NVIDIA: From Gaming Chip to AI Backbone — The Fastest Value Creation in History

NVIDIA headquarters, Santa Clara, California — home of the company that became the backbone of the AI revolution

There is no precedent in the history of publicly traded companies for what NVIDIA has done in the last three years.

In January 2023, NVIDIA was a $360 billion semiconductor company that made graphics chips for gamers and had an interesting but unproven bet on AI. By January 2025, it was a $3.4 trillion company — the second most valuable on Earth, briefly surpassing Apple — that had generated more market capitalization in 24 months than any company in history. Revenue grew from $27 billion in fiscal 2023 to $130 billion in fiscal 2025. Net income grew from $4.4 billion to $72.9 billion. Gross margins expanded from 57% to 75%.

To put that in context: NVIDIA's FY2025 net income of $72.9 billion is larger than Apple's net income, Microsoft's net income, or Google's net income in FY2024. NVIDIA achieved this with one-fifth the headcount of Microsoft and roughly one-tenth the revenue history.

This is the story of how a company that Jensen Huang founded in a Denny's restaurant in 1993 became the pickaxe seller of the AI gold rush — and why the moat it has built may be the most defensible technical monopoly in technology since Microsoft dominated PC operating systems in the 1990s.

The Gaming Foundation Nobody Took Seriously

NVIDIA's origin story is not AI. It's Quake.

In the early 1990s, 3D gaming was the most demanding computational workload in consumer technology. Rendering three-dimensional environments in real-time required specialized hardware — a graphics processing unit — that could perform millions of floating-point calculations per second in parallel. NVIDIA's first product, the NV1, launched in 1995, was mediocre. The NV3 — the RIVA 128 — was good enough to survive. The GeForce 256, launched in 1999, was the product that mattered: it was the first chip to carry the GeForce brand and the first to be marketed as a GPU rather than a graphics accelerator.

What made gaming GPUs technically interesting was the same thing that would make them transformative for AI thirty years later: parallelism. A CPU is designed to execute a single complex task very quickly. A GPU is designed to execute thousands of simpler tasks simultaneously. Rendering a 3D frame requires calculating the color of millions of pixels in parallel — perfect for a massively parallel architecture. That architectural requirement shaped NVIDIA's entire chip design philosophy, and that philosophy turned out to be exactly what deep learning needed.

By FY2022, NVIDIA's Gaming segment generated $12.46 billion in revenue — its largest segment at the time, accounting for 46% of total revenue. GeForce was the dominant discrete GPU brand for PC gaming with roughly 80% market share. AMD had caught up in performance but never achieved parity in software ecosystem. Intel had tried and largely failed to enter discrete GPUs competitively. The gaming moat was real, but the AI moat was the one that mattered.

CUDA — The Technical Moat Built Over 15 Years

In 2006, NVIDIA did something that looked expensive and unnecessary: it launched CUDA.

CUDA — Compute Unified Device Architecture — was a programming framework that let developers write general-purpose programs that could run on NVIDIA GPUs. The problem CUDA solved was that GPUs, despite their raw parallel processing power, were nearly impossible to program for anything other than graphics. CUDA provided the APIs, the compiler toolchain, and the libraries that let researchers and engineers express arbitrary parallel computations in a form that NVIDIA hardware could execute.

The research community adopted CUDA slowly at first. High-performance computing was the early use case — physics simulations, molecular dynamics, fluid dynamics. Then, in 2012, something happened that changed the trajectory of both deep learning and NVIDIA: AlexNet.

Alex Krizhevsky's convolutional neural network, trained on NVIDIA GPUs using CUDA, won the ImageNet image classification competition with error rates so far below the competition that the machine learning research community was forced to reconsider its assumptions about what deep learning could do. AlexNet demonstrated that GPU-accelerated training made deep neural networks practical at a scale that was impossible with CPU training. Every major AI lab in the world went to the same place to buy the hardware they needed: NVIDIA.

