, California, that designs graphics processing units, systems on chips, networking hardware and the software that runs on them. It is the dominant supplier of the accelerators used to train and operate artificial intelligence systems, and in 2025 it became the first company in history to be valued at more than US$4 trillion and then at more than US$5 trillion.

The company was founded in 1993 by Jensen Huang, Chris Malachowsky and Curtis Priem to build graphics chips for personal computers, a market that at the time contained several dozen competitors and that almost all of them left. Nvidia survived, won, and then did something far less common: it converted a position in video game hardware into a position at the centre of scientific computing and then of artificial intelligence, by recognising twenty years in advance that the mathematical structure of computer graphics and the mathematical structure of machine learning are, at the level of the silicon, the same problem.

Nvidia's products include the GeForce line for gaming, the RTX professional line for visualisation and design, the data centre accelerators sold under the Tesla, A100, H100, Blackwell and Rubin names, the Jetson and Drive platforms for robotics and vehicles, and the networking hardware acquired with Mellanox Technologies. Underneath all of them sits CUDA, the programming model Nvidia released in 2006 and has funded continuously since, which is the single most important reason the company's position has proved difficult to attack.

The company is fabless. It designs its chips and has them manufactured, overwhelmingly by Taiwan Semiconductor Manufacturing Company in Taiwan, which places a very large share of the world's artificial intelligence capacity on one island in a contested strait. Nvidia's fortunes are therefore bound up with the geopolitics of semiconductors in a way that few other companies' are, and its sales to China have been repeatedly restricted, permitted and restricted again by governments in Washington and Beijing.

History

Founding, 1993

Nvidia was founded on 5 April 1993 by three engineers who had concluded that the personal computer was about to need a dedicated processor for graphics.

Jensen Huang, born in Tainan, Taiwan, and raised in Thailand, Kentucky and Oregon, had been a microprocessor designer at AMD and then director of CoreWare at LSI Logic. Chris Malachowsky had been an engineer at Sun Microsystems. Curtis Priem had designed graphics chips at IBM and at Sun, where he had built the GX graphics processor. The three had worked together in the Silicon Valley graphics community and had reached the same conclusion separately.

Their initial conversations took place, by the account all three have repeated since, at a Denny's on Berryessa Road in East San Jose, where Huang had worked as a dishwasher and then a waiter as a teenager. The choice of venue was practical rather than sentimental: the coffee was cheap and nobody minded three men occupying a booth for hours with paper napkins.

The thesis was that general-purpose processors were not going to be able to render three-dimensional graphics fast enough, that a specialised parallel processor could, and that the arrival of affordable personal computers capable of multimedia would create a market large enough to support one. The founders were not alone in this view. By 1995 there were roughly seventy companies building graphics chips.

They raised approximately US$20 million from Sequoia Capital, Sutter Hill Ventures and Sierra Ventures, on the strength of the founders' records rather than a product. The name came from invidia, the Latin for envy, and from the placeholder file name NV the founders had used for their work, which they had read as "next version".

The near-death experience, 1995 to 1997

The first product, the NV1, shipped in 1995 and failed.

It was an ambitious part, combining graphics, audio and a game port on one chip, and it rendered three-dimensional objects using quadratic surfaces rather than triangles. The engineering argument for quadratics was that curved surfaces could be described with fewer primitives and therefore less memory bandwidth. The argument was defensible and it was wrong, because Microsoft released Direct3D, which assumed triangles, and the entire software industry followed. The NV1 was incompatible with the standard the market chose.

Nvidia had by then raised money on the NV1, had committed to a follow-on NV2 for Sega's next console, and had approximately thirty days of cash. Huang went to Sega and told its chief executive that the NV2 architecture was wrong, that Sega should not use it, and that Sega should nonetheless pay Nvidia what it had committed. Sega agreed, paid, and went elsewhere for the Dreamcast's graphics. Huang has described this repeatedly as the moment the company survived, and has drawn from it the lesson he restates in almost every interview: that admitting an approach has failed, early and completely, is more valuable than defending it.

The company cut its staff to around forty people and bet everything on a triangle-based, Direct3D-compatible design.

RIVA and the invention of the GPU, 1997 to 2000

The RIVA 128, released in August 1997, was the result. It was fast, it was compatible, it was priced aggressively, and it sold roughly a million units in four months. The RIVA TNT of 1998 and the TNT2 of 1999 followed, and Nvidia went from survivor to leader.

Nvidia listed on Nasdaq on 22 January 1999 at US$12 a share.

In October 1999 the company released the GeForce 256 and described it in its marketing as "the world's first GPU", defining a graphics processing unit as a single-chip processor with integrated transform, lighting, triangle setup and rendering engines capable of processing at least ten million polygons per second. The claim was contested at the time, since several earlier products had performed some of these functions, but the term stuck and is now universal. The genuinely new capability was hardware transform and lighting, which moved the geometry calculations off the main processor.

The same year Nvidia won the contract to supply the graphics processor for Microsoft's first Xbox, a deal worth roughly US$200 million in advance payments that funded a great deal of subsequent development and that ended acrimoniously in arbitration over pricing.

In December 2000 Nvidia acquired the assets of 3dfx, its principal rival, which had made the Voodoo cards that defined the first generation of consumer three-dimensional graphics and which had failed by attempting to manufacture and sell its own boards rather than supplying chips to partners. The purchase was later challenged by 3dfx's bankruptcy trustee, who argued Nvidia had paid less than fair value; Nvidia prevailed, and the Ninth Circuit affirmed in November 2014.

The programmable shader and the road to CUDA, 2001 to 2007

The decisive technical turn came with programmability.

The GeForce 3, released in 2001 with the NV20 chip, introduced programmable vertex shaders, small programs that ran on the graphics processor to determine how geometry was transformed. The GeForce FX and then the GeForce 6 extended programmability to pixel shading and added floating point precision. Each step made the graphics processor less like a fixed pipeline for drawing triangles and more like a parallel computer that happened to be good at drawing triangles.

Academic researchers noticed. By 2003 there were published papers on using graphics hardware for matrix multiplication, fluid simulation, and medical image reconstruction, all performed by disguising the mathematics as a rendering operation: data was loaded as textures, computation was expressed as a shader, and results were read back as pixels. The technique was called general-purpose computation on graphics processing units, or GPGPU, and it was appallingly awkward.

Ian Buck, a doctoral student at Stanford, built a language called Brook that hid the graphics concepts behind ordinary array operations. Nvidia hired him in 2004. Under Buck and with Huang's direct and sustained backing, the company built CUDA, a Compute Unified Device Architecture, which exposed the graphics processor as a parallel processor programmed in a dialect of C. It was released in 2007 alongside the Tesla architecture, whose unified shader design made the hardware genuinely general rather than a graphics pipeline in disguise.

