Gunjo · Business Intelligence for the AI Era
← Sticker Wall JOURNEY · DETAIL

Jensen Huang and NVIDIA: How the CUDA Ecosystem Bet Created an AI Compute Monopoly Empire

Founded: Jensen Huang, Chris Malachowsky, Curtis Priem · NVIDIA Corporation

JOURNEY

Key Fields

FIELD STAMPS
IndustryConsumer Electronics / Semiconductors
RegionUS
ScaleGiant
ChannelOther

Origin

In 1993, when Jensen Huang decided to start a company with two engineer partners at a Denny's diner, 3D PC gaming was just emerging. Huang had previously worked as a chip engineer at AMD and later moved to LSI Logic as a chip design director, witnessing firsthand the explosive demand for graphics processing without any standard solution. The three determined that PC gaming would become mainstream and wanted to build a chip capable of handling 2D and 3D graphics simultaneously, yet their starting capital was only $40,000. To make matters trickier, there were absolutely no industry standards for 3D acceleration technology at the time, and no one knew how it should be done.

Milestones

1993
Inception and Startup Failure
In 1993, at age 30, Jensen Huang resigned from LSI Logic and founded NVIDIA with Malachowsky and Priem in an empty office in Santa Clara, California. Their $40,000 starting capital quickly ran low. In 1994, a $20 million Series A funding round from Sequoia Capital and Sutter Hill kept them afloat, with plans to break into the Windows market using DirectDraw graphics acceleration. However, chip design started almost from scratch, and the team was completely in the dark regarding the 3D architecture direction.
1995
NV1 Chip Release Failure
The NV1 adopted non-mainstream quadratic surface rendering technology, which was incompatible with the emerging triangle/polygon 3D standard. Although Sega invested about $7 million to order the NV1/NV2 chips for its Saturn game console, the project was ultimately canceled because the NV2 solution conflicted with Sega's new console direction. NVIDIA came close to bankruptcy with cash flow capable of lasting only a few months; Huang later recalled, 'We were just weeks away from going out of business.'
1997
RIVA 128 Release Turning Point
In 1997, Jensen Huang poached David Kirk from SGI as Chief Scientist, abandoning quadratic surface technology to pivot fully to triangle polygon rendering. The RIVA 128 shipped 1 million units within four months, becoming NVIDIA's first successful product. When it went public in January 1999, annual revenue was approximately $158 million, and its market cap on listing day was about $470 million, validating Huang's razor-edge strategy of product iteration every six months for the first time.
1999
GeForce 256 Release PMF
In 1999, the GeForce 256 was defined by NVIDIA as the world's first GPU (Graphics Processing Unit), integrating 23 million transistors. For the first time, geometric transformation and lighting calculations were offloaded from the CPU to the graphics card. Priced from $199, it propelled NVIDIA's fiscal 2000 revenue past the $1.035 billion scale, defeating 3dfx and Matrox to become the leader in discrete graphics cards, which allowed Jensen Huang to establish the GPU category mindset.
2006
CUDA Release Inflection Point
In 2006, CUDA transformed the GPU from a dedicated graphics chip into a general-purpose parallel computing platform. Jensen Huang bet approximately $1 billion in R&D investment and internal incentives to make the entire company jump into this arena. However, for the first five years, there was virtually no return. Wall Street repeatedly questioned CUDA as a money-burning black hole. Around 2010, NVIDIA's market cap was chronically undervalued, yet Huang stubbornly refused to cut the CUDA team, requiring all GPUs to be compatible with the CUDA interface.
2012
AlexNet Deep Learning Breakthrough Turning Point
In 2012, Alex Krizhevsky used two NVIDIA GTX 580 graphics cards to train AlexNet and win the ImageNet competition, reducing error rates by over 10 percentage points compared to traditional methods. Deep learning was proven viable for the first time. Sensing the signal, Jensen Huang rapidly reorganized the company into an AI-first strategy, donating the first DGX-1 supercomputer to OpenAI in 2016 where Huang personally signed the chassis, cementing NVIDIA's binding relationship with AI labs.
2020
Data Center Revenue First Surpasses Gaming Growth
In the fourth quarter of fiscal 2020, data center revenue surpassed the gaming business for the first time, reaching a quarterly scale of approximately $1.46 billion. In September 2020, Jensen Huang announced a planned $40 billion acquisition of ARM, which ultimately failed but showcased his BCG chip ambitions, following the 2019 acquisition of Mellanox for $6.9 billion to strengthen data center networking capabilities. After AI training demand exploded, NVIDIA's H100 experienced historic stock shortages and premium pricing phenomena.
2025
Market Cap Exceeds $3 Trillion Growth
In June 2024, NVIDIA's market cap briefly surpassed Apple and Microsoft to become number one globally, intraday breaking $3.3 trillion. Full-year fiscal 2025 revenue reached $130.497 billion, more than doubling year-over-year, with data center revenue accounting for roughly $1152 billion or about 88% of total revenue. The Blackwell architecture supported inference demand upon mass production in 2025. In 2026, Jensen Huang publicly stated that the inference inflection point had arrived, expecting AI chips to bring in a trillion dollars in revenue by 2027.

