# A History of NVIDIA — From the 1993 Founding Through GeForce and CUDA to an AI Infrastructure Company, Read From Primary Sources

> NVIDIA's history from primary sources, from NV1, GeForce and CUDA to Vera Rubin, and how Data Center grew from 4.8% to 89.7% of revenue in its 10-K filings.

- Published: 2026-11-17
- Author: 友田 陽大
- Tags: AI, インフラ, 技術選定, アーキテクチャ設計
- URL: https://tomodahinata.com/en/blog/nvidia-history-from-gpu-to-ai-infrastructure
- Category: NVIDIA, GPUs & AI chips
- Pillar guide: https://tomodahinata.com/en/blog/nvidia-gpu-ai-tensor-core-fp8-hbm-nvlink-explained

## Key points

- NVIDIA was founded on April 5, 1993 by Jensen Huang, Chris Malachowsky and Curtis Priem. Its first product, NV1 (1995), bet on a proprietary approach that drew curved surfaces (quadric patches) instead of triangles. When the industry standardized on Direct3D and OpenGL, NV1 lost its market, and 1996 revenue was only 3.91 million dollars.
- RIVA 128 (1997) switched to the industry-standard Direct3D and kept the company alive. In August 1999 NVIDIA began shipping GeForce 256 and called it, by its own definition, the world's first GPU. In November 2006 it announced CUDA, which let programmers use GPUs for general-purpose computing from C.
- AlexNet (2012) was trained for five to six days on two gaming GTX 580 cards with 3GB of memory each. It was not built on a product NVIDIA designed for deep learning, but the company's investment in general-purpose computing was ready to absorb that turning point.
- In NVIDIA's 10-K revenue by market, Data Center rose from 4.8% of revenue in FY2014 (Jan 2013 to Jan 2014) to 89.7% in FY2026 (Jan 2025 to Jan 2026). Gaming revenue also grew 10.6x: gaming did not shrink, Data Center simply outgrew it by a wide margin.
- It took about 11 years from the CUDA announcement until Data Center revenue first exceeded 1 billion dollars. Generations from Hopper on show up in revenue within one to two years of shipping. This analysis looks only at bets that survived; failures such as NV1 and the abandoned Arm acquisition are part of the same history.

---

Today NVIDIA is usually described as "the AI chip company." When it was founded in 1993, though, its goal was 3D graphics for personal computers, and its first product failed.

This article traces how NVIDIA went **from a graphics-chip company to an AI infrastructure company that sells entire data centers**, from its founding to November 2026. It is a history of the company and its technology, not a biography.

**Quick facts**: founded April 5, 1993 (three founders) / CUDA announced November 2006, Toolkit 1.0 June 2007 / headquarters: Santa Clara, California

**The short answer** is that NVIDIA's history can be explained by three bets.

1. **Ride the industry standard** (1997): after failing with the proprietary NV1, NVIDIA survived by building RIVA 128 around Microsoft's Direct3D.
2. **Turn the GPU into a programmable, general-purpose computer** (2001–2007): programmable shaders, then unified shaders, then CUDA. That investment was ready when the deep-learning boom arrived with AlexNet in 2012.
3. **Sell data centers, not chips** (2016 on): NVIDIA combined HBM, NVLink, Tensor Cores and Mellanox networking, and began selling systems at the scale of a whole rack.

As a result, in the revenue-by-market tables of NVIDIA's 10-K (the annual report filed with the U.S. Securities and Exchange Commission), Data Center rose **from 4.8% of revenue in FY2014 to 89.7% in FY2026**.

**About the numbers in this article.** Every revenue figure is copied from the text of NVIDIA's 10-K filings; shares and growth rates are my own calculations from those figures (the code is in Section 13). NVIDIA's fiscal year (FY) **ends in late January**. FY2026, for example, ran from January 27, 2025 to January 25, 2026, which is roughly calendar 2025. I always distinguish fiscal years from calendar years, and for products I distinguish "announced," "in full production," "shipping" and "generally available."

## 1. What is NVIDIA? The big picture in one timeline

First, the overview. NVIDIA is a "fabless" company: it designs semiconductors but owns no fabs, and outsources manufacturing. According to its FY2026 10-K, it had about 42,000 employees at the end of that year, about 31,000 of them in research and development.

| When (calendar) | Event | Significance |
|---|---|---|
| April 1993 | Founded by three people (April 5) | Founding |
| May 1995 | First product, NV1 | A proprietary bet that failed |
| August 1997 | RIVA 128 ships in volume | Survived by supporting Direct3D |
| January 1999 | Initial public offering (IPO) | 12 dollars per share |
| August 1999 | GeForce 256 begins shipping | "World's first GPU" by NVIDIA's definition |
| January 2001 | GeForce3 begins shipping | Programmable shaders |
| November 2006 | GeForce 8800 (G80) and CUDA announced | Unified shaders; general-purpose computing in C |
| June 2007 | CUDA Toolkit 1.0 | First official release |
| 2009–2010 | Fermi | ECC, double precision, caches: built as a computer |
| 2012 | AlexNet (trained on two GTX 580s) | Turning point for deep learning |
| April 2016 | Tesla P100 announced | HBM2, NVLink |
| May 2017 | Tesla V100 announced | Tensor Cores |
| March 2019 | Mellanox acquisition announced | Bringing networking in-house |
| April 2020 | Mellanox acquisition completed | Becoming a data-center company |
| March 2022 | Hopper (H100) announced | Workhorse of generative AI |
| March 2024 | Blackwell announced | Rack-scale systems |
| January 2026 | Vera Rubin announced as "in full production" | Latest generation as of November 2026 |

On November 11, 2026, co-founder Jensen Huang resigned as CEO, and CFO Colette Kress became interim CEO. I cover the events and the governance questions in [Jensen Huang's resignation as NVIDIA CEO](/blog/jensen-huang-resignation-nvidia-ceo-transition-governance). This article focuses on the company's 33-year path.

