Nvidia
Nvidia Built the Only Road That AI Can Drive On — And Now People Are Building New Roads
Based on 139 related nodes across 8 research explorations in the AI sector
The Basic Situation
Imagine that every car in the world runs on a special kind of fuel, and one company owns the only refineries that make it. That company is Nvidia. The fuel is the combination of their hardware (GPUs — specialized chips) and software (CUDA — the programming system everyone uses to talk to those chips). And right now, essentially all serious AI development runs through Nvidia.
This is not an exaggeration. The research data shows that Nvidia’s GPU economics is the single most connected node in the entire AI infrastructure graph — meaning more things depend on it, push against it, or feed off it than any other single factor. Nvidia is not just a player in AI. It is the structural foundation most other players are built on top of.
The interesting question is how long that lasts, and what threatens it.
How Nvidia Got Here: The Software Trick Nobody Talks About
Most people think Nvidia is powerful because they make the best chips. That is true, but it is not the main reason they are so hard to displace.
The real reason is something called CUDA — a programming language and library system Nvidia has been building for 20 years. Think of it like this: imagine if there was one universal electrical socket standard that every appliance, every factory, every hospital in the world used. Now imagine that standard was patented, deeply embedded in every engineering school curriculum, and had 4 million trained electricians who knew nothing else. That is CUDA.
When AI researchers want to train a model or run an application, they write code in CUDA. When a company wants to hire AI engineers, those engineers know CUDA. When open-source AI communities share tools and libraries, those tools are built on CUDA. Switching away from CUDA does not just mean buying a different chip — it means re-writing years of software, retraining your entire team, and giving up access to a vast library of pre-built tools. The research estimates this takes months to years per workload.
This is what makes Nvidia’s position structurally durable in a way that raw hardware performance alone never could be.
The Money Machine: Why Nvidia Earns 88 Cents of Profit on Every Dollar
Nvidia’s hardware margins are extraordinary. An H100 chip — their current flagship AI processor — costs roughly $3,300 to manufacture and sells for around $28,000. That is an 88% gross margin, which is essentially unheard of for physical hardware.
How does a company charge ten times manufacturing cost? Because there is no real alternative. When a company needs to train a large AI model, they need Nvidia’s chips. When a cloud company needs to offer AI services, they buy Nvidia’s chips. When a government wants to build an AI data center, they buy Nvidia’s chips. There is a near-complete absence of substitutes for the specific combination of hardware performance and software ecosystem that Nvidia provides.
There is also a clever built-in mechanism that keeps demand high: Nvidia releases a new generation of chips roughly every two years. Each new generation is good enough that the previous generation becomes economically inefficient for serious AI work. This is like a treadmill — companies that bought chips two years ago now face a choice between falling behind competitors or buying new ones. The treadmill generates perpetual upgrade demand regardless of whether the overall AI market grows.
The Strangest Strength: Nvidia Wins When AI Gets Cheaper
Here is a counterintuitive finding from the research. You might think that if AI models get cheaper to run, Nvidia sells fewer chips. The data suggests the opposite.
When the price of running an AI request falls — because models become more efficient or competition drives prices down — demand for AI applications explodes. People who couldn’t afford AI at $10 per use suddenly use it at $0.10 per use, a hundred times more. Total compute consumed goes up, not down. Economists call this the Jevons Paradox — cheaper energy historically leads to more energy use, not less.
The same dynamic applies to open-source AI. When Meta releases a free, powerful AI model that anyone can download and run, you might think that hurts Nvidia. But every person or company that downloads and runs that model needs chips to run it on. And those chips almost universally run on CUDA. Nvidia has committed $26 billion over five years specifically to supporting open-source AI development — not out of charity, but because every open-source model deployment is a future CUDA workload. The research calls this the “Open-Source Infrastructure Paradox” — Nvidia benefits from the proliferation of technology it does not control.
The Three Real Threats
1. The Big Tech Companies Are Building Their Own Chips
Google, Amazon, Microsoft, Meta, and Apple are all designing their own specialized AI chips — called ASICs or XPUs. Google’s latest chip (TPU Ironwood) reportedly offers 4.7 times better performance per dollar than Nvidia’s H100 for certain tasks. Amazon’s Trainium chips offer 30-40% better price performance for their specific workloads.
Why does this matter? These companies collectively represent a massive fraction of total AI compute spending. If Google runs Gemini on its own chips, that is revenue Nvidia never sees. If Amazon trains models on Trainium, same story.
