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

Origin

In 1993, Jensen Huang and two co-founders left LSI Logic and Sun Microsystems to found NVIDIA, betting on the then-nonexistent PC graphics acceleration market. The company's first chip, the NV1, took the wrong technical path and nearly burned through its initial capital, surviving only thanks to the RIVA 128 in 1997. In 1999, it invented the GPU concept and launched the GeForce 256, becoming one of the two gaming graphics card giants after acquiring 3dfx's assets in 2000. In 2006, Jensen Huang pushed through widespread opposition to launch CUDA, transforming GPUs from fixed-function graphics pipelines into general-purpose parallel computing. While Wall Street viewed this as a bottomless money pit at the time, it laid the groundwork for the deep learning revolution triggered by AlexNet in 2012.

Milestones

Turning Points

• The 1997 launch of the RIVA 128 sold 1 million units in four months, pulling the company back from the brink of bankruptcy.
• In 2006, Jensen Huang forced through CUDA despite opposition from across the company, causing stock performance to lag the broader market for the next five years.
• In 2012, AlexNet won the ImageNet championship using two gaming graphics cards, making NVIDIA the accidental de facto hardware standard for deep learning.
• In 2016, the company personally delivered the first DGX-1 to OpenAI, shifting from selling graphics cards to gamers to selling computing power to AI labs.
• In 2023, ChatGPT triggered a global GPU buying frenzy, with the data center business surpassing its cumulative scale of the previous 20 years within 9 months.

Failures & Pitfalls

• In 1995, the NV1 was largely ignored due to incompatibility with Direct3D, leaving cash reserves sufficient for only one month.
• In 2008, the financial crisis combined with heavy investments in CUDA dragged down profit margins, causing the stock price to plummet from $37 to $6.
• In 2009, the Tegra series bet on mobile device processors, only to be completely crushed by custom chips from Qualcomm and Apple, leading to a virtual exit from the mobile market after 2015.
• In 2022, an attempted $66 billion acquisition of Arm was jointly blocked by the US Federal Trade Commission and UK regulators, resulting in the collapse of the deal.

Core Data

• 市值：Over $5 trillion
• 2025财年营收：$130.5 billion
• 2025财年数据中心营收：$115.2 billion
• 毛利率：73%
• Blackwell架构Q4营收：Approx. $11.0 billion
• AI芯片市占率：87.4%
• 创始人启动资金：$40,000
• 1997年RIVA128出货量：Over 1 million units
• 2016年数据中心营收：$830 million

Competitors / Peers

In the AI training accelerator market, NVIDIA's primary competition comes from AMD's Instinct MI series, Google's TPUs, and custom chips from Amazon such as Trainium and Inferentia. AMD gradually increased shipments of the MI300X between 2024 and 2025, attempting to break into NVIDIA's customer base with cost-performance advantages, though a massive software ecosystem gap remains. Google's TPUs account for a significant share of internal training workloads but are not sold publicly on the market. Cloud provider custom chips primarily target inference workloads and cannot shake NVIDIA's dominance on the training side in the short term. Intel's Gaudi series has largely retreated to the sidelines. Additionally, startups like Cerebras, Groq, and SambaNova maintain a presence in specific inference and sparse training scenarios, but their shipment volumes remain far behind NVIDIA.