## 2. The 1993 founding: three engineers and the PCI bus

### The founders and the founding date

NVIDIA's Form S-1 for its 1998 IPO gives its inception date as "April 5, 1993." According to the same filing, the founders were:

- **Jensen Huang (Jen-Hsun Huang)**: President and CEO. Before NVIDIA he worked at LSI Logic (1985–1993) and AMD (1983–1985).
- **Chris A. Malachowsky**: Vice President, Engineering. He had worked at Sun Microsystems and Hewlett-Packard, and was a co-inventor of Sun's GX graphics.
- **Curtis R. Priem**: Chief Technical Officer. At Sun Microsystems he architected the GX graphics products.

Two of the three were graphics engineers from Sun, and that shaped the design of the first product.

### Why 1993?

PCs at the time were moving the bus that connects expansion cards from the old ISA (PC/AT) standard to the new PCI standard. David Rosenthal, who worked on the architecture in NVIDIA's early days, recalled in a 2025 blog post:

> "The PC/AT bus' bandwidth was completely inadequate for 3D games, but the PCI bus had considerably more. Whether it was enough was an open question."

Think of it as a road that had just been widened, but might still jam under the heavy traffic of 3D games. **How to get the most out of limited bandwidth** became the design theme of the first product.

## 3. Why NV1 failed: curves versus triangles

### What NV1 tried to do

In May 1995 NVIDIA released its first product, NV1. The S-1 describes it as "a multimedia accelerator that provided 3D graphics, video and audio," "targeted primarily to the game console market."

The surface of a 3D object is normally represented as **a mesh of triangles**. A triangle is the simplest shape that describes a surface, and any curved surface can be approximated with enough small triangles. Making a smooth surface such as a sphere look good, however, takes a great many triangles.

NV1 took a different path. It rendered **curved surfaces directly (quadric patches)**. NVIDIA marketed this as "quadratic texture mapping." Rosenthal explains the reasoning:

- Describing one curved patch takes more data than describing one triangle.
- But you need so many fewer patches for the same visual result that the data per frame drops substantially.
- So the approach makes better use of limited PCI bus bandwidth.

In other words, NV1's approach **made technical sense**. It was a rational answer to the bandwidth constraint of the day.

### The Sega connection

According to Rosenthal, Sega was the only company in the video game industry using quadric patches at the time, and NV1's selling point was that it could run Sega's games on a PC. Jon Peddie, a graphics-industry historian who writes for the IEEE Computer Society, says Sega of America signed an exclusive licensing agreement with NVIDIA and announced that it would port Sega games to PCs equipped with NV1.

### Why it failed

The S-1 states the reason frankly:

> "The NV1 was developed in the absence of industry standards with the goal of establishing the Company's proprietary NV technology as a 3D graphics standard. By the end of 1996, the PC industry had broadly adopted Microsoft's Direct3D and Silicon Graphics Inc.'s ("SGI's") OpenGL 3D APIs. As a result, the Company experienced a significant reduction in revenue from sales of the NV1"

Both Direct3D and OpenGL are triangle-based APIs (the contract software uses to ask hardware to draw). Once game developers built with triangles, a chip designed for curved patches had no use. Peddie writes that NVIDIA and its partners "couldn't get game developers to invest in the NV1 quadratic texture mapping and Nvidia's SDK for it." Rosenthal recalls that "once Microsoft announced Direct X it was obvious to me that Nvidia was doomed if the next chip did quadric patches." The APIs were not the only problem. Peddie also cites criticism that NV1's 2D performance could not compete with other products of the day, especially in DOS, that its audio quality fell short of expectations, and that the board was expensive.

According to the S-1, NVIDIA stopped selling NV1, also ceased development of NV2, a chip for a game console platform, and moved on to RIVA 128. Peddie writes that NV2 was for Sega, and that losing it nearly put the company out of business (the S-1 itself names the NV2 customer only as "a game console platform").

The numbers show it. Per the S-1, revenue was 1.18 million dollars in 1995 and 3.91 million dollars in 1996. All revenue in those two years came from NV1, and in 1995 the cost of revenue exceeded revenue. Customers were concentrated, too: the board maker Diamond alone accounted for 86% of revenue in 1995 and 82% in 1996.

### What survived from NV1

Not all of NV1 was wasted, however. Rosenthal argues that NV1's real success was its **I/O architecture**: hardware functions were exposed as "objects," and when a chip lacked a function, software (a resource manager) executed it instead. Because software could fill in for hardware that was not finished, NVIDIA could release new chips faster.

**The lesson of NV1 is that technical merit does not sell if the ecosystem (developers, APIs, software) does not follow.** That lesson carries through to the design of CUDA later on.

## 4. RIVA 128: surviving by riding the standard

NVIDIA introduced RIVA 128 in April 1997 and began shipping it in volume in August 1997 (per the S-1; the same filing's MD&A says it was "introduced" in August, so the wording is inconsistent). The S-1 says the chip "is designed to be compatible with Microsoft's Direct3D" — the switch to the standard that NV1 had missed.

The S-1 lists 1.5 million polygons per second and a fill rate of 100 million texture-mapped pixels per second (both company specifications). RIVA 128 sold, and 1997 revenue reached **29.07 million dollars**, about 7.4 times the previous year's 3.91 million (my calculation).

NVIDIA went public in January 1999 (3.5 million shares at 12 dollars each). In between, in 1998, it moved its fiscal year-end from December 31 to January 31 (effective January 1, 1998), and then, effective February 1, 1998, to a 52- or 53-week year ending on the last Sunday in January (FY1999 10-K). That is the origin of the late-January fiscal year the company still uses.

