4 trillion NVIDIA, why?

In July 2025, history was refreshed once again.

In July 2025, history was refreshed once again. NVIDIA, a company founded by a Chinese-American who loves leather jackets, saw its market value rocket past $4 trillion, leaving traditional giants behind and becoming the absolute core of the global capital market.

In an instant, applause, gasps, bubble theories, and doubts intertwined. Media headlines were dominated by Jensen Huang's quotes, astonishing wealth effects, and the grand narrative of AI devouring everything. But for every decision-maker caught in the wave of the industry—whether investors, corporate strategists, or technology leaders—the real questions were far more important than the fluctuations in stock prices:

What exactly supports this vast empire? Is it the GPUs that are being frantically snapped up? When AMD, Intel, and even major cloud providers claim to have their own AI chips, why does NVIDIA's "throne" seem unassailable? After reaching $4 trillion, how will its growth story continue?

To find answers, we decided to adopt an "old-fashioned" but most effective method—engaging in deep conversations with those who truly shape this industry. Silicon Rabbit utilized our expert network in Silicon Valley to communicate with several anonymous experts on the front lines of AI. Among them were former heads of AI infrastructure from top cloud providers, chief architects leading large model training, and top VC partners who assess the next technological trends on the Silicon Valley road.

Now, please allow us to present these precious firsthand insights to you. This is not only a dissection of a company but also a deep analysis of the core driving forces of an era.

01 The deepest moat lies in the invisible code

When we asked nearly all interviewed experts the same question—"What is NVIDIA's core barrier?"—not a single person's answer was "chip performance." Instead, they all pointed to a product born nearly twenty years ago—CUDA.

A senior technical director who once built AI platforms at FAANG opened our conversation with a vivid metaphor:

"The biggest cognitive bias from the outside is still viewing NVIDIA as a hardware company. It's like thinking Coca-Cola's success lies only in its bottle. Since Jensen Huang officially launched CUDA in 2006, he has not been selling chips but 'preaching.' He built a 'NVIDIA cult,' and CUDA is its scripture.

Today, any customer who buys an H100 or B200 is not just paying for the silicon; they are also purchasing a 'ticket' to enter this cult ecosystem. This is an intangible 'ecological tax' that almost everyone must pay."

CUDA (Compute Unified Device Architecture) is the somewhat awkwardly named starting point of all NVIDIA's myths. When GPUs were merely "treasures" for gamers, Jensen Huang foresaw the need to invest heavily to open up the heart of GPUs—the thousands of parallel computing cores—for general scientific and commercial computing This grand chess game has been in play for nearly 20 years.

It is not just a product, but an ecosystem. CUDA is not merely a programming interface; it encompasses a complete set of rich, highly optimized mathematical libraries (such as cuDNN for deep neural networks, cuBLAS for linear algebra), powerful compilers, intuitive debugging tools (like NVIDIA Nsight), and a vast developer community.

It creates a perfect example of network effects. The more developers use CUDA, the more CUDA-based applications and frameworks (such as TensorFlow, PyTorch) will emerge; these killer applications will attract more users and developers to engage in the CUDA ecosystem. Once this positive flywheel starts turning, the gravitational pull it generates will be immense.

Today, there are over 4 million developers worldwide using CUDA. For any AI PhD student, their first line of model code is almost always run on CUDA. This has formed a powerful "muscle memory" that has spread from academia to industry, becoming the de facto industry standard.

02 Invisible Costs, Visible Barriers

"Since CUDA is so powerful, can't competitors like AMD's ROCm or Intel's oneAPI create a better alternative?" This was the question we posed to a chief AI architect responsible for large model training. He smiled and countered:

"Do you know what the real cost is of migrating the core AI business of a leading company from the NVIDIA platform to another platform? It's not just the hardware cost of purchasing tens of thousands of new chips, but a long and despairing 'technical bill' that could amount to several times, or even more than ten times, the hardware costs."

With the help of this expert, we got a glimpse of the iceberg of this "technical bill":

Code Refactoring and Migration: This is far from a simple "find-and-replace." Countless engineers have painstakingly hand-written computation kernels optimized for NVIDIA GPUs, which must be almost entirely rewritten for AMD or Intel chips. The underlying hardware architecture differences involved are unimaginable to outsiders.

Performance Optimization Hell: Even if the code is successfully migrated, the new hardware cannot achieve the performance of the NVIDIA platform "out of the box." Engineers need to spend months or even years on tedious performance tuning, resolving various unexpected bugs, to slowly "approach" the original efficiency. For the time-sensitive AI competition, this time cost is fatal.