Turning Points

  • After the NV1 failure in 1995, Jensen Huang negotiated with Sega to preserve the NV2 contract to no avail, forcing him to abandon quadratic surface technology and pivot to the triangle/polygon route; this was the closest NVIDIA ever came to death.
  • The release of CUDA in 2006 meant NVIDIA cast its general-purpose computing vision onto the market, judged by Wall Street for consecutive years as a money-burning black hole while Huang stubbornly maintained internal resource allocation without cutting the CUDA team.
  • In 2012, AlexNet won the competition trained on two NVIDIA graphics cards, and Jensen Huang rapidly reorganized the entire company into an AI-first strategy, deciding to bind itself tightly to the AI compute track.
  • The $6.9 billion acquisition of Mellanox in 2019 filled in data center networking capabilities, upgrading NVIDIA from a single-card supplier to a full-stack AI cluster supplier.

Failures & Pitfalls

  • Because the NV1 chip adopted quadratic surface rendering technology that was incompatible with industry standards, the product became unsellable upon launch, pushing NVIDIA's cash flow close to collapse.
  • The NV2 custom chip project for the Sega game console was ultimately canceled by Sega, worsening NVIDIA's revenue blow and pushing it to the brink of bankruptcy.
  • The roughly $40 billion acquisition deal for ARM in 2020 was ultimately blocked by antitrust reviews worldwide, announced as abandoned in February 2022, with all due diligence costs and time windows completely sunk.
  • After CUDA was released, there was almost no commercial return for the first five years, and around 2010 Wall Street repeatedly and publicly questioned CUDA as a management-deluded black hole project.

关键成功要素

  • Jensen Huang insisted that all GPUs must be compatible with the CUDA interface, refusing to cut general-purpose computing R&D for short-term profits—this is the source of NVIDIA's ecosystem moat.
  • Forcing out competitors such as 3dfx and Matrox through a razor-edge product iteration pace of once every six months, establishing market dominance in discrete graphics cards.
  • Acquiring Mellanox in 2019 to complete data center networking capabilities, upgrading NVIDIA from a single-card supplier to a full-stack AI cluster supplier.
  • In 2024, data center revenue accounted for about 88% of the company's total revenue, completing NVIDIA's thorough transformation from a gaming company into an AI compute infrastructure company.
  • Fiscal 2025 revenue reached $130.497 billion, more than doubling year-over-year, with the data center business at roughly $115.2 billion validating the commercial explosion of AI training and inference demand.

Lessons

  • If a company's technical standard bets run counter to mainstream industry standards, even the most brilliant product will be directly eliminated by the market; NV1's quadratic surface technology is a bloody lesson.
  • Infrastructure investments that look like money burners may only yield returns a decade later. The delayed-return logic of CUDA demonstrates that foundational ecosystem investments must be supported by extremely strong founder conviction.
  • External critical events (the AlexNet championship) determine a company's destiny, but Huang's persistence in the general-purpose computing path before the event occurred is what allowed him to capture that dividend.
  • Large-scale M&A carries extremely high uncertainty risk; the failure of the ARM acquisition illustrates that in the face of geopolitics and antitrust pressure, even $40-billion-scale acquisitions cannot be completed.
  • The transitions from gaming to data centers to AI infrastructure are each dangerous inflection points. Huang's choice to enter early across three bets shows that critical strategic windows take precedence over technology itself.

Core Data

  • Fiscal 2025 Revenue：$130.497 billion (public data basis, independent review not verified)
  • Fiscal 2024 Revenue：$609.22 billion (public data basis, independent review not verified)
  • Data Center Fiscal 2025 Revenue：Approximately $115.2 billion (public data basis, independent review not verified)
  • Peak Market Capitalization：Approximately $3.33 trillion (public data basis, independent review not verified)
  • Gross Margin：Approximately 73% (public data basis, independent review not verified)
  • Team Size：Approximately 29,600 people (public data basis, independent review not verified)

Competitors / Peers

AMD launched the Instinct MI300X in 2024 attempting to capture the AI training and inference market, but its software ecosystem ROCm still has a significant gap with CUDA, resulting in limited market share; Intel launched the Gaudi series of AI accelerators but received lukewarm market response; Google's self-developed TPUs support Gemini training internally but have limited external sales; Huawei's Ascend 910B absorbs partial substitution demand in the Chinese market but is constrained by advanced process supply; startups like Cerebras and Groq follow a wafer-scale engine route with limited commercialization; the core threat stems from major clients' self-developed chip trends, such as Google TPUs and Amazon Trainium.

Sources

Related

Same industry / scale
Consumer Electronics / SemiconductorsUS

Qualcomm: From Lab-Born CDMA to Three Decades of Global Mobile Dominance via Patent Royalties and Chips

Consumer Electronics / SemiconductorsUS

Fender: How the Telecaster and Stratocaster Defined Rock History

Consumer Electronics / SemiconductorsUS

NVIDIA: From Gaming GPU Maker to AI Computing Hegemon, Achieving Market Cap Supremacy Through the CUDA Ecosystem and Data Center GPUs

Consumer Electronics / SemiconductorsUS

AMD: The Chip Challenger that Rallied from the Brink of Bankruptcy to Counter Intel with the Zen Architecture

⚠️ This entry draws on self-reported accounts and public reporting, not independently verified, and is not investment, startup, or legal advice; revenue figures are case-specific and not representative of typical results.