Nvidia: From Graphics Cards to the Most Valuable Company on Earth - Example Timeline | Boromlia
Three men started a graphics chip company at a Denny's in 1993 and nearly went bankrupt twice before anyone had heard of them. The decision that made Nvidia the most valuable company in the world was taken in 2006, for a market that did not exist, and cost the company margin for a decade. This is the whole arc, from the NV1 to a 96 billion dollar quarter. Compiled 5 September 2026.
Three Men in a Denny's
Jensen Huang was thirty and working at LSI Logic. Chris Malachowsky and Curtis Priem were engineers at Sun Microsystems, tired of building workstation graphics for a market that was about to be eaten by the personal computer. They met in a Denny's on Berryessa Road in San Jose, which was the booth Huang had waited tables in as a teenager, and agreed on a proposition: that the PC would become a machine people bought for entertainment, and that entertainment would require a dedicated chip for three-dimensional graphics. They incorporated on 5 April 1993 with about forty thousand dollars between them and no product. Sequoia Capital and Sutter Hill put in roughly twenty million over the following two years. The name came from the internal file prefix NV, for next version, and from invidia, which is Latin for envy.
The NV1 Bets on the Wrong Shape
The first product, the NV1, shipped in 1995 and drew images out of curved quadratic surfaces rather than triangles. The engineering argument for it was good: a curve takes fewer primitives to describe than the triangles that approximate it. The commercial argument against it turned out to be decisive. Microsoft released Direct3D and standardised on triangles, and every game written for Direct3D was therefore written against Nvidia's hardware rather than for it. Sega had built the Saturn around the same quadratic approach and put five million dollars into the company, which bought time but not a market. Nvidia cut its staff from about a hundred to roughly forty. It had one more product in it.
RIVA 128, and Four Months of Payroll
With about four months of cash left, Huang abandoned the architecture the company had been founded on and rebuilt around triangles and Direct3D. He also abandoned the normal way of making a chip. Rather than fabricate silicon and test it, which the company could not afford to do twice, Nvidia built a software simulator of the chip and taped out from the simulation, betting that the emulator was right. The RIVA 128 shipped in August 1997 and sold about a million units in four months. The method survived the crisis and became the company's practice: simulate first, tape out once, ship on a six-month cadence rather than the industry's eighteen.
The IPO
Nvidia listed on the Nasdaq on 22 January 1999 at twelve dollars a share, raising about forty-two million dollars. The company was five years old, had come within months of failing twice, and was one of perhaps thirty firms fighting over a graphics market that most analysts expected Intel to absorb.
The Company Invents a Word
The GeForce 256 launched in the autumn of 1999 with hardware transform and lighting, which moved geometry work off the processor and onto the graphics card. Nvidia marketed it as the world's first GPU, a term the company coined and then defined in its own press release. The engineering was real. The naming was a commercial act, and it worked twice over: it separated Nvidia from the field of graphics accelerators at the time, and it left the industry using a word that Nvidia owned the definition of. Every subsequent argument about what a GPU is has been conducted on the company's terms.
Buying What Was Left of 3dfx
3dfx had defined PC gaming graphics with Voodoo and had been Nvidia's most serious rival. In December 2000 Nvidia bought its assets for about seventy million dollars in cash and a million shares, and 3dfx filed for bankruptcy. A number of its engineers moved across. The competitive field for discrete graphics narrowed to two companies within four years, and it has stayed that way for twenty-five.
The Xbox, and the Argument About the Price
Microsoft's first Xbox shipped in November 2001 with an Nvidia graphics processor, the NV2A, and a contract that made Microsoft the company's largest customer almost overnight. Console volume is a different business from add-in cards: high, predictable, and priced to a schedule agreed years in advance. The schedule was the problem. Microsoft wanted the price cut faster than the contract allowed, the dispute went to arbitration in 2002, and it was settled in February 2003 on terms that left Nvidia off the second Xbox entirely. Microsoft went to ATI. The lesson Nvidia took from it, and repeated for twenty years, was that a single dominant customer sets your margin for you.