The complication for Nvidia is that these companies are not switching chips for everything — custom silicon is particularly good at a specific type of AI work called “inference” (running an already-trained model to answer questions), as opposed to “training” (the expensive process of building the model in the first place). Nvidia’s strongest advantage has historically been in training. As the AI industry matures, more compute shifts toward inference. That shift plays to Nvidia’s competitors.
2. China Is Now Permanently Off the Table
The US government banned Nvidia from selling its most capable chips to China. Then, in April 2025, it banned even the downgraded chips Nvidia had designed specifically to comply with earlier restrictions. This resulted in a $5.5 billion charge to Nvidia’s earnings in a single quarter.
China was a large market. It is now largely closed to Nvidia hardware. Meanwhile, Huawei has been building its own AI chip ecosystem (called Ascend) that currently delivers roughly 60% of Nvidia’s performance, with its own software system. That ecosystem is growing within China, creating a separate AI hardware world that Nvidia cannot participate in.
This is not a temporary trade dispute. The research treats the US-China AI chip split as a permanent structural fracture.
3. Nvidia Is Lending Money to Its Own Customers
This is the strangest and most hidden risk. Nvidia has committed over $100 billion in financing to its own customers — including $100 billion to OpenAI — structured so that those customers use the money to buy Nvidia chips. The revenue Nvidia reports partially comes from money Nvidia itself provided.
The research draws an explicit parallel to Lucent Technologies, a telecommunications equipment company that collapsed in 2001 after it vendor-financed its customers to buy its own equipment. When customers couldn’t repay and stopped buying, Lucent’s revenue collapsed simultaneously with its balance sheet. The mechanism is identical. This risk is within Nvidia’s control — it is a choice they are making — but it inflates reported revenue in a way that is difficult to assess from the outside.
The Non-Obvious Finding: The Export Controls Might Be Backfiring
The US restricted China’s access to powerful Nvidia chips partly to slow down Chinese AI development. The research identifies a paradox: those restrictions may have accelerated it.
China’s AI researchers, unable to use the most powerful training hardware, had strong incentives to make their models work with less. DeepSeek — a Chinese AI lab — released models that matched or exceeded Western models while using a fraction of the compute. The efficiency innovations came directly from the pressure of hardware scarcity.
Meanwhile, China is building domestic rare earth processing into a potential deterrent. China controls over 90% of global rare earth processing — materials critical to semiconductor manufacturing. The research characterizes this as a lever China has not yet fully deployed, but that structurally constrains how far the US can escalate chip restrictions before triggering supply chain retaliation that would hurt Nvidia’s own manufacturing.
Where Nvidia Has Real Options
Inference software. Nvidia knows inference is where it is most exposed. It has been building software tools (called NIM — Nvidia Inference Microservices) specifically designed to make its GPUs more competitive for inference workloads. If these tools can close the price-performance gap with custom silicon, the training-to-inference shift becomes less threatening.
Sovereign AI. Nations — not just companies — are buying GPU clusters as a matter of national strategy. France, Japan, Saudi Arabia, and others want domestic AI infrastructure. These government buyers are price-insensitive in a way that corporate customers are not, they lack the engineering depth to build custom silicon alternatives, and they are buying Nvidia because CUDA is the de facto standard. This market is structurally favorable for Nvidia.
Supply chain insurance. Nvidia currently gets roughly 90% of its high-bandwidth memory (a critical chip component) from a single South Korean supplier. That is a concentration risk. Micron, a US company, is building US-based manufacturing capacity. Locking in long-term supply agreements with multiple suppliers would reduce a vulnerability Nvidia currently cannot fully control.
The Bottom Line
Nvidia is the most structurally central company in AI — more things depend on it than on any other single player. Its CUDA software ecosystem is a genuine 20-year moat that protects it even when hardware alternatives emerge. Its margins are extraordinary. Its position in the open-source ecosystem is cleverly constructed to convert industry growth into CUDA demand regardless of which model or company wins.
But three forces are pushing against it simultaneously. The big tech companies are building around it in the inference layer where future compute will concentrate. China is permanently gone as a market, and Chinese alternatives are catching up. And Nvidia has structured its financing in a way that makes its revenue partially circular and exposed to the same credit dynamics that have historically preceded hardware industry collapses.
The core question the research leaves open is whether CUDA’s software depth can hold the inference layer — the growing center of gravity — as custom silicon closes in. If it can, Nvidia’s position is durable for years. If it cannot, the monopoly economics erode faster than the Architecture Treadmill can compensate.
Nvidia built the only road AI can drive on. The threat is not that anyone tears up that road — it is that people gradually build new roads around it, one lane at a time, starting with the busiest section.