The critical insight about CUDA is not that it was a better API. It's that NVIDIA invested fifteen years building the ecosystem around it. cuDNN — the CUDA Deep Neural Network library — optimizes fundamental operations that every neural network uses: convolutions, matrix multiplications, activation functions. The deep learning community built their frameworks — TensorFlow, PyTorch, JAX — to use cuDNN. Those frameworks now power essentially all serious AI research and production deployment globally. Switching away from CUDA doesn't just mean switching chips — it means porting your entire framework stack, rewriting your optimization code, and potentially losing years of accumulated performance tuning.

AMD has competitive GPU hardware (the MI300X competes with NVIDIA's H100 on paper) and has invested heavily in its ROCm software stack. Google has custom TPUs. Intel has Gaudi accelerators. None of them have CUDA's ecosystem depth. The software moat that NVIDIA built between 2006 and 2020 is the reason that when the AI supercycle arrived, only one company could supply it at scale.

The Data Center Pivot

NVIDIA began breaking out its Data Center revenue segment in FY2018. That year it was $1.93 billion — meaningful but dwarfed by Gaming's $5.51 billion. The first signal of what was coming arrived in FY2021: Data Center overtook Gaming for the first time, at $6.70 billion versus $5.52 billion.

The inflection point was ChatGPT.

OpenAI launched ChatGPT on November 30, 2022. Within five days it had 1 million users. Within two months it had 100 million — the fastest consumer application adoption in history. More importantly, it demonstrated to every major corporation and government in the world that large language models were commercially viable, and that training and running them required enormous amounts of GPU compute.

Every major technology company launched an AI arms race simultaneously. Microsoft committed $10 billion to OpenAI and pledged to deploy AI across its entire product portfolio. Google scrambled to respond with Gemini. Meta announced $65 billion in AI capital expenditure for 2025 alone. Amazon, Apple, Oracle, and every major cloud provider began ordering NVIDIA H100s — and later H200s and the Blackwell architecture — at quantities that exceeded NVIDIA's manufacturing capacity.

NVIDIA's Data Center revenue in FY2024 was $47.5 billion — up 217% from $15.0 billion in FY2023. In FY2025, Data Center revenue hit $115.2 billion, accounting for 88% of total company revenue. Gaming, once NVIDIA's core business, contributed $11.4 billion — nearly irrelevant by comparison.

The price dynamic was extraordinary. The NVIDIA H100 — the GPU at the center of the AI training boom — was selling for $25,000-$40,000 per card on secondary markets at peak demand in 2023. Entire companies were founded on the premise of securing H100 allocation. Hyperscalers paid multi-billion-dollar deposits to TSMC and NVIDIA to secure future supply. The H100 was simultaneously the most valuable and most scarce industrial input in the global economy.

The Blackwell Architecture — Building the Next Moat

In March 2024, NVIDIA unveiled Blackwell — its next GPU architecture, succeeding Hopper (which powered the H100). The naming convention itself signals Jensen Huang's ambition: he names architectures after pioneers in computing and mathematics. Turing. Volta. Ampere. Hopper. Blackwell (after David Harold Blackwell, the statistician).

The Blackwell B200 GPU delivers 20 petaflops of FP4 AI performance — 5x the performance of the H100's 4 petaflops of FP8. But the more important innovation is the NVLink switch fabric that connects Blackwell GPUs in massive clusters. A single Blackwell rack-scale system — the NVL72 — connects 72 GPUs over NVLink, enabling the entire cluster to behave as a single logical accelerator for training the largest frontier models.

This matters because the trend in frontier AI is toward larger models trained on more compute. GPT-4 reportedly required approximately 10,000 A100s running for 90-100 days. The next generation models require substantially more. Training at that scale requires not just fast chips but fast interconnects — and NVIDIA's NVLink fabric is proprietary, fast, and deeply integrated with its software stack.

The upgrade cycle is already underway. NVIDIA reported that Blackwell generated $11 billion in revenue in its first quarter of production — the fastest product ramp in the company's history. NVIDIA's own guidance and analyst projections suggest Blackwell revenue will substantially exceed Hopper's FY2024 trajectory, potentially hitting $200 billion in FY2026 if supply constraints ease.