The decision was expensive and, for years, unpopular with investors. CUDA required that every consumer GeForce chip carry transistors devoted to general computation that most gamers would never use, which reduced gaming performance per dollar. It required a large permanent software organisation writing libraries, compilers, debuggers and documentation for a market that did not yet exist. Huang defended the spending through several years in which analysts asked, on earnings calls, when it would stop.

It did not stop, and the reason Nvidia holds the position it does today is that it did not.

The mobile detour and the Tegra years, 2008 to 2014

Nvidia's judgement has not been uniformly good, and its attempt to enter mobile computing is the clearest example.

Tegra, introduced in 2008, combined an Arm processor with Nvidia graphics on a single chip for phones and tablets. The technical execution was reasonable and the strategic position was hopeless. Nvidia had no cellular modem, which Qualcomm did, and integrated modems were what phone manufacturers wanted. It acquired Icera in 2011 to build one and closed the business in 2015 having failed to make it competitive. Tegra won places in the Microsoft Surface RT, the Nexus 7 and various Android tablets, most of which sold poorly.

The business found its footing in two places rather than the one intended. The Nintendo Switch, released in 2017 and built around a Tegra X1, has sold more than 150 million units across its generations and has been a durable and profitable design win. And the automotive platform, which grew out of Tegra, became the Drive family.

The same period brought the chipset defect, one of the more expensive engineering failures in the company's history. In July 2008 Nvidia took a write-down of approximately US$200 million after certain mobile chips and graphics processors were found to have abnormal failure rates caused by a packaging material that cracked under thermal cycling. The parts had been sold to Apple, Dell and HP and were in laptops that failed in the field. A class action followed in September 2008 and was settled in September 2010 with reimbursement for repairs or replacement. The cost, counting warranty and reputation, was considerably more than the write-down.

Litigation with Intel over chipset licensing ran through the same years and ended on 10 January 2011 with a six-year cross-licensing agreement under which Intel paid Nvidia US$1.5 billion. The settlement also confirmed that Nvidia would not build chipsets for Intel processors, closing a business that had been material.

Deep learning, 2012

In September 2012 a neural network called AlexNet, built by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton at the University of Toronto, won the ImageNet Large Scale Visual Recognition Challenge with an error rate roughly ten percentage points below the next entrant. It had been trained on two Nvidia GeForce GTX 580 cards in a bedroom, using CUDA, over about a week.

The result is generally treated as the beginning of the modern era of artificial intelligence, and its significance for Nvidia was immediate and specific: the method that worked was one that consumed enormous quantities of parallel floating point arithmetic, and Nvidia was the only company selling that arithmetic in a form researchers could program.

Huang reorganised the company around the observation with a speed that is the strongest single piece of evidence for his judgement. Nvidia began designing hardware features for neural networks rather than for graphics: Tensor Cores, introduced with the Volta architecture in 2017, perform the mixed-precision matrix multiplications at the heart of training and are of no use whatsoever in rendering a video game. It built cuDNN, a library of neural network primitives, and gave it away. It built the DGX systems, complete computers rather than chips, and sold the first one by hand to OpenAI in 2016, delivered personally by Huang with a message written on the case.

The bet took six years to appear in the financial statements and then appeared all at once.

The Arm attempt, 2020 to 2022

On 13 September 2020 Nvidia agreed to buy Arm from SoftBank Group for approximately US$40 billion in cash and stock, with SoftBank retaining about a tenth of Nvidia. It would have been the largest semiconductor acquisition ever attempted.

The logic was that Arm's instruction set is in essentially every phone and a growing share of servers, and that owning it would give Nvidia a central processor to pair with its accelerators. The objection, raised immediately by Arm's licensees, was that Arm's business model depends on neutrality: Qualcomm, Apple, Samsung, Broadcom and others license Arm designs and compete with Nvidia, and would have been licensing from a competitor.

The United Kingdom's Competition and Markets Authority raised significant competition concerns in August 2021. The European Commission opened an investigation in October 2021, stating that the acquisition could restrict competitors' access to Arm products and give Nvidia information about competitors' plans. The United States Federal Trade Commission sued to block it in December 2021. In February 2022 SoftBank and Nvidia abandoned the transaction, Nvidia forfeiting a US$1.25 billion break fee that SoftBank kept and applied to Arm's own listing in 2023.

In retrospect the failure cost Nvidia very little. Within two years the value of the accelerator business had made the entire Arm question a footnote, and Nvidia built its own Arm-based central processor, Grace, under an architecture licence it already held.

The AI boom, 2023 to 2026

Nvidia's revenue for the financial year ending January 2023 was approximately US$27 billion, of which the data centre segment was about US$15 billion. Two years later the annual figure was above US$130 billion, with the data centre segment approaching US$115 billion, and gross margin in the mid-seventies as a percentage of revenue, a figure more typical of software than of hardware.

The cause was the arrival of large language models in general use. ChatGPT, released in November 2022, made the capability legible to people outside research, and within months every large technology company, every well-funded startup and a considerable number of governments were attempting to buy accelerators. Training a frontier model requires tens of thousands of them running for months; operating one for millions of users requires many more. The supply was constrained by TSMC's advanced packaging capacity and by the availability of high bandwidth memory, not by Nvidia's willingness to sell.

Nvidia passed a US$1 trillion valuation in May 2023, becoming the seventh American company to do so. It closed above US$2 trillion on 1 March 2024, having taken 180 days to double where Apple and Microsoft had each taken more than 500. On 18 June 2024 it became the most valuable company in the world, passing both. It was added to the Dow Jones Industrial Average in November 2024. In 2025 it passed US$4 trillion and then US$5 trillion, the first company to reach either figure.

The period has not been uninterrupted. On 27 January 2025 Nvidia lost approximately US$600 billion of market value in a single session, the largest one-day loss by any American company, after DeepSeek, a Chinese laboratory, released a model whose reported training cost was a small fraction of what Western laboratories were spending. The market read this as evidence that demand for accelerators might be far lower than assumed. The reading was at least premature: DeepSeek's efficiency gains lowered the cost of a unit of capability, and demand for capability proved elastic enough that total spending continued to rise. The episode nonetheless established the central risk in the investment case, which is that the value of the company depends on an assumption about how much computation intelligence requires, and that nobody knows the answer.

By 2026 the concentration had become a subject of comment in its own right. Charles Schwab reported in August 2026 that Nvidia and Micron together accounted for roughly a third of the aggregate earnings growth of the entire S&P 500 for the year, and the ten largest contributors for about two-thirds. An index intended to represent the breadth of American business had come to depend on the capital expenditure plans of a handful of buyers of one company's products.