## 5. GeForce 256 and the word "GPU" (1999)

### Defining "GPU"

In August 1999 NVIDIA began shipping GeForce 256. Its FY2000 10-K calls it "The world's first GPU" and explains:

> "GeForce 256 was the first to incorporate hardware transform and lighting engines along with four-pixel per clock rendering pipelines in a single chip."

Step by step:

- **Why it was needed**: a 3D game transforms the coordinates of each vertex into a position on screen (transform) and computes how light falls on it (lighting). Until then, the CPU did this work.
- **An analogy**: in a kitchen, the CPU had been doing the chopping (transform and lighting) while the graphics chip only did the cooking (drawing pixels). GeForce 256 took over the chopping as well.
- **The definition**: NVIDIA defined a GPU (Graphics Processing Unit) as "a processor that integrates T&L and rendering on one chip."

Keep in mind that **"the world's first GPU" is a claim based on NVIDIA's own definition**. A 2008 IEEE Micro paper written by NVIDIA engineers also says "the first GPU was the GeForce 256," but that too is the company's perspective. This article does not claim that GeForce 256 was the first use of the term "GPU."

Per the 10-K filings, revenue was 374.51 million dollars in FY2000 (February 1999 to January 2000) and 735.26 million dollars in FY2001 (January 2000 to January 2001), roughly doubling in a year.

### Back into game consoles: Xbox

NV1 and NV2 had missed the console market, but on March 5, 2000 NVIDIA signed an agreement with Microsoft to develop graphics chips for its Xbox console. According to the FY2001 10-K, Microsoft paid 200 million dollars in April 2000 as an advance against chip purchases.

## 6. Programmable shaders (2001): from fixed processing to code you can write

### The limits of fixed function

GPUs up to GeForce 256 were **fixed-function**. The way transform and lighting were computed was baked into the circuits; developers could only change parameters.

GeForce3, which began shipping in January 2001, changed that. Developers could write small programs called **shaders** and decide for themselves what to compute for each vertex or pixel.

### Was it "fully programmable"?

The sources disagree here. The FY2001 10-K calls GeForce3 "the first GPU to offer full vertex and pixel shader programmability." A 2008 IEEE Micro paper written by NVIDIA's own engineers is more careful:

> "In 2001, the GeForce 3 introduced the first programmable vertex processor executing vertex shaders, along with a configurable 32-bit floating-point fragment pipeline … The Radeon 9700, introduced in 2002, featured a programmable 24-bit floating-point pixel-fragment processor"

So with GeForce3, the vertex side was truly programmable while the pixel side was only "configurable." A processor that could freely compute pixels in floating point came first from competitor ATI (Radeon 9700, 2002). **The GPU's evolution was not NVIDIA's work alone; it advanced through competition.**

### Researchers start using GPUs for things other than graphics

Once shaders were programmable, researchers began using GPUs for computation unrelated to graphics. They treated matrices and simulation data as "images" and had shaders compute on them. The field became known as GPGPU (general-purpose computing on GPUs), and by 2007 it was mature enough for a survey paper in the journal Computer Graphics Forum.

The method was awkward, though: every computation had to be translated into the language of "drawing." CUDA removed that awkwardness.

## 7. G80 and CUDA (2006–2007): computing without the language of graphics

### Unified shaders: removing the workload imbalance

In November 2006 NVIDIA announced the GeForce 8800 (code name G80). The IEEE Micro paper explains the goal.

Earlier GPUs had separate processors for vertices and for pixels. Because there are many more pixels than vertices, pixel processors outnumbered vertex processors by roughly three to one. Real games, however, shift the ratio of work from scene to scene. In a scene full of small triangles, vertex processing becomes the bottleneck while pixel processors sit idle.

G80 therefore used **unified shaders**: the same processors handle both vertices and pixels. It is like a store that used to have separate cashiers and shelf stockers, and instead trains everyone to do both so staff can move to wherever the line is longest. GeForce 8800 organized 128 processors (streaming processors) into 16 SMs (streaming multiprocessors).

There is prior art here, too. The same paper notes that "the XBox 360 introduced an early unified GPU in 2005." NVIDIA's Fermi whitepaper describes G80 as "the first GPU to replace the separate vertex and pixel pipelines with a single, unified processor," but once game consoles are included, G80 was not the first unified design.

### CUDA: using the GPU from C

With unified shaders, the GPU became "a collection of processors that can compute anything." **CUDA** (Compute Unified Device Architecture) made it possible to program that hardware directly, without going through the language of graphics.

The FY2007 10-K says:

> "In November 2006, we began seeding the industry with our GeForce 8 GPU, enabled with CUDA, a technology and compiler that, for the first time, allows for programmers to write C language applications for GPUs."

Here is the timeline:

| Stage | When | Source |
|---|---|---|
| Announcement | November 2006 (Q4 FY2007) | FY2007 10-K, IEEE Micro paper |
| First official toolkit | June 2007 (CUDA Toolkit 1.0) | NVIDIA CUDA Toolkit Archive |
| Compute products (such as Tesla C870) begin shipping | Q3 FY2008 (July 30 to October 28, 2007) | FY2008 10-K |

The Fermi whitepaper lists what G80 introduced: C language support, SIMT (an execution model in which one instruction drives many threads at once), and memory shared among threads with barrier synchronization. These remain basic concepts of CUDA programming in 2026. CUDA Toolkit 13.5.0 shipped in November 2026; the line has continued for more than 19 years.

**CUDA's significance is the mirror image of the NV1 lesson.** NV1 failed by forcing a new approach on developers. CUDA let developers start from C, which they already knew, and kept the same programs running on later generations of GPUs. The code and know-how developers accumulated became, in themselves, a reason to keep choosing NVIDIA GPUs.

