Case Study: CUDA, Nvidia's Disruptive Platform Strategy

The usual story about Nvidia is a hardware story: better chips, bought faster than anyone can make them, at the center of the AI boom. It's true, and it misses the more important thing. Nvidia's most durable asset isn't a chip. It's CUDA, the software platform that sits between its hardware and the people who build on it, and the moat around CUDA was paid for almost entirely by other people.

That last part is what makes it worth studying. For close to two decades, researchers, framework builders, tool makers, and a generation of graduate students have poured their own time and money into building on CUDA. That accumulated investment, the libraries, the published results, the courses taught in it, the code that already runs, is the switching cost. A competitor can build a faster chip. It cannot build an ecosystem's nearly two decades of invested effort.

CUDA is a clear and current example of what we call a PIVA, a Platform of Increasing Value of Adoption: a platform whose value compounds as more people adopt it, because each new adopter and each new thing built on it makes the whole more valuable to everyone else. But the sharper lesson isn't that Nvidia built a compounding platform. It's where it built one: in the gap between two of the most successful platform incumbents in the history of computing, Intel and Microsoft, and in a system of use that neither of them was adequately defending.

What makes CUDA a platform, and not just software

It's easy to call anything with an API a platform. In our breakdown of what actually defines a platform strategy, we lay out the elements that separate a real PIVA from a product with an interface bolted on. CUDA is worth walking through against them, because it hits every one.

An evolving system of use. A platform earns its name by performing an essential function inside a system of use that keeps changing, and CUDA's has changed dramatically. It began as a way to run general-purpose math on graphics chips. The system of use then moved to scientific and high-performance computing, then to deep learning, then to the large language models driving today's demand. CUDA got repointed at each new system of use rather than being welded to the first one. That is the single most important thing it did right, and we'll come back to it, because it's also where the risk now lives.

Adopter critical mass, reached the hard way. Nvidia didn't will an ecosystem into being from nothing. It followed the same ignition path the iPhone did: it sold real, standalone products first, GPUs people bought for their own sake, and converted that installed base into a platform second. Every gaming and workstation GPU that shipped was a potential CUDA target, which is how Nvidia got past the chicken-and-egg problem that kills most platform plays before they start.

Complementors with a better deal. The compounding engine of any platform is its complementors, the outside parties who build on it and, in doing so, extend its reach without the owner paying for all of that work. The question that decides whether they show up is whether their return is higher on your platform than going it alone. For AI researchers and the builders of frameworks like PyTorch and TensorFlow, it plainly was: CUDA gave them mature tooling, working examples, and an enormous base of others solving the same problems. Their investment became Nvidia's moat.

Value capture through the hardware. CUDA is free. Nvidia captures the value the ecosystem creates through the thing it sells, the silicon. This is a clean example of the asymmetric capture that platforms run on: you earn on one side of the system while subsidizing another to keep the flywheel turning. Amazon does the same with a subsidized Kindle that earns back through book sales; Nvidia gives away the software layer and monetizes the demand it generates in chips.

Defensibility from accumulated investment. Platforms compete platform-against-platform, and CUDA's defense isn't primarily technical. It's the sheer accumulated mass of what already runs on it, the multi-link chains of dependency, and the widely shared belief that CUDA is where the ecosystem is. That belief pulls the next wave of complementors in ahead of any feature comparison. This is exactly the structural defensibility that makes a platform, once ahead, so hard to unseat, and it sets up the more interesting question of why the two incumbents best positioned to contest it couldn't.

The layer nobody was defending

Here's the part that turns a good platform story into a strategic one.

When Nvidia began building CUDA in the mid-2000s, accelerated computing for scientific and, later, AI workloads was nobody's crown jewel. Intel's business, and Microsoft's, was the PC and, increasingly, the data center, and both were pouring resources into cloud. But the bread and butter, the thing that dominated priorities and set the metrics, was still the general-purpose CPU and the software that ran on it. Massively parallel computation for a niche of researchers was a rounding error in that world.

That is precisely why Nvidia could build there undisturbed. It chose a system of use that sat beneath the incumbents' attention at the moment it mattered most, the years when a platform is fragile and a determined incumbent could still smother it. By the time that system of use grew into the center of the computing industry, the accumulated ecosystem investment was already too deep to dislodge. This is a textbook disruptive entry from below: not a better product aimed at the incumbent's best customers, but a foothold in a segment the incumbent had no reason to defend, which then became the main event.