The Toolchain Gap: NVIDIA provides extremely mature performance analysis and debugging tools like Nsight and NVProf, which help engineers quickly identify bottlenecks. In contrast, competitors' toolchains still lag years behind in stability, usability, and feature richness. This architect candidly stated, "A problem that can be solved in an afternoon on NVIDIA might take a week on other platforms, and you still don't know where the problem lies." The talent pool gap: A harsh reality is that the number of engineers proficient in CUDA in the market may be hundreds or thousands of times greater than those proficient in ROCm. For companies, this means higher recruitment costs, longer training cycles, and significant risks of project delays.

Ecological inertia: Model communities like Hugging Face have the vast majority of open-source models pre-trained and optimized for NVIDIA GPUs. When a team wants to quickly validate a new idea, the fastest path is always to "download the model and run it on NVIDIA GPUs."

"To summarize," the architect said at last, "NVIDIA's moat was not dug by itself, but built over the past fifteen years by millions of developers around the world, line by line of code, debugging time after time, and project after project. Filling this moat requires not money, but time, and a similarly large and loyal army of developers. As it stands, no one can achieve that."

03 Upward Integration: From Selling Shovels to Selling "Gold Mining Factories"

If CUDA is NVIDIA's "software soul," then its "hardware" evolution strategy is equally wise. A top VC partner with 20 years in Silicon Valley provided us with a unique business perspective:

"To understand NVIDIA's business model, you can't just look at GPUs; you need to see how its 'average transaction value' has been gradually increasing. This is a textbook case of 'Upward Integration.' It is essentially not selling products, but continuously solving larger and more valuable problems for customers."

This top VC partner depicted NVIDIA's strategy as a four-stage rocket:

Stage One: Selling "parts" - GPU chips. This is the starting point. From G80 to Fermi, and now to today's Blackwell architecture, NVIDIA has always maintained a leading position in single-card performance. This is the foundation of all its business.

Stage Two: Selling "equipment" - DGX/HGX servers. NVIDIA quickly realized that what customers needed was not eight independent GPUs, but a "monster" that could efficiently coordinate these eight GPUs. Thus, it tightly coupled GPUs using high-speed interconnect technology NVLink and NVSwitch, launching the DGX server. It was no longer selling parts, but a "plug-and-play AI supercomputer." The average transaction value jumped from thousands of dollars to hundreds of thousands of dollars.

Stage Three: Selling "production lines" - SuperPOD clusters. When customers need to train models with hundreds of billions or trillions of parameters, a single DGX is no longer sufficient. NVIDIA connected hundreds or thousands of DGX servers into a massive cluster using the InfiniBand high-speed networking technology acquired from Mellanox, providing a complete set of software to manage it. This is SuperPOD. It is no longer selling equipment, but a complete blueprint for an "AI model production line." The average transaction price has soared to tens of millions or even hundreds of millions of dollars.

Level 4: Selling "factories" - data center-level solutions. Today, NVIDIA is moving towards its ultimate form. It has partnered with cloud service providers to launch DGX Cloud, allowing customers to rent a complete "AI factory" on demand. It even directly participates in the design of customer data centers. What it sells is the "AI capability" itself.

Through this layered strategy, NVIDIA has transformed itself from a chip supplier into an indispensable "general contractor" providing full-stack solutions in customers' AI strategies. Each integration addresses deeper pain points for customers, resulting in higher profit margins and stronger customer loyalty.

Conclusion

At this point, the story seems legendary enough. But for a $4 trillion empire, its ambitions go far beyond this. Take NVIDIA AI Enterprise (NVAIE) as an example; it is like the "Windows operating system" of the AI era. After purchasing NVIDIA's hardware, enterprises can subscribe to NVAIE services in exchange for the stability, security, technical support, and performance guarantees necessary for running critical business operations.

This not only opens up a brand new, high-profit software subscription market for NVIDIA but, more importantly, transforms its relationship with customers from one-time transactions into long-term service partnerships.

When this "hardware + software + services" full-stack capability is refined to perfection, it perfectly aligns with one of the most important new trends of the 21st century: Sovereign AI.

An expert focused on geopolitical technology revealed the final chapter of NVIDIA's story to us:

"We are entering an era of 'Sovereign AI.' Every country will realize that having its own independent AI infrastructure, its own foundational large models, and AI trained on its own data is part of national sovereignty in the 21st century, as important as having its own currency and military. And who can provide these countries with the complete set of tools to build 'Sovereign AI'? Today, the answer is only one - NVIDIA."

This elevates NVIDIA beyond the realm of a commercial company; its products have become a strategic resource in 21st-century geopolitics. This not only opens up a new blue ocean market defined by "nations" but also enhances the certainty and irreplaceability of its business to unprecedented heights.

$4 trillion. This number is not a myth, nor a bubble.

Author of this article: GuiTuJun, Source: 36Kr, Original title: "Why NVIDIA is Worth $4 Trillion? We Discussed the Answer with Several Core Experts from Silicon Valley"