CUDA, for a Market That Did Not Exist
Researchers had been coaxing general arithmetic out of graphics cards for years by disguising it as pixel shading, which worked but was miserable. Ian Buck's Brook project at Stanford had shown a cleaner way, and Nvidia hired him. In November 2006 the company launched the G80 architecture, and in 2007 it released CUDA, which let ordinary C code run across the card's parallel processors. The decision that mattered came next. Nvidia put CUDA support into every GPU it manufactured, including the cheap consumer parts, and kept a software organisation employed on it. That cost die area, engineering and gross margin on every card sold, for a customer base that in 2007 consisted of a few thousand academics. Huang defended it to analysts for the better part of a decade. It is the reason that when the demand arrived, the software was already installed on every machine that mattered.
Tesla: A Graphics Card With No Screen Output
In 2007 Nvidia began selling the Tesla line, which was a GPU with the display connectors removed. There was nothing to look at, because the customer was a computational chemist or a seismic modeller who wanted the arithmetic. It was the first admission, in the product catalogue rather than in a speech, that the company was no longer only in the graphics business. For six years it was a rounding error on the accounts.
A 196 Million Dollar Packaging Fault
In July 2008 Nvidia disclosed a charge of 196 million dollars to cover notebook and desktop chips failing in the field. The cause was the packaging material used to bond the die, which cracked under the heating and cooling cycles of a laptop. The shares fell about thirty per cent in a day and class actions followed. The episode is worth keeping in view because the company that would later be valued in trillions spent 2008 and 2009 being sued over solder, losing its chipset business to an Intel licensing dispute, and watching a global financial crisis take the PC market down with it.
Tegra, the Bet That Did Not Come In
Nvidia's answer to the smartphone was Tegra, a system on a chip that won the Motorola Xoom, the Nexus 7 and a run of Android tablets. It lost the war for a reason that had nothing to do with graphics: Qualcomm integrated the cellular modem onto the same die, and a phone maker will take a worse GPU over a second chip and a second power budget every time. Nvidia withdrew from phones in 2014 and redirected Tegra into cars and, eventually, the Nintendo Switch. Two bets were running in parallel through these years. The mobile one was the obvious one, and it failed. The other was CUDA.
AlexNet, in Which Nvidia Had No Part
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton entered the 2012 ImageNet competition with a convolutional neural network and won it by a distance that ended the argument: a top-five error rate of 15.3 per cent against 26.2 per cent for the runner-up. The network was trained on two GeForce GTX 580 cards bought at retail, in a bedroom, because the university could not supply the compute. Nvidia had nothing to do with the paper. It had simply spent six years making sure that a graduate student with three hundred dollars could program a consumer graphics card in C. The most consequential thing that ever happened to the company happened to it rather than because of it, and the company recognised it faster than anyone else did.
Huang Delivers the First DGX-1 by Hand
Nvidia announced the DGX-1 in April 2016: eight Tesla P100s, a hundred and twenty-nine thousand dollars, and a machine that existed only to train neural networks. In August, Huang carried the first one into OpenAI's offices in San Francisco and signed the case. The photograph of that delivery is now used to date the beginning of the modern AI industry, which is a retrospective flattery. What it actually shows is a supplier making sure that the most interesting customer in his market started on his architecture.
Crypto, and the Revenue Nobody Could Attribute
Ethereum mining absorbed consumer graphics cards through 2017 and into 2018, emptied the retail channel and raised prices. Nvidia told investors it could not reliably separate the demand from gaming demand, which was true, and reported it inside the gaming line, which the Securities and Exchange Commission later decided was inadequate. In May 2022 the company paid 5.5 million dollars to settle those charges without admitting them. The correction came first. When mining stopped, the channel was full of cards nobody wanted. Nvidia missed its own guidance in November 2018 and the shares lost roughly half their value in the last quarter of the year.
Mellanox, Which Turned Chips Into Racks
In March 2019 Nvidia agreed to buy Mellanox for 6.9 billion dollars, outbidding Intel. Mellanox made InfiniBand and high-speed Ethernet, which is to say it made the wiring that lets a room full of accelerators behave as one machine. The deal closed in April 2020. It is the least glamorous acquisition in the company's history and probably the most important. Without it Nvidia sells chips. With it Nvidia sells the rack, the network and the reference design, and the customer buys the whole system or none of it.