Products and technology

Architectures

Nvidia names each generation of its graphics architecture after a scientist, and the sequence is a reasonable summary of the company's technical history.

Tesla (2006) introduced the unified shader, making the same processing units available for vertex, pixel and general computation, and shipped alongside CUDA. Fermi (2010) added error-correcting memory, a proper cache hierarchy and much stronger double-precision arithmetic, which is what scientific computing requires; it was the architecture that made Nvidia credible in supercomputing, and it ran extremely hot. Kepler (2012) prioritised performance per watt after the complaints about Fermi. Maxwell (2014) continued that emphasis and was unusually efficient.

Pascal (2016) brought the move to 16 nanometre manufacturing and a large generational jump; the GTX 1080 and the data centre P100 were both significant. Volta (2017) introduced Tensor Cores and was aimed almost entirely at data centres. Turing (2018) added hardware ray tracing cores to consumer parts and introduced DLSS, an upscaling technique that renders at low resolution and reconstructs detail with a neural network, which was the first consumer application of Nvidia's own AI research to its own graphics products.

Ampere (2020) produced the A100, the accelerator on which a great deal of the first wave of large language models was trained, and the GeForce RTX 3000 series, which was almost impossible to buy for two years because of cryptocurrency mining and pandemic supply constraints. Hopper (2022) produced the H100, the part most associated with the AI boom, with a Transformer Engine that handles the specific arithmetic of the attention mechanism and support for eight-bit floating point. Ada Lovelace (2022) served consumers and professional visualisation.

Blackwell (2024) is built from two dies joined by a ten terabyte per second interconnect and presented to software as one processor, an approach adopted because reticle limits prevent single dies growing further. The GB200 NVL72 rack combines 72 Blackwell processors and 36 Grace central processors into one liquid-cooled unit treated as a single very large accelerator. Morgan Stanley reported in November 2024 that the whole of 2025 production was already sold. Blackwell Ultra followed in 2025 and Rubin, named for the astronomer Vera Rubin, was announced for 2026 with a companion central processor named Vera.

The cadence has shifted. Nvidia moved from a roughly two-year architecture cycle to an annual one in 2024, a decision that places enormous strain on its own engineering and on TSMC, and that is intended to make it unprofitable for a competitor to catch a moving target.

CUDA and the software position

CUDA is the reason Nvidia's position has survived the arrival of well-funded competition, and it is widely misunderstood as a programming language when it is better described as an accumulated stock of working software.

The language itself is not especially difficult to replicate. What is difficult to replicate is nineteen years of libraries, of compiler optimisations for particular loop shapes, of numerical routines whose edge cases have been found and fixed, of profiling tools, of framework integrations, of documented behaviour, of Stack Overflow answers, of graduate students who learned it and took it to industry, and of code in production that works and that nobody will rewrite without a very strong reason.

The layered stack includes cuBLAS for linear algebra, cuDNN for neural networks, cuFFT for transforms, NCCL for communication between many processors, TensorRT for optimised inference, RAPIDS for data science, and a growing set of domain libraries for genomics, computational lithography, weather and quantum simulation. PyTorch and TensorFlow, through which most machine learning is actually written, are tuned first and best against CUDA.

Competitors attack this in two ways. AMD offers ROCm, a broadly comparable stack that is improving and that has been adopted at some large sites; its difficulty has historically been reliability across the long tail of operations rather than peak performance on the common ones. The second approach is to bypass CUDA entirely by having frameworks compile to multiple back ends, which is the strategy behind OpenAI's Triton, Google's JAX and XLA, and PyTorch's compiler. This is the more serious threat over time, because it attacks the dependency rather than the product.

Nvidia's response has been to keep moving up: to sell not chips but racks, not racks but networked clusters, and to make the value increasingly reside in the interconnect and the systems software where substitution is hardest.

Networking

Nvidia's acquisition of Mellanox Technologies for US$6.9 billion, announced on 11 March 2019 and completed in April 2020, is the most consequential purchase the company has made and was widely underestimated when it happened.

Mellanox was an Israeli company that made InfiniBand and high-speed Ethernet networking for data centres. The purchase looked at the time like a sensible adjacency. It turned out to be the thing that allowed Nvidia to sell clusters rather than chips.

Training a large model is not one computation but tens of thousands of processors performing a computation together, and every processor must exchange its results with the others at every step. The speed of the whole is set by the slowest part of that exchange. Owning both the accelerators and the network let Nvidia design them against each other: NVLink connects processors within a rack at very high bandwidth, NVSwitch aggregates those links, and Quantum InfiniBand or Spectrum Ethernet connects racks across a building. BlueField data processing units move networking, storage and security work off the central processors. The combination is why an Nvidia cluster is faster than the sum of its accelerators, and why a competitor who matches the accelerator still does not match the system.

Nvidia outbid Intel and others for Mellanox. The Israeli operation has since become one of the company's largest engineering centres, and in 2023 Nvidia built Israel-1, described as the country's most powerful AI supercomputer, around it. Nvidia also bought Cumulus Networks in May 2020 for its network operating system, folding it into the same unit.

Systems and supercomputing

DGX systems are complete computers containing eight accelerators, the networking to connect them and a full software stack, sold as an appliance. The first DGX-1 was delivered by Huang in person to OpenAI in August 2016. DGX SuperPOD is a reference design for a cluster of them, and DGX Cloud rents the same systems through third-party data centres, which puts Nvidia into partial competition with its largest customers.

Nvidia hardware occupies a large share of the TOP500 list of the world's fastest supercomputers and a larger share of the Green500 list ranked by efficiency. Systems including Switzerland's Alps, Italy's Leonardo, Japan's ABCI and the United States' Perlmutter and Selene are built on it. The company operates its own machines, Selene and later Eos, both for internal research and as engineering proofs of the reference designs it sells.

The Grace central processor, an Arm-based design released in 2023, is paired with accelerators in the Grace Hopper and Grace Blackwell superchips, connected by NVLink rather than by the slower PCI Express bus used to attach an accelerator to a conventional processor. This closes the loop Nvidia had tried to close by buying Arm.

Automotive and robotics

Nvidia Drive provides computing for driver assistance and autonomous driving, combining the Orin and later Thor processors with perception, planning and simulation software. Customers have included Mercedes-Benz, Volvo, Jaguar Land Rover, BYD, Nio, Xpeng, Li Auto, Lucid and Toyota, the last announced in May 2017. The business is significant strategically and modest financially, in the low single-digit billions annually, because automotive design cycles are long and volumes take years to arrive.

Isaac is the equivalent platform for robotics, and Jetson is the family of small modules used in robots, drones, cameras and industrial equipment. Nvidia has invested heavily in humanoid robotics through the Project GR00T foundation model effort, on the argument that a robot requires three computers: one to train the model, one to simulate the world it will be trained in, and one inside the machine itself. Nvidia sells all three.