### It still did not sell right away

The revenue figures reveal something surprising. Total revenue was 4.10 billion dollars in FY2008 (January 2007 to January 2008) and 4.28 billion dollars in FY2013 (January 2012 to January 2013): almost no growth in five years. The market-platform table, first disclosed in the FY2016 10-K and going back to FY2014, shows that even in FY2014 (January 2013 to January 2014) Data Center (then written "Datacenter") revenue was just **199 million dollars, 4.8% of the total**. For years after its announcement, CUDA was not a business that moved the company's revenue.

## 8. Fermi (2009–2010): building it as a "computer"

NVIDIA's 2009 Fermi whitepaper lists what GPU-computing users had asked for: double-precision (64-bit) performance, ECC (a mechanism that detects and corrects bit errors in memory), a true cache hierarchy, more shared memory, faster atomic operations, and more.

ECC matters because when thousands of GPUs run in a data center, memory bit errors stop being negligible. In a game, an error means a small glitch in one frame; in scientific or financial computing, it corrupts the result.

Fermi responded with the following (as stated in the whitepaper):

- "First GPU with ECC memory support"
- 8x the peak double-precision performance of the previous generation (GT200)
- Full IEEE 754-2008 (the international floating-point standard) 32-bit and 64-bit precision
- A unified address space that supports C++
- The first Fermi GPU: 3.0 billion transistors and up to 512 CUDA cores

**Fermi is the generation that turned the GPU from "a graphics chip that can also compute" into "a computer you put in a data center."** The change came from the requests of supercomputing and scientific users, not gamers.

## 9. AlexNet (2012): what two gaming GPUs changed

### What the paper says

In 2012, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton of the University of Toronto presented an image-recognition paper at NeurIPS (then NIPS): the network now known as AlexNet. The paper establishes these facts:

- The network has 60 million parameters and 650,000 neurons, with five convolutional layers and three fully connected layers.
- It won the ILSVRC-2012 image-recognition competition with a top-5 test error of **15.3%**. The second-best entry scored **26.2%**. The 15.3% figure is for the entered variant that averaged the predictions of seven CNNs, not for a single network.
- Training used "**two GTX 580 3GB GPUs**" and took **five to six days**.
- A single GTX 580 had only 3GB of memory and the network did not fit on one card, so it was split across two GPUs.
- The code was released under the name "cuda-convnet."

The GTX 580 is one of NVIDIA's **gaming** products (GeForce). NVIDIA's list of CUDA GPUs puts it at compute capability 2.0, the same class as the data-center Tesla C2050, that is, the Fermi generation.

### Why it was a turning point

AlexNet was not built on a product NVIDIA designed for deep learning. Researchers built it with off-the-shelf gaming GPUs and CUDA, which anyone could use. The accurate way to put it is that **NVIDIA's long investment in general-purpose computing was ready to absorb a turning point that happened outside the company.**

The paper even contains lines that foreshadow where NVIDIA's products would go:

> "All of our experiments suggest that our results can be improved simply by waiting for faster GPUs and bigger datasets to become available."

It also explains that the model did not fit in 3GB and had to be split across two cards, and that it relied on GPUs being able to read and write each other's memory without going through host memory. **Memory capacity and GPU-to-GPU communication**, the problems that HBM and NVLink would later tackle, were already visible here.

## 10. Pascal (2016) and Volta (2017): designing for deep learning

From here on, NVIDIA designed its GPUs with deep learning as a premise. The same series covers how each technology works, with formulas, in [Why NVIDIA GPUs are so strong at AI](/blog/nvidia-gpu-ai-tensor-core-fp8-hbm-nvlink-explained); here I focus on their place in the history.

### Tesla P100 (announced April 2016): HBM2 and NVLink

The whitepaper for Tesla P100, announced on April 5, 2016, highlights two things:

- **HBM2**: "the world's first GPU architecture to support HBM2 memory." Memory dies are stacked and placed in the same package as the GPU, giving 180GB/s per stack, or a theoretical 720GB/s with four stacks (my calculation).
- **NVLink**: a dedicated link connecting GPUs at up to 160GB/s (bidirectional total), five times PCIe Gen 3 x16.

It was a direct answer to the "memory" and "GPU-to-GPU communication" problems that AlexNet had exposed. According to the 10-K for FY2017 (February 2016 to January 2017), NVIDIA began shipping the DGX-1 AI server that year, and Data Center revenue grew 145% over the previous year.

### Tesla V100 (announced May 2017): Tensor Cores

With V100, announced on May 10, 2017, NVIDIA introduced **Tensor Cores**: dedicated units that compute the matrix multiplications that dominate deep learning, a small matrix at a time. Per the whitepaper, V100 has 640 Tensor Cores and delivers up to 125 Tensor TFLOPS (theoretical peak, FP16 multiply with FP32 accumulate), up to 12 times the peak of standard FP32 operations on P100.

According to the 10-K for FY2018 (January 2017 to January 2018), V100 began shipping in the first half of FY2018, and Data Center revenue rose 133% to 1.932 billion dollars. **Within the period for which market-platform figures exist (FY2014 on), FY2018 is the first year Data Center revenue exceeded 1 billion dollars.**

From this point, the GPU began changing from "a chip whose graphics circuits can also be used for computing" into "a chip with circuits dedicated to AI computation."

## 11. The Mellanox acquisition (announced 2019, completed 2020): from GPU company to data-center company

### Dates and amounts

From the 8-K filings and press releases:

| Stage | Date | Details | Source |
|---|---|---|---|
| Agreement signed | March 10, 2019 | 125 dollars in cash per Mellanox share | 8-K |
| Announced | March 11, 2019 | Enterprise value of approximately 6.9 billion dollars | NVIDIA press release |
| Completed | April 27, 2020 | Approximately 7.0 billion dollars | 8-K, NVIDIA press release |

The difference between 6.9 and 7.0 billion dollars reflects two different measures: enterprise value at announcement and transaction value at completion.