The Arm Deal
In September 2020 Nvidia agreed to buy Arm from SoftBank for up to forty billion dollars, which would have been the largest semiconductor acquisition ever made. Arm's designs sit inside almost every phone on earth and are licensed to companies that compete with Nvidia, and that was the whole objection. The Federal Trade Commission sued to block it in December 2021. Regulators in the United Kingdom, the European Union and China raised their own concerns. In February 2022 the parties abandoned the transaction and Nvidia forfeited a prepayment of 1.25 billion dollars. Eighteen months of executive attention bought a write-off and a licensing agreement.
ChatGPT, Trained on Somebody Else's Hardware
OpenAI released ChatGPT on 30 November 2022. It reached a hundred million users in two months, which is a consumer statistic, but the consequence was an industrial one. Every company operating data centres rewrote its capital expenditure plan, and the thing all of those plans needed was the accelerator that the models had been trained on. Nvidia's own 2022 had been poor. Crypto had collapsed again, the gaming channel was full, and in September the United States had restricted exports of the A100 and H100 to China. The company went into the year that changed it looking like a business with a cyclical problem.
The Guidance That Moved 184 Billion Dollars
On 24 May 2023 Nvidia reported a quarter that was merely good and then gave guidance that was not. It told the market to expect about 11 billion dollars of revenue in the following quarter. The consensus estimate was 7.2 billion. The shares rose about 24 per cent the next day and added roughly 184 billion dollars of market value in a single session, at the time the largest one-day gain any company had recorded. What had changed was not the technology, which had been shipping for a year. What had changed was that the demand had finally shown up in a forecast, and the forecast was public.
One Trillion Dollars
On 30 May 2023 Nvidia touched a market capitalisation of one trillion dollars, six days after the guidance. It was the sixth American company to do it, and the first that most people outside the industry had never heard of.
Two Trillion, Nine Months Later
Nvidia passed two trillion dollars on 23 February 2024. Apple had taken two years to go from one trillion to two, and Microsoft about the same. Nvidia took nine months. The figure that explains it is not the share price. In the fiscal year that ended in January 2024, data centre revenue was 47.5 billion dollars, more than three times the year before, in a segment that had been a rounding error a decade earlier.
Blackwell, and the Rack as the Unit of Sale
At GTC in March 2024 Nvidia announced the Blackwell architecture and, more to the point, the GB200 NVL72: seventy-two GPUs and thirty-six CPUs wired together with NVLink and sold as one liquid-cooled rack. Mellanox is what made that possible, four years after the acquisition closed. The commercial effect is that the price of the smallest thing Nvidia sells to a serious customer went from thousands of dollars to millions, and the competitive question stopped being whose chip is faster and became whose rack a data centre is built around.
The Most Valuable Company on Earth
Nvidia crossed three trillion dollars on 5 June 2024, split its stock ten for one on 10 June to make the shares purchasable by ordinary people again, and on 18 June closed as the most valuable public company in the world, ahead of Microsoft and Apple. It held the position for a matter of days on that occasion. It has traded in and out of first place ever since, which is the more interesting fact: the three largest companies in the world are now a software firm, a consumer hardware firm, and their common supplier.
DeepSeek, and 589 Billion Dollars in a Day
A Chinese laboratory published a model that performed comparably to the American frontier systems and claimed to have trained it for a fraction of the compute. On 27 January 2025 Nvidia fell about 17 per cent and lost roughly 589 billion dollars of market value, the largest single-day loss any company has recorded. The argument the market was having was whether efficiency reduces demand for the hardware or increases it. Nvidia's position, that cheaper inference means more inference, was not testable that week. Revenue over the following eighteen months supported it, which is not the same as proving it.