Omniverse, introduced in 2019 and steadily extended, is a platform for building physically accurate simulations of factories, warehouses and cities, used both for industrial design and to generate synthetic training data for robots. It is the basis of what Nvidia calls digital twins, and its adoption by manufacturers including BMW, Foxconn and Siemens has been one of the company's more credible attempts to sell into industry rather than technology.

Corporate affairs

Leadership and culture

Jensen Huang has been chief executive since the company was founded in 1993, which makes him one of the longest-serving founder chief executives of any large public company, and the culture is unusually a direct expression of one person's preferences.

Huang has roughly sixty direct reports, against a typical figure of eight or ten, on the explicit argument that layers of management exist to filter information and that he would rather have the information. He does not hold one-to-one meetings in the conventional sense and does not give private feedback, preferring to correct people in front of their colleagues on the grounds that the lesson is then available to everyone. He describes the company as having "no org chart" in any meaningful sense and as being organised around projects. He answers a very large volume of email personally and requires from his senior people a weekly written note listing the five most important things they are working on, a practice he has said replaces status meetings.

The style is demanding and has been described as such by people who have worked under it. Huang has spoken publicly about tolerance for suffering as a prerequisite for success and has said he wishes "pain and suffering" on employees, a remark that reads badly out of context and that in context is an argument about resilience rather than a description of practice. Retention is unusually high for the industry, helped considerably by the fact that equity granted at almost any point in the past fifteen years has appreciated enormously.

He wears a black leather jacket in public appearances to the point that it is a recognised signature, and his keynotes at the company's GTC conference have become events of a kind previously associated with consumer product launches rather than with semiconductor companies.

Colette Kress has been chief financial officer since 2013, joining from Cisco and previously Microsoft. Other long-serving executives include Jay Puri in worldwide field operations, Debora Shoquist in operations, and Ian Buck, the author of the research that became CUDA, who runs the accelerated computing business.

The board has included Dawn Hudson, Harvey Jones, Michael McCaffery, Stephen Neal, Mark Stevens, Aarti Shah, Tench Coxe and Persis Drell. Huang serves as both chairman and chief executive, a combination governance specialists generally criticise and that shareholders have consistently declined to change.

Finances

Nvidia's financial year ends in late January, so the year labelled 2026 covers most of calendar 2025.

Revenue was US$4.7 billion in financial 2017, US$10.9 billion in 2020, US$26.9 billion in 2023, US$60.9 billion in 2024 and approximately US$130.5 billion in 2025, with the following year higher again. Net income followed a steeper curve, reaching approximately US$72.9 billion in financial 2025. Gross margin has run between 73 and 78 per cent through the boom.

The segment split is the significant number. Data centre, which was a minority of revenue as recently as 2022, is now approximately seven-eighths of it. Gaming, historically the entire company, is now a modest business in absolute terms and a small one in relative terms, though still larger than most semiconductor companies in total. Professional visualisation and automotive are each a low single-digit percentage.

Customer concentration is extreme and is disclosed as such: a small number of hyperscale cloud providers and large AI laboratories account for a very large share of sales, and in some quarters two customers have each exceeded a tenth of revenue. Nvidia has also taken equity positions in customers and partners including CoreWeave, Nebius, Applied Digital and, announced in 2025, a very large investment in OpenAI, which has drawn scrutiny over whether the company is helping to finance the demand it reports.

The company pays a nominal dividend and has returned very large sums through buybacks. It executed a four-for-one stock split in 2021 and a ten-for-one split in June 2024.

Manufacturing

Nvidia does not own a fabrication plant. It designs chips and contracts their manufacture, and the overwhelming majority of that manufacture is performed by Taiwan Semiconductor Manufacturing Company at its facilities in Taiwan. Samsung has built some generations, notably the Ampere consumer parts.

The constraint on supply through the boom has generally not been the transistors but the packaging. Modern accelerators combine a logic die with stacks of high bandwidth memory on a silicon interposer, a technique TSMC calls chip-on-wafer-on-substrate, and the capacity for that step has been the bottleneck. High bandwidth memory itself comes from SK Hynix, Samsung and Micron, and Nvidia has at times pre-purchased years of output.

The concentration is a strategic exposure without a good answer. A very large share of the world's capacity to build advanced artificial intelligence sits on an island that the government in Beijing claims. TSMC is building plants in Arizona, Japan and Germany, and Nvidia said in 2025 that Blackwell wafers had begun production in Arizona, but leading-edge capacity outside Taiwan remains a small fraction of the total and advanced packaging is only beginning to move. Nvidia has also announced manufacturing and testing work in Texas with Foxconn and Wistron, as part of a commitment to build AI infrastructure in the United States.

China, export controls and geopolitics

Nvidia's relationship with China is the clearest illustration of a semiconductor company becoming an instrument of foreign policy, and it has cost the company a great deal of money.

In October 2022 the United States restricted exports of the A100 and H100 to China on national security grounds, reasoning that the same chips that train commercial models train military ones. Nvidia responded by designing the A800 and H800, cut down to fall below the thresholds. In October 2023 the rules were tightened to capture those parts as well. Nvidia designed the H20, which used 96 gigabytes of HBM3 memory and four terabytes per second of bandwidth, higher than the H100, but delivered roughly 296 teraflops against the H100's 1,979. Despite the much lower raw figure it performed more than a fifth faster than an H100 on some large language model inference work, because inference is often limited by memory bandwidth rather than arithmetic.

In April 2025 the H20 was itself brought under licence, and Nvidia wrote off approximately US$4.5 billion of inventory and commitments. In August 2025 the position reversed again, with an arrangement under which sales to China could resume in exchange for a share of the revenue paid to the United States government, an unusual instrument with contested legal basis. Within the same month the Chinese government advised domestic companies against buying the H20 on security grounds, and Nvidia instructed TSMC, Samsung and Amkor to stop production, having ordered 300,000 units from TSMC in July. A Blackwell-derived B30A was under development for the market.

In 2026 the United States said it would permit export of the newer H200 under conditions. China responded with restrictions on importing certain foreign chips, using the situation to push domestic alternatives from Huawei, Cambricon and others. Nvidia has estimated the potential cost of Chinese refusal at around US$30 billion. Huang has argued publicly and repeatedly that export controls have failed on their own terms, in that they have accelerated Chinese domestic development while removing American influence over the ecosystem, and that the correct objective is for the world to build on American technology. Critics respond that this is the argument a seller would make.

The controls have also produced smuggling. Chips have been routed through Singapore, Malaysia and the Middle East, and in August 2026 employees of Nvidia and Supermicro were indicted over the smuggling of artificial intelligence servers to China, including a senior Nvidia manager alleged to have forged documents to conceal illegal exports.