### Why buy a networking company?

Mellanox was an Israeli company that made InfiniBand, the high-speed interconnect used in supercomputers, as well as high-speed Ethernet products. Its 2019 revenue was 1.3306 billion dollars (Mellanox 10-K).

In the announcement press release, Huang said:

> "Addressing this demand will require holistic architectures that connect vast numbers of fast computing nodes over intelligent networking fabrics to form a giant datacenter-scale compute engine."

When a large AI model is trained, thousands of GPUs constantly exchange results. No matter how fast each GPU is, the whole system is only as fast as the network between them allows. If NVLink connects the GPUs inside a server, Mellanox connected the servers.

According to the 10-K for FY2021 (January 2020 to January 2021), Mellanox contributed 10% of the company's total revenue that year. The FY2026 10-K says NVIDIA has about 6,000 employees in Israel and the surrounding region, who mainly support research and development and other work on its networking products.

### An acquisition that did not work: Arm

Not every acquisition succeeded. In September 2020 NVIDIA agreed to buy Arm from SoftBank Group. On February 8, 2022, the agreement was terminated "because of significant regulatory challenges." According to the FY2023 10-K, NVIDIA wrote off the prepayment made at signing and recorded a 1.35 billion dollar charge.

## 12. Hopper, Blackwell and Vera Rubin: separating "announced," "in production" and "shipping"

Recent generations move from announcement to shipment quickly, and press coverage often blurs the stages. I use only the wording in official documents. For specifications, see [the series pillar article](/blog/nvidia-gpu-ai-tensor-core-fp8-hbm-nvlink-explained).

| Generation | Announced | Full production | Shipping | Source |
|---|---|---|---|---|
| Hopper (H100) | March 22, 2022 | September 20, 2022 (H100 announced as "in full production") | Began shipping during FY2023 (January 31, 2022 to January 29, 2023) | Two press releases, FY2023 10-K |
| Blackwell | March 18, 2024 | — | Production systems began shipping in Q4 FY2025 (October 28, 2024 to January 26, 2025) | Press release, FY2025 10-K |
| Blackwell Ultra (GB300) | — | — | Production units began shipping in Q2 FY2026 (April 28 to July 27, 2025) | FY2026 10-K |
| Vera Rubin | January 5, 2026 (CES) | Announced as "in full production" on January 5 and March 16, 2026 | Production shipments began in Q3 FY2027 (between July 27, 2026, when the quarter began, and August 26, when the 10-Q was filed) | Press releases, 10-Q filed August 26, 2026 |

A "—" in the table means this article has not confirmed an official document giving the date for that stage (not that the stage did not happen).

Vera Rubin deserves a closer look:

- The January 2026 CES announcement said "NVIDIA Rubin is in full production, and Rubin-based products will be available from partners the second half of 2026."
- The FY2026 10-K, filed in February 2026, said Rubin "is expected to commence production shipments in the second half of fiscal year 2027" (July 27, 2026 to January 31, 2027).
- The 10-Q filed on August 26, 2026 says "Vera Rubin, began production shipments in the third quarter of fiscal year 2027."
- An NVIDIA blog post on November 16, 2026 reports MLPerf results (an industry benchmark for AI performance) for Vera Rubin NVL72. Its MLPerf Endpoints (inference) result is noted as provisional and pending verification by MLCommons; the MLPerf Training v6.1 results have been published.
- Whether general availability (anyone can rent it) has begun at the major cloud providers is **unconfirmed as of November 17, 2026**.

"Full production" means the factory has started making the product. It does not mean customers have received it, or that anyone can use it.

## 13. Original analysis 1: how the revenue mix changed, using 10-K figures

From here on, this is my own analysis of 10-K figures.

### Revenue by market over time

In the notes to its 10-K, NVIDIA discloses revenue by market (Data Center, Gaming, Professional Visualization, Automotive, OEM & Other). The table shows selected years and the shares I calculated from them (in millions of dollars; shares are my calculation).

| FY | Period (calendar) | Data Center | Gaming | Total | DC share | Gaming share |
|---|---|---|---|---|---|---|
| FY2014 | 2013-01-28 to 2014-01-26 | 199 | 1,511 | 4,130 | 4.8% | 36.6% |
| FY2016 | 2015-01-26 to 2016-01-31 | 339 | 2,818 | 5,010 | 6.8% | 56.2% |
| FY2018 | 2017-01-30 to 2018-01-28 | 1,932 | 5,513 | 9,714 | 19.9% | 56.8% |
| FY2020 | 2019-01-28 to 2020-01-26 | 2,983 | 5,518 | 10,918 | 27.3% | 50.5% |
| FY2021 | 2020-01-27 to 2021-01-31 | 6,696 | 7,759 | 16,675 | 40.2% | 46.5% |
| FY2022 | 2021-02-01 to 2022-01-30 | 10,613 | 12,462 | 26,914 | 39.4% | 46.3% |
| FY2023 | 2022-01-31 to 2023-01-29 | 15,005 | 9,067 | 26,974 | 55.6% | 33.6% |
| FY2024 | 2023-01-30 to 2024-01-28 | 47,525 | 10,447 | 60,922 | 78.0% | 17.1% |
| FY2025 | 2024-01-29 to 2025-01-26 | 115,186 | 11,350 | 130,497 | 88.3% | 8.7% |
| FY2026 | 2025-01-27 to 2026-01-25 | 193,737 | 16,042 | 215,938 | 89.7% | 7.4% |

Sources: each year's 10-K (FY2014 and FY2016 from the FY2016 10-K, FY2018 from the FY2018 10-K, FY2020 from the FY2020 10-K, FY2021 from the FY2021 10-K, FY2022 and FY2023 from the FY2023 10-K, FY2024 and FY2025 from the FY2025 10-K, FY2026 from the FY2026 10-K). Totals include Professional Visualization, Automotive and OEM & Other. Where the same year appears in more than one 10-K, I cross-checked the filings and found no restatements within the range of this table.