The China Write-Down
In April 2025 the United States required a licence for the H20, the deliberately reduced accelerator Nvidia had designed to stay inside the previous set of export rules. The company took a charge of 4.5 billion dollars, said 2.5 billion of revenue had gone unshipped in the quarter, and warned of roughly 8 billion more in the next. By July the administration had signalled that licences would be granted and sales could resume. The episode established the pattern that has governed the China business since: the product is redesigned to the rule, the rule moves, and the company writes off the difference.
Four Trillion
On 9 July 2025 Nvidia became the first company of any kind to be worth four trillion dollars. Thirty-two years earlier the same company had been three men with forty thousand dollars and an idea about triangles that turned out to be wrong.
Five Trillion, and Half a Trillion in Bookings
Nvidia closed above five trillion dollars on 29 October 2025, again the first company ever to do it. Huang told the audience at the company's own conference that Nvidia was carrying about half a trillion dollars of Blackwell and Rubin bookings through the end of 2026. That figure is the one worth watching rather than the market capitalisation. A share price is an opinion. An order book is a contract.
Fiscal 2026: 215.9 Billion Dollars
Reported on 25 February 2026. Revenue for the year was 215.9 billion dollars, up 65 per cent. Data centre revenue was 193.7 billion, up 68 per cent, and the fourth quarter alone was 62.3 billion. Gaming, the business the company was founded to be in, was 16 billion for the year, up 41 per cent, and now accounts for about seven per cent of revenue. Diluted earnings per share for the year were 4.90 dollars. Huang's framing on the call was that the inference inflection point had arrived and that Grace Blackwell delivered an order of magnitude lower cost per token.
Vera Rubin, and a One Trillion Dollar Projection
Huang opened GTC 2026 at the SAP Center in San Jose on 17 March. The announcement was Vera Rubin, a multi-rack system pairing Rubin GPUs with Nvidia's own Vera CPUs, quoted at ten times the inference throughput per watt of Blackwell at a tenth of the cost per token, with availability from the second half of 2026. The number that carried the keynote was not a specification. Huang raised the company's cumulative revenue projection to one trillion dollars through 2027. A supplier forecasting a trillion dollars of its own sales is a claim about the entire capital cycle of an industry, made by the party with the most to gain from believing it.
A Trillion Dollars, Backwards
The shares peaked on 14 May 2026, at a record of about 5.5 trillion dollars. Over the following eight weeks they fell roughly 16 per cent and gave back something close to a trillion dollars of market value. Nothing in the business had deteriorated. Analysts kept raising their earnings estimates and Nvidia still held around 97 per cent of the server GPU market. What moved was where investors wanted to own the boom: money rotated into memory and storage, where high bandwidth memory prices were rising and Micron was up 229 per cent on the year. The bottleneck had shifted one component to the left, and the market repriced accordingly.
96.2 Billion Dollars in Three Months
Reported on 26 August 2026, for the quarter that ended on 26 July. Revenue was 96.2 billion dollars, up 106 per cent on the same quarter a year earlier. Data centre revenue alone was 89.0 billion, up 117 per cent. Diluted earnings per share were 2.46 dollars on a GAAP basis. Guidance for the following quarter is 108 billion dollars. For scale: the company now books more revenue in one quarter than it did in the whole of fiscal 2024, and its data centre segment alone is larger than the entire semiconductor industry was in the year Nvidia was founded. Huang's summary on the call was that compute is revenue, and he forecast about 70 per cent revenue growth in fiscal 2028, well above what analysts had modelled.
Where This Came From, and When
Compiled 5 September 2026. Financial figures come from Nvidia's own quarterly and annual results releases at nvidianews.nvidia.com. Market capitalisation milestones and single-session moves are taken from contemporaneous reporting by CNBC, TechCrunch, Forbes and Reuters, and are stated as of the day they occurred. The Vera Rubin specifications are the company's own claims from GTC 2026 and have not been independently benchmarked here. The 2026 drawdown figures come from market reporting rather than from Nvidia. The early history draws on published accounts of the founding, the NV1, the RIVA 128 and the 3dfx acquisition. Photographs are Wikimedia Commons files of the campus, the hardware and the people. There is no free photography of most of the events on this page.