Competition

Nvidia's competition falls into three categories, and only one of them is a conventional competitor.

AMD is the direct rival, with the Instinct MI300 and MI350 accelerators and the ROCm software stack. Its hardware is competitive on paper, its memory capacity has at times exceeded Nvidia's, and it has won places at Microsoft, Meta and Oracle. Its problem is the software and the networking, which is to say that it competes with the chip rather than with the system.

The hyperscalers building their own are the more serious structural threat, because they are simultaneously Nvidia's largest customers. Google has built Tensor Processing Units since 2016 and is on its seventh generation; Gemini is trained on them, and Google has begun offering them to others. Amazon has Trainium and Inferentia, and has built very large Trainium clusters for Anthropic. Microsoft has Maia. Meta has MTIA. OpenAI is designing its own with Broadcom. None of these need to beat Nvidia on any general measure; each needs only to be cheaper than Nvidia for that company's own workloads, which is a far lower bar, and each is designed with knowledge of exactly what those workloads are.

Broadcom and Marvell sit behind that trend, designing custom accelerators for the hyperscalers and taking a share of the value that would otherwise be Nvidia's. Broadcom's AI revenue has grown to the point of being discussed as a genuine alternative to Nvidia's franchise.

Beyond these are the startups, including Cerebras with a wafer-scale processor, Groq with deterministic inference hardware, SambaNova and Tenstorrent, all of which have credible technical arguments and none of which has yet achieved scale. Intel has made repeated attempts with Nervana, Habana Gaudi and Falcon Shores, none of which has established a durable position, and remains the most obvious candidate for a revival given its manufacturing.

In consumer graphics the competition is AMD's Radeon and, since 2022, Intel's Arc, in a market where Nvidia holds roughly four-fifths of discrete desktop units.

Litigation and regulation

Nvidia and AMD received subpoenas from the United States Department of Justice in December 2006 over possible antitrust violations in graphics cards.

The 2008 chipset defect produced a class action, settled in September 2010 with repairs or replacement for owners of affected Apple, Dell and HP laptops. The 3dfx bankruptcy trustee's challenge to the 2000 asset purchase failed, affirmed by the Ninth Circuit in November 2014.

In May 2022 the Securities and Exchange Commission charged Nvidia with inadequate disclosure about the contribution of cryptocurrency mining to its gaming revenue growth in 2018. Nvidia paid US$5.5 million without admitting or denying the findings. A related shareholder suit, E. Ohman J:or Fonder AB v. Nvidia, reached the Supreme Court, which in December 2024 dismissed the appeal as improvidently granted, leaving the case to proceed.

The French Competition Authority searched Nvidia's Paris offices on 26 September 2023 as part of an investigation into suspected anti-competitive practices in graphics cards, and later signalled charges. In June 2024 the Department of Justice opened an antitrust investigation into Nvidia while the Federal Trade Commission examined Microsoft and OpenAI, the inquiries directed at conduct rather than mergers, including allegations that Nvidia penalises customers who buy from competitors and bundles networking with accelerators. The European Commission and Chinese regulators have opened their own reviews, China's focused on conditions attached to the Mellanox clearance.

In January 2026 a court filing in copyright litigation indicated that Nvidia developers had contacted the shadow library Anna's Archive to evaluate pirated material for model training and that management had approved it despite concerns raised. This followed reporting in 2024, based on internal communications, that Nvidia had scraped large quantities of video from YouTube without permission and by circumventing protections. Nvidia denied the allegations and moved to dismiss.

Controversies

The GTX 970 memory dispute

Shortly after the GeForce GTX 970 was released in 2014, users observed that the card, advertised with four gigabytes of memory, rarely used more than three and a half.

Investigation established that the card's specifications differed from those announced. It had 1.75 megabytes of level two cache rather than the two megabytes of the GTX 980, and 56 render output units rather than 64. More significantly, its memory was divided into a fast 3.5 gigabyte segment on a 224-bit bus and a slow 0.5 gigabyte segment on a 32-bit bus, the latter roughly seven times slower, because the unit managing two memory controllers shared its read and write channels between them. Using the last half gigabyte therefore caused a sharp performance drop.

Nvidia acknowledged that the published specifications had been altered without notice, promised a driver modification to mitigate the problem, then said that promise had been a miscommunication and that no such update would come. It offered to assist customers seeking refunds. Huang apologised on the company blog on 26 February 2015. A class action alleging false advertising was filed in February 2015 in the Northern District of California against Nvidia and Gigabyte Technology, and was settled in July 2016 with a US$30 refund per card.

GeForce Partner Program

Announced on 1 March 2018, the GeForce Partner Program offered manufacturers public relations support, game bundles and marketing funds. Reporting indicated that participation required a partner's gaming brand to be aligned exclusively with Nvidia, so that a manufacturer could not sell an AMD card under an established gaming label. The arrangement was widely criticised as anti-competitive and the programme was cancelled on 4 May 2018.

Hardware Unboxed

On 10 December 2020 Nvidia told Steven Walton of the review channel Hardware Unboxed that it would no longer supply Founders Edition review samples, because his reviews focused on rasterisation performance rather than ray tracing. An email from senior public relations manager Bryan Del Rizzo said that the channel's recommendations "have continued to focus singularly on rasterization performance" and that "it is very clear from your community commentary that you do not see things the same way that we, gamers, and the rest of the industry do", and that the decision would be revisited "should your editorial direction change".

The reaction across technology media was strongly negative. Linus Sebastian devoted an episode of the WAN Show to it; Gamers Nexus called it a self-inflicted wound that would make any reviewer covering ray tracing suspect. TechSpot, Hardware Unboxed's partner site, said the incident raised serious questions about journalistic independence. Nvidia reversed two days later, and Del Rizzo sent an apology saying that "to withhold samples because I didn't agree with your commentary is simply inexcusable and crossed the line".

Supply, pricing and mining

Nvidia's consumer business has been repeatedly disrupted by cryptocurrency mining, which in 2017 and again in 2021 absorbed most available supply and pushed street prices far above list. The company's attempts to manage this, including the Lite Hash Rate limiter on RTX 3000 cards and separate mining products, were partially circumvented. When mining demand collapsed the resulting inventory correction was severe, and the failure to explain the exposure to investors is what produced the SEC settlement.

Pricing of the top consumer cards has risen substantially across the RTX generations and the company has been criticised for allocating capacity away from gaming toward data centre products, for naming conventions that attach a high tier number to a substantially cut-down chip, and for shipping cards whose memory capacity is low relative to price.

Open source and developer relations

Nvidia's history with open source software has been contentious, particularly in the Linux community, where the company's proprietary driver was long the main obstacle to a fully open graphics stack. Linus Torvalds made an obscene gesture at a public event in 2012 while calling Nvidia "the single worst company we've ever dealt with", a moment that circulated widely.