### The code used for the calculation

I calculated the shares and growth rates with the Python script below. It first checks that the components add up to the 10-K total, to catch transcription errors.

- Language: Python 3.9.6 (standard library only, no extra dependencies)
- Environment: Apple M4 Pro (macOS), interpreter /usr/bin/python3 (Python 3.9.6). No GPU needed.
- Run: `python3 nvidia_revenue_mix.py`

The labels and messages printed by the script are in Japanese; the code is identical to the Japanese version of this article so that the output matches. 期間 = period, その他 = other markets combined, 倍 = times.

```python
"""NVIDIA の市場別売上（10-K の "revenue by specialized markets" 表）から構成比と成長率を計算する。

数値の単位は百万米ドル。出典は各年度の Form 10-K（EDGAR）。
FY は1月下旬に終わる会計年度（例: FY2026 = 2025-01-27〜2026-01-25）。
"""

# Year: (Data Center, Gaming, Professional Visualization, Automotive, OEM & Other/IP, total, fiscal-year period)
REVENUE = {
    "FY2014": (199, 1511, 789, 99, 1532, 4130, "2013-01-28〜2014-01-26"),
    "FY2016": (339, 2818, 750, 320, 783, 5010, "2015-01-26〜2016-01-31"),
    "FY2018": (1932, 5513, 934, 558, 777, 9714, "2017-01-30〜2018-01-28"),
    "FY2020": (2983, 5518, 1212, 700, 505, 10918, "2019-01-28〜2020-01-26"),
    "FY2021": (6696, 7759, 1053, 536, 631, 16675, "2020-01-27〜2021-01-31"),
    "FY2022": (10613, 12462, 2111, 566, 1162, 26914, "2021-02-01〜2022-01-30"),
    "FY2023": (15005, 9067, 1544, 903, 455, 26974, "2022-01-31〜2023-01-29"),
    "FY2024": (47525, 10447, 1553, 1091, 306, 60922, "2023-01-30〜2024-01-28"),
    "FY2025": (115186, 11350, 1878, 1694, 389, 130497, "2024-01-29〜2025-01-26"),
    "FY2026": (193737, 16042, 3191, 2349, 619, 215938, "2025-01-27〜2026-01-25"),
}


def main() -> None:
    # 1) Check that the components add up to the 10-K total (catches transcription errors)
    for fy, (*parts, total, _span) in REVENUE.items():
        assert sum(parts) == total, f"{fy}: 内訳の合計 {sum(parts)} != 合計 {total}"

    # 2) Shares of total revenue
    print(f"{'FY':<7}{'期間':<24}{'DC':>8}{'Gaming':>8}{'その他':>8}{'DC/Gaming':>11}")
    for fy, (dc, gaming, pv, auto, oem, total, span) in REVENUE.items():
        other = (pv + auto + oem) / total
        print(f"{fy:<7}{span:<24}{dc / total:>8.1%}{gaming / total:>8.1%}{other:>8.1%}{dc / gaming:>11.2f}")

    # 3) First fiscal year in which Data Center exceeded Gaming
    first = next(fy for fy, v in REVENUE.items() if v[0] > v[1])
    print(f"\nData Center が Gaming を初めて上回った年度（この表の範囲）: {first}")

    # 4) Compound annual growth rate (CAGR): (end / start) ** (1 / years) - 1
    def cagr(start: float, end: float, years: int) -> float:
        return (end / start) ** (1 / years) - 1

    dc14, g14, t14 = REVENUE["FY2014"][0], REVENUE["FY2014"][1], REVENUE["FY2014"][5]
    dc26, g26, t26 = REVENUE["FY2026"][0], REVENUE["FY2026"][1], REVENUE["FY2026"][5]
    print(f"FY2014→FY2026（12年）の CAGR: Data Center {cagr(dc14, dc26, 12):.1%}, "
          f"Gaming {cagr(g14, g26, 12):.1%}, 合計 {cagr(t14, t26, 12):.1%}")
    print(f"倍率: Data Center {dc26 / dc14:.0f}倍, Gaming {g26 / g14:.1f}倍, 合計 {t26 / t14:.1f}倍")


if __name__ == "__main__":
    main()
```

Output (copied verbatim from running it in this environment):

```text
FY     期間                            DC  Gaming     その他  DC/Gaming
FY2014 2013-01-28〜2014-01-26       4.8%   36.6%   58.6%       0.13
FY2016 2015-01-26〜2016-01-31       6.8%   56.2%   37.0%       0.12
FY2018 2017-01-30〜2018-01-28      19.9%   56.8%   23.4%       0.35
FY2020 2019-01-28〜2020-01-26      27.3%   50.5%   22.1%       0.54
FY2021 2020-01-27〜2021-01-31      40.2%   46.5%   13.3%       0.86
FY2022 2021-02-01〜2022-01-30      39.4%   46.3%   14.3%       0.85
FY2023 2022-01-31〜2023-01-29      55.6%   33.6%   10.8%       1.65
FY2024 2023-01-30〜2024-01-28      78.0%   17.1%    4.8%       4.55
FY2025 2024-01-29〜2025-01-26      88.3%    8.7%    3.0%      10.15
FY2026 2025-01-27〜2026-01-25      89.7%    7.4%    2.9%      12.08

Data Center が Gaming を初めて上回った年度（この表の範囲）: FY2023
FY2014→FY2026（12年）の CAGR: Data Center 77.4%, Gaming 21.8%, 合計 39.1%
倍率: Data Center 974倍, Gaming 10.6倍, 合計 52.3倍
```

The last three lines read: Data Center first exceeded Gaming in FY2023 (within this table); the FY2014→FY2026 (12-year) CAGR is 77.4% for Data Center, 21.8% for Gaming and 39.1% for the total; and the multiples are 974x for Data Center, 10.6x for Gaming and 52.3x for the total.

