Graphcore raised $767 million. Intel spent $2 billion on Habana Labs. Intel tried five separate GPU architectures over 27 years. AMD has been working on ROCm for a full decade and still trails by double digits. The graveyard of NVIDIA challengers keeps growing, and every headstone tells the same story. The moat is not the chip. It never was.
The $4.7 Billion Graveyard
In 2016, a company called Graphcore set out to build a chip that would dethrone NVIDIA. They were not delusional. They had brilliant engineers, a genuinely novel architecture called the Intelligence Processing Unit, and backing from Microsoft, Dell, Samsung, and the government of the United Kingdom. Over eight years, they raised $767 million. By several hardware benchmarks, their chip was competitive.
They could not get a single major customer to switch.
When Microsoft tried to run Graphcore chips on Azure, the team ran into so many software bugs that Microsoft cancelled the partnership. Not because the hardware failed. Because the software ecosystem did not exist at the depth customers needed. Graphcore eventually sold what was left to SoftBank for less than the total capital invested.
And Graphcore was not alone.
Intel acquired Habana Labs in 2019 for $2 billion. That was Intel’s fifth attempt at building a competitive GPU in 27 years. Five architectures. Billions of shareholder dollars. Each attempt failed for a variation of the same reason: CPU-centric thinking applied to a problem that requires purpose-built parallel compute silicon and a mature software stack. The entire Gaudi initiative was effectively abandoned.
Then there is AMD. Unlike Graphcore and Intel, AMD actually has competitive hardware. Their MI300X matches NVIDIA’s specs on certain benchmarks. They landed a multi-year supply agreement with OpenAI. They have been building their CUDA alternative, ROCm, since 2016.
After 10 years of investment by a company that knows how to build semiconductor products, ROCm still benchmarks 10 to 30% behind CUDA in compute performance, with a fraction of the developer community. AMD holds 6 to 8% market share against NVIDIA’s 75 to 87%.
Stop and think about the size of that gap. A well-funded, technically sophisticated competitor has been building a direct alternative for a full decade and remains meaningfully behind. Ten years is not an early-stage gap. It is evidence that the moat has a time dimension that money cannot compress.
So what is going on? Why does every well-funded, technically credible competitor crash into the same invisible wall?
The answer is fourteen words long. Jensen Huang said it on the Q2 FY2026 earnings call: “Accelerated computing is unlike general-purpose computing. It is a full-stack co-design problem.”
You cannot just build a chip. You have to build the chip, the software libraries, the developer tools, the debugging environment, the optimization framework, and the deployment pipeline. And they all have to work together seamlessly. That integration, refined over 20 years and $76.7 billion in cumulative R&D investment, is the actual moat. Not the silicon. The ecosystem.
Four million developers know CUDA. 1.5 million AI models on Hugging Face are built on it. Every one of those models would need to be substantially rewritten to run on a different platform. The switching cost is not a fee or a contract. It is millions of hours of accumulated code that becomes worthless the moment you leave.
And the pricing power confirms it is real, not theoretical. NVIDIA’s gross margins expanded from 56.9% in fiscal year 2023 to 75.0% in fiscal year 2025. During that same period, revenue grew 384%. Customers paid dramatically more per unit while simultaneously ordering dramatically more units. Price up, volume up, at the same time. When a company raises prices and demand increases, you are looking at a moat that customers cannot walk around.
Why Hardware Alone Is Never Enough
Think about your bank. You have your checking account there. Your savings account. Your mortgage. Your auto loan. Your credit card. Maybe your kids’ college fund. Each product is fine on its own, nothing special. But together, they create a tangle of direct deposits, autopays, linked accounts, and routing numbers that makes switching to a new bank feel like rewiring your house.
That is exactly what CUDA does to the AI industry, except the wiring is orders of magnitude more complex. Every AI model, every training pipeline, every inference deployment, every optimization routine is built on CUDA’s libraries and tools. Switching is not a weekend project. It is a multi-month engineering overhaul with genuine performance risk during the transition. And unlike your bank, where the new institution will happily help you move, there is no ROCm or custom ASIC onboarding team that can guarantee your models will run the same way on day one.
This is why hardware specifications are not the battleground. Products compete with platforms, and platforms win. You can pour a beautiful concrete foundation on an empty lot with no roads, no power, and no plumbers, but nobody is going to move in. Graphcore built the foundation. NVIDIA built the city.
Why This Matters for Investors
I use a concept at OIA called the Capitalism Guarantee. The question is simple: if someone had unlimited money, could they destroy this moat? For NVIDIA, we ask: if someone had $4.26 trillion in cash and nothing else to do but attack, could they destroy the CUDA ecosystem?
The answer: not quickly. The CUDA ecosystem has a 20-year time moat that capital alone cannot compress. You could hire engineers, fund alternative software stacks, subsidize customers to switch. But you cannot compress 20 years of accumulated developer adoption into five years by spending faster. AMD has been proving that for a decade.
This matters because the default assumption in technology investing is that a better product with enough funding will eventually win. The NVIDIA competitor graveyard is evidence that this assumption breaks down when the advantage is an ecosystem rather than a product. Graphcore had competitive hardware. Intel had $2 billion and five tries. AMD has a decade of direct effort. None of it was sufficient, because the advantage they were attacking is not a product to be out-engineered. It is a community to be out-adopted, and communities built over 20 years do not respond to capital the way products do.
(I have analyzed competitive dynamics across many industries. Very rarely do you find a gap this wide after this long of direct competition. Usually, 10 years is enough for a well-funded competitor to at least reach parity. The fact that AMD has not tells you something fundamental about the nature of this advantage.)
I should be honest about where uncertainty lives. There is a version of this analysis where I am wrong, where custom silicon from hyperscalers captures a larger share than I expect, where the CUDA ecosystem’s importance diminishes as AI workloads become more standardized. The moat is deep, but it operates in a fast-changing industry that requires $18.5 billion in annual R&D just to maintain the walls. This is not Coca-Cola, where the moat runs itself. NVIDIA’s moat requires constant, exceptional rebuilding.
One Thing to Watch
Watch AMD’s ROCm adoption metrics. After 10 years, ROCm still trails CUDA by 10 to 30% in benchmarks and has a fraction of the developer community. If that gap starts closing meaningfully (not in press releases, in actual developer adoption and model compatibility), it would be the first real evidence that the software ecosystem moat can be replicated with enough time and investment. So far, every year of data says it cannot. But “so far” is not “forever,” and the gap is the single most important competitive signal in AI infrastructure.
Go Deeper
This analysis is one finding from the full GAMMERS deep dive on NVIDIA. If you want the complete moat breakdown, including all six moat types, the 5 Stages of Decline check, and the custom silicon threat analysis, listen to Episode 2 of the NVIDIA podcast series.