The position has changed considerably. In May 2022 Nvidia published open source GPU kernel modules under dual GPL and MIT licensing, and by 2024 these had become the default for recent architectures. The company maintains a large number of open projects, including the Linux kernel modules, the container toolkit, DALI for data loading, NeMo for language model development, Megatron-LM, Cutlass for matrix kernels, Warp, and the Open Source Physics and Isaac simulation components, and it has contributed to Kubernetes, PyTorch and the Nouveau driver through published firmware and documentation.

The commercial core, including CUDA itself, the compiler and the principal libraries, remains proprietary, which is the point of it.

The company's developer relations are conducted largely through GTC, the GPU Technology Conference, which began in 2009 as a modest technical event and has become one of the largest gatherings in computing, with Huang's keynote treated as a statement of direction for the industry. The Inception programme supports several thousand startups with hardware credits, technical help and introductions to investors, a scheme that both seeds demand and gives Nvidia early visibility of what is being built.

Gaming and consumer graphics

Gaming was Nvidia's entire business for its first fifteen years and remains the foundation of its brand, its driver engineering and a very large developer community, even though it is now a small share of revenue.

GeForce

The GeForce line has run continuously since 1999 and is the most recognised brand in computer hardware. Its generations have followed the architectures: the GeForce 256 and 2 on the original design, the 3 and 4 with programmable vertex shading, the FX, which was a commercial disappointment against ATI's Radeon 9700 and remains the company's clearest product-level defeat, the 6, 7, 8, 9, then the numbered GTX series from 200 through 900, the GTX 10 series on Pascal, and the RTX series from 2018.

The GTX 1080 of 2016 and the GTX 1060 that followed it are often cited as the best-regarded generation the company has produced, because the jump from 28 to 16 nanometre manufacturing delivered an unusually large improvement at prices that did not rise to match. The RTX 3000 series of 2020 was technically strong and almost impossible to obtain for two years, first because of pandemic supply constraints and then because Ethereum mining absorbed the supply. The RTX 4000 and 5000 series have been criticised for price and for memory capacity even as they have extended the technical lead.

Ray tracing and DLSS

The RTX branding introduced in 2018 marked two changes that have shaped the last decade of real-time graphics.

Ray tracing simulates light by following individual rays from the camera into the scene, bouncing them off surfaces and accumulating what they encounter. It is how film visual effects have been produced for decades and it produces reflections, shadows and indirect illumination that are correct rather than approximated. It had been considered impossibly expensive for interactive frame rates. Turing added dedicated hardware for the ray-triangle intersection tests that dominate the cost, and Microsoft added DirectX Raytracing to Windows at the same time. Early implementations were slow and the feature was widely dismissed as marketing; a generation later, path traced modes in games including Cyberpunk 2077, Alan Wake 2 and the 2020 release of Minecraft produce images that are qualitatively different from anything rasterisation achieves.

DLSS, Deep Learning Super Sampling, renders the image at a lower resolution and reconstructs a higher-resolution result with a neural network running on the Tensor Cores. The first version, trained per game, was soft and unconvincing. The second, using temporal information from previous frames and a single general network, was good enough that reviewers began finding reconstructed images preferable to native rendering in some scenes because the network also resolves aliasing. The third generates intermediate frames entirely, and the fourth replaced the convolutional network with a transformer. Competing techniques, AMD's FSR and Intel's XeSS, followed, and reconstruction is now how essentially all demanding games are rendered.

The strategic significance is that Nvidia used its artificial intelligence research to make its graphics products better in a way its competitors could not immediately match, and in doing so made the Tensor Cores that consumers were paying for actually useful to them.

Services and devices

GeForce Now streams games from Nvidia data centres to televisions, phones and low-powered computers, one of the few cloud gaming services to survive after Google closed Stadia. It has repeatedly been complicated by publishers withdrawing titles over licensing.

Shield is a family of Android devices, of which the Shield TV set-top box has been the durable product, valued in the home theatre community for its video processing and its AI upscaling of standard definition content. The Shield Portable and Shield Tablet were discontinued.

GeForce Experience, replaced in 2025 by the Nvidia App, handles driver updates, game settings, recording and streaming. Nvidia's driver engineering, including per-game optimisation profiles delivered before major releases, is a substantial and generally underestimated part of its consumer advantage.

Professional visualisation

The professional line, sold as Quadro until 2020 and as RTX since, serves computer-aided design, film and television, medical imaging, architecture and scientific visualisation. The hardware overlaps heavily with consumer parts; the difference is certified drivers, error-correcting memory on higher models, longer availability and support arrangements that matter to organisations whose work stops when a driver fails.

The segment is worth a few billion dollars a year and has been reshaped twice, first by the collapse in demand during the pandemic and its recovery, and then by the arrival of generative tools in creative work, which has pushed buyers toward parts chosen for memory capacity rather than rendering throughput.

Omniverse is the platform Nvidia has built around this business and is a more ambitious idea than the hardware it runs on. It uses Universal Scene Description, the open format developed at Pixar, as a common representation so that models created in different tools can be combined into one live simulation. The applications are industrial: BMW has simulated whole factories before building them, Foxconn has planned production lines, Siemens has connected it to its industrial software, and Amazon has used it for warehouse robotics. The second use is generating synthetic data, since a physically accurate simulation can produce labelled training examples for a robot far more cheaply than the physical world can.

Nvidia's argument is that every physical process worth automating will first be simulated, that the simulation requires the same hardware as the training, and that the company therefore sells into industrial automation without needing to build robots. The argument is coherent. Adoption outside a set of large and sophisticated manufacturers has been slower than the company's presentations suggest.

Healthcare, science and other markets

Clara and the related Holoscan platform serve medical imaging, genomics and surgical robotics, providing accelerated reconstruction for CT and MRI, real-time processing for endoscopy and ultrasound, and the computing inside medical devices. Parabricks accelerates genomic analysis, reducing whole-genome secondary analysis from a day to under an hour. BioNeMo applies language model techniques to proteins and molecules for drug discovery, and Nvidia has taken positions in and partnered with a number of computational biology companies including Recursion and Schrödinger.

Earth-2 is an effort to build a high-resolution digital twin of the planet's climate, using machine learning emulators including FourCastNet and CorrDiff to produce forecasts far faster than conventional numerical weather prediction. It has been adopted by meteorological agencies in several countries and is one of the clearer demonstrations that learned models can substitute for parts of physical simulation.

cuLitho accelerates computational lithography, the extremely expensive calculation that determines what patterns must be written onto a photomask so that the desired pattern appears on the wafer. TSMC, ASML and Synopsys have adopted it, which places Nvidia's products inside the process that manufactures Nvidia's products.