Jensen Huang — The Founder-CEO Factor

Not every company can execute on a thirty-year technical bet. The reason NVIDIA could is Jensen Huang.

Huang co-founded NVIDIA in 1993 and has led it continuously since. He is 61 years old and shows no sign of succession planning — nor should he. The compounding value of a founder-CEO who understands the technical roadmap as deeply as the capital allocation decisions is rare in large-cap technology. Jeff Bezos built AWS because he understood that Amazon's logistics infrastructure could be generalized into a cloud service. Jensen Huang built CUDA because he understood that parallel computing was a general-purpose resource, not just a graphics rendering tool.

His bet on CUDA in 2006, when it had no near-term revenue justification, reflects a pattern: invest heavily in platform infrastructure before the market demands it, then own the market when demand arrives. The same logic applies to NVIDIA's DGX systems — purpose-built AI supercomputers that NVIDIA began selling to research labs in 2016 — which established deep relationships with the AI research community before commercial demand existed. Those relationships became purchasing decisions when OpenAI and every major AI lab needed to scale.

Huang's compensation is notable: he takes a relatively modest salary by CEO standards but owns approximately 3.5% of NVIDIA's shares outstanding, worth roughly $120 billion at peak valuations. His incentives are perfectly aligned with long-term value creation, not quarterly earnings management.

Ten-Year Financial Performance (FY2020–FY2025)

The numbers tell a story that requires no narrative embellishment:

Note: EPS shown on post-split (10:1, June 2024) basis. FY2023 net income and EPS were impacted by $1.35B Arm acquisition termination fee and $1.22B inventory charges. Gross margin expansion from FY2023 trough (56.9%) to FY2025 (74.6%) represents 17.7 percentage points of structural improvement driven by Data Center mix shift.

The gross margin story is the most important. NVIDIA's Data Center GPUs carry gross margins substantially above the company average — industry analysts estimate H100/Blackwell margins at 70-80%+. As Data Center grows from 46% of revenue in FY2022 to 88% in FY2025, the overall margin structure improves structurally. This is not cost-cutting — it is product mix driving permanent margin expansion.

Free cash flow conversion is extraordinary: NVIDIA converted 47% of revenue to free cash flow in FY2025. For comparison, Microsoft converts approximately 30% and Apple approximately 25%. The FCF generation reflects a capital-light model — NVIDIA designs chips, TSMC manufactures them, and NVIDIA captures the intellectual property margin without owning fabs.

The Capital Return Machine

The NVIDIA campus sign — a company Jensen Huang co-founded in a Denny's restaurant in 1993, now worth over $3 trillion

NVIDIA completed a 10-for-1 stock split on June 10, 2024 — the company's sixth split since its 1999 IPO. Pre-split, the stock had traded above $1,200.

On dividends: NVIDIA pays a quarterly dividend of $0.01 per share (post-split) — essentially symbolic at current valuations. This is not an income stock and was never intended to be.

On buybacks: NVIDIA returned $15.4 billion to shareholders through buybacks in FY2025 and announced a $50 billion buyback authorization in August 2024. At NVIDIA's current free cash flow generation rate ($60+ billion annually), the company could theoretically repurchase its entire float in approximately five years. The buyback pace will likely accelerate as capital expenditure needs stabilize.

The total shareholder return is the headline: NVIDIA's stock returned approximately 2,400% over the five years ending FY2025 — a 88% compound annual growth rate. To put that in context: $10,000 invested in NVIDIA in January 2020 was worth approximately $250,000 in January 2025. No other large-cap stock in the S&P 500 came close.

The Risks — What Could Break the Machine

Export Controls and Geopolitical Risk

NVIDIA's most immediate structural risk is the US government. The Biden administration's export controls on advanced AI chips — initially targeting the A100 and H100, then tightening with additional restrictions in October 2023 and again in 2024 — prohibited NVIDIA from selling its most advanced chips to China.

