CUDA Software MoatWide moat

Nvidia (NVDA) — moat facet

The real fortress isn't measured in copper but in code and trained human beings — twenty years of the whole field's labor and 7.5 million developers, accumulated on Nvidia's behalf.

CUDA is the real fortress, and it is worth understanding why, because it is so easily overlooked by people who fixate on the chips. CUDA is the software layer that lets programmers actually harness Nvidia's hardware, and because researchers have written against it for the better part of two decades1, an enormous body of code, tooling, and — most importantly — human skill now assumes Nvidia sits at the bottom of the stack; the FY2026 annual report counts more than 7.5 million developers using CUDA and Nvidia's other tools3, up from 4.7 million two years earlier4. That is not an advantage you can photograph or put on a spec sheet, which is exactly why it is so underrated and so powerful.

Developers using CUDA and NVIDIA tools, per 10-K (millions)1.2MFY191.6MFY202.2MFY213.0MFY223.8MFY234.7MFY245.9MFY257.5MFY26NVIDIA Forms 10-K FY2019-FY2026, business section; each an 'over' figure
The developer base grew more than sixfold in seven years, to over 7.5 million by the FY2026 annual report.

The crucial consequence is that a competitor can match the silicon and still lose the sale. Suppose a rival ships a chip that is just as fast and rather cheaper. To adopt it, a customer must rewrite or re-optimize a great deal of software, re-validate that everything still works, and retrain the very people whose expertise is the customer's scarcest resource. Faced with that cost and that risk, most customers conclude that the cheaper chip is not actually cheaper at all once the switching costs are counted — and they stay put. The moat, in other words, is measured not in copper but in code and in trained human beings.

Every major AI framework in wide use assumes Nvidia beneath it2, tuned and optimized for its hardware over years of collective effort. That framework dependence means the default path for anyone starting new work runs straight through Nvidia, without a decision ever being consciously made. The tools simply expect it, the tutorials assume it, and the path of least resistance leads there.

It helps to see why this is a fundamentally different and sturdier kind of advantage than a fast chip. A chip lead is a lead in a thing you sell; a software ecosystem is a lead in a thing your customers build upon and depend on. The first can be answered by a rival who simply builds a better thing. The second can only be answered by persuading the customer to tear down and rebuild years of their own work — which is why hardware leads are so often temporary and ecosystem leads so often prove permanent. Nvidia was shrewd, or lucky, or both, to invest so heavily in the software when the chips alone were the fashionable thing to sell.

And switching costs of this kind compound quietly and relentlessly. Every library painstakingly optimized for CUDA, every graduate student who cut their teeth on it, every production model tuned to run just so on Nvidia hardware is one more brick in a wall that no one deliberately built and no rival can quickly tear down. It is the accumulated residue of twenty years of the whole field's labor, and that is why the deepest part of Nvidia's moat is not something it manufactures each quarter, but something the entire industry has been unwittingly building on its behalf for two decades.

Moat trajectory: Widening

Widening. CUDA is the reason Nvidia's lead is so much more durable than a chip lead alone: nearly two decades of libraries, tools, and optimizations that the entire AI field is written against. Every new model, framework, and research paper built on CUDA adds another layer a competitor would have to replicate — not just the silicon, but the whole software world sitting on top of it. As AI development accelerates, more code is written to CUDA every day. The software moat widens with the field itself.

The number that tests this moat
Third-party estimate
Share of AI accelerators / CUDA lock-in
~90%+ of AI training compute

CUDA's moat shows up as market share: an estimated ~90%+ of AI training runs on Nvidia because the software stack is where the switching cost lives. If that share eroded as rivals' software matured, the fortress would be cracking.

Nvidia does not disclose accelerator market share; the ~90%+ figure is an analyst estimate of AI-training compute, corroborated by rivals' disclosed revenue (AMD's whole Data Center segment is a fraction of Nvidia's).
Source: Industry analyst estimates of AI-accelerator share (IDC / TechInsights coverage) ↗
Aspects of the moat
⚠ Threats to the moat
References
  1. ReportedCUDA launched in 2006-2007, roughly two decades of accumulated code and skill.
    NVIDIA — CUDA platform (introduced 2006–2007; millions of registered developers, management-disclosed) — 2006–present · publ. 2006–2026 · source ↗
  2. ReportedMajor AI frameworks (PyTorch, TensorFlow/JAX) ship CUDA-first, NVIDIA-optimized backends.
    PyTorch project — CUDA-first backend and release optimization for NVIDIA hardware — Current · publ. 2016–2026 · source ↗
  3. ReportedOver 7.5 million developers use CUDA and NVIDIA's other software tools (FY2026 10-K).
    NVIDIA Corporation, Form 10-K (FY2026), business section - 'There are over 7.5 million developers worldwide using CUDA and our other software tools' — Fiscal year ended 25 January 2026 · publ. Filed Feb 2026 · source ↗
  4. ReportedOver 4.7 million developers used CUDA and NVIDIA's other tools two years earlier (FY2024 10-K).
    NVIDIA Corporation, Form 10-K (FY2024) - 'There are over 4.7 million developers worldwide using CUDA and our other software tools' — Fiscal year ended 28 January 2024 · publ. Filed Feb 2024 · source ↗
Sources
Generated September 18, 2026