Telecommunications work covers virtualised radio access networks with Ericsson, Nokia and others, and the AI-RAN initiative, which proposes running radio processing and artificial intelligence inference on the same hardware at cell sites. Nvidia invested approximately US$1 billion in Nokia in 2025 in connection with this.

Quantum computing is addressed through CUDA-Q, a programming model for hybrid classical and quantum systems, and NVQLink, announced in 2025, which connects quantum processors to accelerators with the very low latency that error correction requires. Nvidia's position, which Huang has stated repeatedly and which once moved quantum computing stocks sharply when he expressed scepticism about timelines, is that useful quantum machines will be co-processors attached to conventional supercomputers rather than replacements for them.

Sovereign AI is the name Nvidia gives to sales to national governments building domestic capacity, a category that barely existed in 2023 and that the company has described as worth tens of billions annually. Projects have been announced with France, Italy, Japan, India, Singapore, Denmark, Canada, Saudi Arabia and the United Arab Emirates, the last two on a scale that has drawn security objections in Washington. The argument governments find persuasive is that a country dependent on foreign infrastructure for machine intelligence has surrendered something it will want back.

Business model and economics

Nvidia's economics are unusual enough to be worth setting out directly, because they explain both the size of the profits and the intensity of the effort to displace the company.

Where the margin comes from

An H100 accelerator carried a list price in the range of US$25,000 to US$40,000 depending on configuration and moment, against a manufactured cost widely estimated at a few thousand dollars. Gross margin of roughly three-quarters on a physical product is not normal in semiconductors, where mature businesses run between forty and sixty per cent, and it is not explained by the silicon.

It is explained by three things. The first is that the customer is not buying a chip but the ability to train a model, and the alternative to buying is not a cheaper chip but a model that is worse or later than a competitor's. When the value of the outcome is very large and the window is short, the price of the input is a second-order consideration. The second is that the software and the networking are bundled into the price of the hardware, so margin that in another structure would be recognised as software revenue appears as hardware margin. The third is scarcity: through the peak of the shortage, allocation rather than price cleared the market, and Nvidia's ability to decide who received capacity was itself a source of leverage.

None of these is permanent. The first weakens as the economics of artificial intelligence products are established and buyers become price-sensitive. The second is what the compiler projects attack. The third disappears when supply catches demand, which advanced packaging capacity additions at TSMC are intended to achieve.

The customer problem

Nvidia's largest customers are Microsoft, Amazon, Google, Meta and a small number of specialised cloud operators and laboratories. In several recent quarters two customers have each accounted for more than a tenth of revenue, a concentration the company is required to disclose and which would be regarded as alarming in most businesses.

Each of these customers is also building a competing accelerator, and each has the two things a competitor most needs: complete knowledge of the workloads, and guaranteed internal demand that removes the chicken-and-egg problem a merchant vendor faces. Google's TPUs are the demonstration that this can work, having reached a seventh generation and trained frontier models. That Nvidia nonetheless retains the great majority of the market against customers who would prefer not to pay it is the strongest available evidence for the durability of the software and networking position.

The relationship is further complicated by Nvidia's investments. The company has taken equity in CoreWeave, Nebius, Applied Digital, xAI and, in 2025, committed very large sums to OpenAI, and has entered arrangements to lease back capacity. Critics describe this as circular, in that Nvidia supplies capital to buyers who use it to buy Nvidia products, which flatters reported demand. Nvidia's answer is that the investments are small relative to revenue, that the capacity is genuinely used, and that seeding infrastructure operators is how a market of this kind is built. The truth is probably that both are correct and that the practice becomes dangerous only if demand turns.

Inception and the developer economy

The Inception programme, which supports startups with hardware credits, engineering assistance, marketing and introductions to venture investors, counts many thousands of members. It costs Nvidia relatively little and performs three functions: it seeds demand among companies that will eventually buy at scale, it ensures new software is written against CUDA rather than against an abstraction layer, and it gives Nvidia unusually early sight of what is being built, which informs both product planning and its investments.

The same logic applies to the company's academic programme, its donation of systems to universities, and its funding of the libraries that researchers use. The effect is that a very large proportion of people entering the field learn Nvidia's tools first, and tools learned first are rarely replaced.

The GPU Technology Conference

GTC began in 2009 as a technical conference for people doing general-purpose computation on graphics hardware, attended by a few thousand researchers. It is now among the largest events in the industry, held principally in San Jose with regional editions in Washington, Taipei, Paris and elsewhere, and Huang's keynote is treated by the technology press and by financial markets as a statement of where computing is going.

The format is idiosyncratic. Huang speaks for two hours or more without a teleprompter, moves between product announcements, technical explanation and argument about the direction of the field, and has on occasion carried physical objects on stage, including a Blackwell wafer and a mock-up rack, to make a point about scale. The 2024 keynote was held in the SAP Center, an ice hockey arena, and was compared by attendees and by reporters to a concert, a comparison the company did not discourage.

The substance is that GTC is where Nvidia publishes its roadmap years ahead, which is a deliberate competitive instrument: a customer considering an alternative accelerator must weigh it not against what Nvidia sells today but against what Nvidia has already committed to shipping in eighteen months. Announcing an annual cadence, and then meeting it, is a way of making that calculation unattractive.

The conference also serves as the principal venue for the company's research organisation, which publishes on graphics, simulation, language models, robotics and chip design, and which functions as a recruiting instrument as much as a scientific one.

Workforce, facilities and philanthropy

Nvidia employs roughly 36,000 people, a figure that has roughly doubled since 2020 and that remains small relative to revenue: on a per-employee basis it is among the most productive large companies ever measured, a consequence of being fabless, of selling to a small number of very large customers, and of software leverage.

Compensation is high and heavily weighted toward equity. Reporting in February 2024 indicated that half of Nvidia's employees earned more than US$228,000 in 2023, and the appreciation of restricted stock over the past decade has made a substantial number of long-tenured engineers wealthy, with the attendant retention benefit and the attendant problem that people who no longer need to work sometimes stop. The company was ranked third on Forbes' Best Places to Work list in 2024 and was described in the same period as the most sought-after employer in Silicon Valley at a time when much of the industry was reducing headcount.

The headquarters campus in Santa Clara comprises two buildings, Endeavor, opened in 2017, and Voyager, opened in 2022, both triangular in plan because triangles are the primitive from which the company's products build images. Voyager encloses a landscaped interior described as an indoor park. The design, by Gensler, uses very large skylights arranged so that daylight reaches most interior workspaces. The company has significant operations in Israel, following Mellanox, and engineering centres in India, Taiwan, China, Germany, the United Kingdom and Armenia, the last announced in April 2022 in Yerevan.