China represented approximately 17% of NVIDIA's data center revenue before controls tightened. NVIDIA developed a China-specific variant (the H800, L40S) designed to comply with export limits, but successive rule tightening effectively blocked those as well. The lost China revenue represents a real headwind — but the more significant risk is escalation. If US-China technology decoupling accelerates and extends to TSMC production for NVIDIA chips, the entire semiconductor supply chain faces stress.

The TSMC Concentration Risk

NVIDIA does not own fabs. Every advanced NVIDIA GPU is manufactured by TSMC in Taiwan on its 4nm and 3nm process nodes. Taiwan faces ongoing geopolitical risk from Chinese territorial claims. A Taiwan Strait conflict — even a non-kinetic escalation — would disrupt NVIDIA's entire production capacity.

TSMC is building fabs in Arizona, Japan, and Germany. Intel and Samsung are building competing advanced-node fabs. But none of these facilities will reach TSMC Taiwan's scale for years, and none currently produce the leading-edge nodes that NVIDIA's Blackwell architecture requires.

Demand Concentration and CapEx Cycle Risk

In FY2025, four hyperscalers — Microsoft, Google, Amazon, Meta — likely accounted for more than 50% of NVIDIA's Data Center revenue. This is extraordinary customer concentration for a $130 billion revenue business.

Hyperscaler capital expenditure follows cycles. The current AI infrastructure buildout is the steepest CapEx cycle in technology history — Microsoft committed $80 billion in AI infrastructure for FY2025 alone, Meta committed $65 billion. If AI model returns disappoint — if enterprise AI adoption is slower than expected, if the killer application doesn't materialize at consumer scale — hyperscaler CapEx could slow sharply. NVIDIA's revenue would follow.

The Open Ecosystem Challenge

Every major technology company is investing in alternatives to NVIDIA. Google's TPU v5 is competitive for its own training workloads. AMD's MI300X offers compelling price-performance. Intel's Gaudi 3 targets inference workloads. Qualcomm and Apple are investing in on-device AI acceleration. Most importantly, the major AI labs — particularly those backed by Google and Amazon — are developing custom silicon that reduces NVIDIA dependence.

The CUDA moat is deep but not impenetrable. ROCm (AMD's equivalent) has improved significantly. JAX and TensorFlow have TPU support. The ecosystem is diversifying. NVIDIA's pricing power — which supports those 75% gross margins — depends on its software moat remaining high enough that alternatives are meaningfully worse, not just slightly worse.

Verdict

NVIDIA is the rarest thing in investing: a business where the competitive moat is getting deeper at the same time the addressable market is getting larger.

The CUDA software ecosystem took fifteen years to build and would take a decade to replicate. The NVLink interconnect architecture creates switching costs at the cluster level that go beyond individual chip decisions. The relationship between NVIDIA's hardware, software stack, and the AI research community is symbiotic — every new architecture that ships, every new optimization in cuDNN, every new model trained on NVIDIA infrastructure deepens the ecosystem advantage.

The financial results are simply the output of that moat working. Revenue of $130 billion with 75% gross margins and 47% free cash flow conversion is not an accident — it is what happens when a monopoly supplier faces effectively unlimited demand from the most capital-intensive infrastructure buildout in the history of computing.

The risks are real. Geopolitical risk is structural and not priceable. Customer concentration creates cyclical exposure. Competition is not standing still. The current valuation — at its peak, roughly 30x trailing revenue — prices in sustained hyperscaler demand, continued margin expansion, and successful Blackwell ramp.

But the core thesis is simple: training and running frontier AI models requires massive GPU compute, NVIDIA's CUDA ecosystem makes its GPUs the default choice, and frontier AI demand is not a trend but a permanent structural shift in how computing resources are allocated globally. Every dollar spent on AI infrastructure in the next decade flows, in some meaningful fraction, through NVIDIA's P&L.

Jensen Huang's Denny's restaurant bet just paid off. Massively.

Photo credits

All photos are sourced from Wikimedia Commons under their respective licenses:

• NVIDIA headquarters, Santa Clara, California — Coolcaesar, CC BY-SA 4.0, via Wikimedia Commons
• NVIDIA sign at headquarters — Dheeraj Jain, CC BY-SA 4.0, via Wikimedia Commons