The compound annual growth rate (CAGR) uses the formula below, with start value $V_s$, end value $V_e$ and number of years $n$.

$$
\mathrm{CAGR} = \left(\frac{V_e}{V_s}\right)^{1/n} - 1
$$

For Data Center from FY2014 to FY2026 (12 years):

$$
\left(\frac{193{,}737}{199}\right)^{1/12} - 1 \approx 0.774
$$

### What the table shows

1. **Data Center first exceeded Gaming in FY2023** (January 2022 to January 2023). In FY2021 and FY2022, Data Center was about 0.85 times Gaming.
2. **Gaming did not shrink.** FY2026 Gaming revenue of 16.042 billion dollars is the largest in the table, 10.6 times FY2014. The FY2026 10-K says Gaming revenue rose 41% from the previous year. Its share fell to 7.4% because Data Center grew at a completely different pace: 974 times in 12 years.
3. **The inflection point was FY2024** (January 2023 to January 2024). Data Center revenue grew 3.17 times in one year ("up 217%" in the 10-K's words). The FY2024 10-K also gives the company's estimate that about 40% of that year's Data Center revenue was for AI inference (running trained models).

### Caveats when reading this table

- **Category names and contents have changed.** Through the FY2018 10-K the labels were "Datacenter" and "OEM & IP"; later they became "Data Center" and "OEM & Other." Starting with the FY2025 10-K, Data Center is broken out into Compute and Networking.
- **The large "other" share in FY2014 (58.6%)** likely reflects, in addition to OEM products, income from the patent cross-license with Intel (signed in January 2011, 1.5 billion dollars in total over six years). According to the FY2016 10-K, that license revenue was a flat 264 million dollars in each of FY2014 through FY2016, about 6.4% of FY2014 total revenue (my calculation). The 10-K does not state which line of the market table it falls in, so placing it in "OEM & IP" is my inference.
- **FY2022 includes cryptocurrency-mining demand.** According to the FY2022 10-K, the increase in OEM & Other came mainly from mining-specific products (CMP), whose revenue was 550 million dollars. The company also says some gaming GPUs were used for mining, but that the scale is "difficult … to estimate."
- **Networking revenue cannot be compared with Mellanox revenue.** Per the FY2026 10-K, growth in networking includes the NVLink compute fabric used in GB200 and GB300 systems. Comparing FY2026 networking revenue of 31.376 billion dollars with Mellanox's 2019 revenue of 1.3306 billion dollars and calling the difference "the payoff of the acquisition" would be wrong.
- **Customers are concentrated.** According to the FY2026 10-K, one direct customer accounted for 22% of total revenue and another for 14%. Their names are not disclosed.

## 14. Original analysis 2: how long each technical bet took to pay off

Next, I line up each major technical bet with "when the bet was made" and "when it first showed up clearly in 10-K figures." The lags are rough indications and do not prove causation.

| Bet | When (calendar) | Payoff visible in 10-K and similar figures | Approximate lag |
|---|---|---|---|
| Switching to the industry standard (Direct3D) | Stopped selling NV1 in 1996; RIVA 128 announced April 1997 (volume shipments from August) | 1997 revenue of 29.07 million dollars (about 7.4x the previous year) | 1–2 years |
| Integrating T&L (GeForce 256) | Shipping from August 1999 | FY2001 revenue of 735.26 million dollars (about 2x FY2000) | 1–2 years |
| CUDA (general-purpose computing) | Announced November 2006, official release June 2007 | Data Center revenue first exceeded 1 billion dollars in FY2018 (1.932 billion) | About 11 years |
| Ready for deep learning (AlexNet was outside research) | 2012 | Data Center overtook Gaming in FY2023 | About 10 years |
| HBM2, NVLink, Tensor Cores (P100, V100) | 2016–2017 | Data Center +145% in FY2017 and +133% in FY2018 (the 10-K says growth was "led by" Volta) | 0–1 year |
| Mellanox acquisition | Announced 2019, completed April 2020 | Mellanox was 10% of company revenue in FY2021 | Same fiscal year (acquired revenue is added directly) |
| Hopper | Announced March 2022, shipping during FY2023 | Data Center 3.17x in FY2024; the FY2025 10-K says "demand for our Hopper architecture drove our significant growth" | 1–2 years |
| Blackwell | Announced March 2024, shipping from Q4 FY2025 | The FY2026 10-K attributes 59% growth in Data Center computing revenue to "demand for our Blackwell computing platform" | About 1 year |
| Vera Rubin | Announced January 2026, production shipments from Q3 FY2027 | Not yet determined (as of November 17, 2026) | — |

### What the table tells us

**The lag has shortened from about ten years to about one.** But two different kinds of bets are mixed together here.

- **Bets that build the foundation** (CUDA, building the GPU into a general-purpose computer) took around ten years to pay off. When CUDA was announced in 2006, there was no deep-learning demand yet.
- **Bets that add generations on top of the foundation** (each generation from Volta on) bring products into a market where demand already exists, so they show up in revenue within one to two years of shipping.

Put differently, the fast payoffs from Hopper on were only possible because of a foundation, CUDA, that took ten years to build.

### Limits of this analysis

- **Survivorship bias**: the table lists bets that turned out well. Failures such as NV1 and NV2, and the Arm acquisition (a 1.35 billion dollar charge writing off the prepayment), are part of the same company's history.
- **Demand-side factors are not separated out**: the FY2024 surge owes a great deal to an external factor, demand for generative AI. Technical bets alone cannot explain revenue.
- **The timing is coarse**: 10-K disclosures are mainly annual, so causation cannot be traced month by month.