Nvidia's philanthropy runs through the Nvidia Foundation, which matches employee giving at a high ratio and runs Compute the Cure for cancer research, and through donations of hardware and software to universities, which is also an effective way of ensuring that the next generation of researchers learns CUDA. The company has committed to renewable electricity for its own operations, though the far larger question is the electricity consumed by the systems it sells.

Energy and environmental questions

The power consumption of artificial intelligence infrastructure has become the principal environmental question attached to Nvidia, and the company's position on it is contested.

A single Blackwell accelerator draws around a kilowatt, and a GB200 NVL72 rack draws on the order of 120 kilowatts, roughly ten times the density of a conventional server rack and enough to require liquid cooling. Data centre electricity demand in the United States, essentially flat for a decade, has begun rising sharply, and utilities in Virginia, Texas, Ohio and Georgia have faced connection queues of years. Several retired fossil generating stations have been kept open or reopened, and large technology companies have contracted directly for nuclear output, including the restart of a reactor at Three Mile Island for Microsoft.

Nvidia's answer is that its products are the efficient way to perform the computation and that the correct measure is energy per unit of work rather than energy per rack. On that measure the argument is strong: each generation has improved performance per watt by a large multiple, an accelerator performs work that would take a very large number of conventional processors far more energy to perform, and Nvidia systems dominate the Green500 efficiency ranking. The counter-argument is the rebound effect, that cheaper computation increases total computation, and that this is precisely what has happened.

Water consumption for cooling, the embodied carbon of manufacturing at advanced nodes, and the electronic waste implications of an annual product cadence are each raised and are each less thoroughly quantified than the electricity figures. Nvidia reports Scope 1, 2 and 3 emissions and has set targets for renewable procurement, but the emissions attributable to the use of its products by others are an order of magnitude larger than its own footprint and are not within its direct control.

Timeline

1993 Founded on 5 April by Jensen Huang, Chris Malachowsky and Curtis Priem.
1995 The NV1 ships and fails, having bet on quadratic surfaces against the triangle-based Direct3D standard.
1997 The RIVA 128 succeeds, selling roughly a million units in four months and saving the company.
1999 Lists on Nasdaq in January; releases the GeForce 256 in October and coins the term GPU.
2000 Acquires the assets of 3dfx; supplies the graphics processor for Microsoft's first Xbox.
2001 The GeForce 3 introduces programmable vertex shaders.
2006 Releases CUDA with the Tesla architecture, opening the graphics processor to general computation.
2008 Launches Tegra; takes a US$200 million write-down over defective mobile chips.
2011 Settles with Intel for US$1.5 billion in a six-year cross-licence.
2012 AlexNet, trained on two GeForce cards, wins ImageNet and begins the deep learning era.
2016 Delivers the first DGX-1 to OpenAI; releases the Pascal architecture.
2017 Volta introduces Tensor Cores; the Nintendo Switch ships with a Tegra X1.
2018 Turing adds hardware ray tracing and DLSS to consumer graphics.
2019 Agrees to acquire Mellanox Technologies for US$6.9 billion.
2020 Completes Mellanox and Cumulus; agrees to buy Arm for US$40 billion; releases Ampere and the A100.
2022 Abandons the Arm acquisition; releases Hopper and the H100; the United States restricts exports of advanced accelerators to China; ChatGPT is released.
2023 Passes a US$1 trillion valuation in May; revenue roughly triples year on year.
2024 Becomes the world's most valuable company in June; joins the Dow Jones Industrial Average in November; announces Blackwell.
2025 Loses approximately US$600 billion of value in one day following DeepSeek's release in January; passes US$4 trillion and then US$5 trillion; the H20 is restricted, permitted and then halted.
2026 Announces the Rubin architecture; the United States permits conditional H200 exports to China while Beijing restricts imports.

Reception and legacy

Nvidia occupies an unusual place in the history of computing because it was, for most of its life, a component supplier that consumers knew by name, and because the component turned out to matter far beyond its original purpose.

The judgement that draws the most comment is Huang's insistence on funding CUDA through years when it earned nothing. It is cited in business schools as an example of a founder's ability to sustain an unpopular position, which is a real advantage of founder control and is usually discussed only in its failure modes. It is worth noting that the same insistence produced Tegra, which failed, and the pursuit of Arm, which was blocked; the record is of conviction applied consistently rather than of conviction that was always right.

Critics point to pricing behaviour in consumer graphics, to the GeForce Partner Program and the Hardware Unboxed episode as evidence of a company willing to use leverage over partners and press, to the GTX 970 specifications and the cryptocurrency disclosures as failures of candour, and to the antitrust investigations now open on three continents. Supporters point to two decades of continuous and mostly correct technical bets, to an execution cadence its competitors have not matched, and to the fact that a very large proportion of the scientific computing performed in the world now runs on software the company gave away.

What is unarguable is the scale of the outcome. A company founded by three engineers in a diner, which nearly failed twice in its first four years, became the most valuable company in the world and the supplier of the physical substrate on which an entire technological era is being built. Whether that era sustains the valuation is not yet known, and the history of the semiconductor industry contains a considerable number of companies that were briefly indispensable.

Assessment

Nvidia is the clearest case in modern business of a company whose position was earned by a technical judgement made long before it paid.

The decision to spend heavily on CUDA from 2006, through years in which the general-purpose computing market was small and the cost was visible in gaming products, was taken on an argument about what parallel computing would eventually be for. It was correct. When deep learning arrived it needed precisely the thing Nvidia had spent a decade building, and no competitor had an equivalent because no competitor had been willing to fund one against no demand.

The second judgement, made after 2012, was to reorganise around neural networks rather than to treat them as an additional market for graphics chips, and to design hardware for them. The third, made in 2019, was to buy Mellanox and thereby to move the unit of sale from the chip to the cluster.

The risks are as clear as the achievement. Demand is concentrated in a handful of buyers whose capital expenditure is discretionary and whose own economics depend on artificial intelligence generating revenue in proportion to its cost, which is not yet demonstrated. Those same buyers are building alternatives, and have every reason to succeed. Manufacturing depends on one company on one island. The software advantage is being attacked from above by compilers that abstract it away. Gross margins in the mid-seventies invite entry by definition. And the valuation embeds an assumption about the quantity of computation intelligence requires that a single efficient model from an unexpected direction can reprice by hundreds of billions of dollars in an afternoon, as January 2025 demonstrated.

What is not in dispute is that for roughly three years Nvidia has been the supplier of a capability that a large part of the world's technology industry has decided it must have, that it has executed on an annual cadence which its competitors have not matched, and that the profits from it have been among the largest ever recorded by a company of its size. Whether that is a durable position or the peak of a cycle is the most consequential open question in the industry, and it will not be settled by argument.