### Three lessons (my interpretation)

From this history I draw three lessons. I offer them as interpretation, not fact.

1. **Ecosystems beat technology**: NV1 made technical sense, but developers and APIs did not follow, and it failed. CUDA did the opposite: the code developers accumulated became a barrier to entry in itself.
2. **Foundation bets start before the demand arrives**: CUDA was announced six years before deep-learning demand became visible. Being able to keep investing in something that took ten years to pay off is what created the later gap.
3. **The unit of sale moved up from "chip" to "data center"**: memory (HBM), connections inside the server (NVLink) and connections between servers (Mellanox) — NVIDIA pulled one bottleneck after another inside its own products.

## 15. Original analysis 3: from "GPU company" to "AI infrastructure company" in the company's own words

How a company describes itself also reflects change. Here are the opening sentences of the business description (Item 1) of selected 10-Ks (original text, excerpted).

| FY | Opening self-description in the 10-K |
|---|---|
| FY2000 | "We design, develop and market a "top-to-bottom" family of award-winning 3D graphics processors, graphics processing units…" |
| FY2008 | "the worldwide leader in visual computing technologies and the inventor of the graphic processing unit" |
| FY2016 | "NVIDIA is the world leader in visual computing. … power breakthroughs in new areas like artificial intelligence, virtual reality and autonomous cars." |
| FY2017 | "…We have extended our emphasis in recent years to the revolutionary field of artificial intelligence" |
| FY2020 | "NVIDIA pioneered accelerated computing…" |
| FY2024 | "NVIDIA is now a full-stack computing infrastructure company with data-center-scale offerings…" |
| FY2026 | "NVIDIA is now a data center scale AI infrastructure company reshaping all industries." |

Through FY2016 the subject was "visual computing." The FY2016 10-K already mentions artificial intelligence as an example application ("power breakthroughs in new areas like artificial intelligence"), but FY2017 is the first year it calls AI a focus of the business ("We have extended our emphasis"). In FY2020 the subject becomes "accelerated computing" (speeding up computation with GPUs and similar hardware), and in FY2024 "infrastructure." By FY2026, AI is part of the definition of the company itself: "AI infrastructure."

Set against the revenue table in Section 13, **the self-description changed slightly ahead of the revenue mix**. When the FY2016 10-K mentioned artificial intelligence as an application, Data Center was 6.8% of revenue; even in FY2017, when it first called AI a focus, Data Center was still only 12% (FY2017 Data Center revenue of 830 million dollars divided by total revenue of 6.91 billion dollars, my calculation). The company had started describing itself in new terms before the revenue followed.

## 16. Limits and counterarguments: how to read this history

Finally, here are counterarguments to, and limits of, the view in this article.

- **Company documents are written in the company's favor**: phrases such as "world's first GPU" and "first unified shader" are claims based on NVIDIA's definitions. As Sections 6 and 7 showed, competitors and other products, such as ATI's Radeon 9700 and the Xbox 360, got there first in some respects.
- **The role of chance**: AlexNet came out of research outside NVIDIA, and the fact that deep learning suited GPUs so well involved some luck. Writing that "NVIDIA foresaw the AI era" tends to be a story told with hindsight.
- **The early history relies partly on recollection**: NV1's design philosophy and the Sega relationship rest on a participant's blog (Rosenthal) and an industry historian (Peddie). The S-1 does not name Sega.
- **Concentration risk**: about 90% of revenue comes from Data Center, and two customers together account for 36% (FY2026). A company that has already lived through sudden swings in gaming and cryptocurrency demand now carries a different kind of concentration.
- **Change in leadership**: in November 2026, the co-founder who had been CEO for 33 years resigned. Many of the technical bets in this article were sustained under the same leadership. Future direction will have to wait for the appointment of a permanent CEO.

## Summary

- NVIDIA was founded by three engineers on April 5, 1993. Its first product, NV1, rendered curved surfaces instead of triangles, an approach that made technical sense, but it failed when the industry standardized on Direct3D and OpenGL.
- RIVA 128 (1997) survived by riding the standard, and GeForce 256 (1999) became the "world's first GPU" by NVIDIA's own definition. Programmable shaders (2001), then unified shaders and CUDA (2006), turned the GPU into a general-purpose parallel computer.
- AlexNet (2012) was trained on two gaming GTX 580s. NVIDIA's investment in general-purpose computing absorbed a deep-learning turning point that happened outside the company. HBM2 and NVLink (P100), Tensor Cores (V100) and the Mellanox acquisition then widened the unit of business from the GPU to the whole data center.
- In 10-K revenue by market, Data Center rose from 4.8% of revenue in FY2014 to 89.7% in FY2026. Gaming grew 10.6x as well: gaming did not shrink, Data Center grew by orders of magnitude.
- It took about 11 years from the CUDA announcement until Data Center revenue passed 1 billion dollars, while generations from Hopper on show up in revenue within one to two years. That picture has to be read together with failures such as NV1 and the abandoned Arm acquisition.
- NVIDIA's August 2026 10-Q reported that Vera Rubin production shipments had begun, but general availability in the cloud is unconfirmed as of November 17, 2026.

For the technical reasons GPUs are so strong at AI computation, see [Why NVIDIA GPUs are so strong at AI](/blog/nvidia-gpu-ai-tensor-core-fp8-hbm-nvlink-explained) in the same series, which works through the four layers: Tensor Cores, low-precision arithmetic, HBM and NVLink.

For the same events seen through the founder's decisions, see [Jensen Huang's life and technology leadership](/blog/jensen-huang-biography-technology-leadership); for competition and the future from here, see [NVIDIA's moat is more than a CUDA monopoly](/blog/nvidia-competitive-advantage-cuda-ecosystem-future-scenarios).