In our language, that under-defended layer is the Priorities leg of Company Fit, the resources, processes, and priorities that determine what an organization can actually execute. Intel and Microsoft did not lack the resources or the engineering talent to build a CUDA, and they did invest in enabling open-standard alternatives, just not with the life-or-death urgency they applied to the PC. What they lacked, for a critical decade, was the priority: an organizational reason to treat a peripheral computing model as worth serious, sustained, ahead-of-revenue investment. Priorities, not capability, decided it.

The two platform kings it went around

What makes CUDA unusually instructive is that Nvidia didn't route around two weak companies. It routed around Intel and Microsoft, the two firms that had defined platform dominance for the previous thirty years.

Intel: a necessary axis, but not the decisive one

Intel met the threat first where it was most comfortable: hardware. It competed on raw performance, and on the performance-per-watt-per-dollar metrics that had governed the CPU business for years. That is the axis Intel knew how to win on, but while a necessary one, it was insufficient. The contest wasn't being decided by just the best chip. It was being decided by the platform the chips needed to support and enable.

When Intel did move to contest the platform layer, it did so through open standards positioned explicitly as an alternative to CUDA's lock-in: OpenCL early on, and later its oneAPI initiative built on SYCL, alongside dedicated AI accelerators from its Habana acquisition (the Gaudi line). The logic was sound in the abstract: commoditize the complement, deny Nvidia its proprietary hold, rally the rest of the industry around a neutral standard nobody owns. But it came too late. An open standard's whole value proposition is a shared ecosystem, and by the time these efforts had momentum the ecosystem had already sunk a decade of investment into CUDA. A standard with better governance but little accumulated complementor investment can't out-compound a proprietary platform that has it. Openness is not, by itself, an ignition strategy. This is element five of a platform strategy, competing platforms and defensibility, playing out in real time: the defender's advantage wasn't just Nvidia's cleverness, it was everyone else's investment, and an open alternative had no answer to that.

Microsoft: the Wintel arrangement, inverted

Microsoft is the subtler and, in some ways, more striking case, because Microsoft wrote the book on platform dominance and still ended up on the wrong side of this one.

In the PC era, the real platform was never Intel's silicon. It was Windows and its APIs. Developers built to Windows; the operating system was the platform, and Intel supplied the hardware substrate beneath it. That was the "Wintel" arrangement: two complementary layers, with the platform value concentrated in the software on top and the hardware as the powerful-but-commoditizable layer underneath.

The AI compute stack inverted that structure, and Nvidia captured both layers at once. It owns the hardware substrate and the developer platform that sits directly on it, the thing Intel never managed to do with x86, and the thing Microsoft had always done from the software side. So one of the most sophisticated platform operators in history now finds itself, in AI compute, an adopter and complementor on someone else's platform. Azure is one of Nvidia's largest customers. Microsoft has built its own AI silicon, the Maia accelerator announced in 2023, and pursued its own model partnerships, and still, for the working ecosystem, CUDA is the layer underneath. Building a chip did not make Microsoft a platform owner at that layer, because the platform was never only the chip.

Neither company had missed the pattern for lack of exposure to it. Both had already made a version of the same platform mistake in smartphones a decade earlier: Intel never established its chips in mobile the way it had in the PC, and Microsoft's Windows Phone never attracted the developer ecosystem that iOS and Android built, so both watched a new platform layer form around them there too. The lesson isn't that Microsoft failed; it remains one of the most valuable companies on earth, and riding a platform you don't own can be an excellent position. The lesson is that platform dominance in one system of use does not transfer to the next when the system of use evolves to a layer you don't control. Owning the PC platform bought Microsoft no automatic standing in the AI-compute platform. Each PIVA has to be won on its own terms.

The part competitors can't copy is the decision, not the software

Strip away the retrospective inevitability and CUDA looks, for most of its life, like a strange thing for a chip company to keep funding. For years it was a software platform bankrolled by a hardware business, absorbing real engineering investment to serve a market that didn't yet exist at scale. On any core-business financial yardstick in, say, 2012, CUDA was an underperforming line item.

That is exactly the trap our companion work on measuring innovation is about: judged by the metrics of the established business, long-horizon platform investment always looks like a loser, right up until it's the whole company. Nvidia's data-center business grew from roughly $15 billion in its fiscal 2023 to more than $115 billion two years later. The platform that made that possible was the one that spent a decade failing the core-business test.

What competitors cannot copy, then, isn't the software; much of CUDA's functionality can be and has been reproduced. It's the organizational fit that let Nvidia sustain the investment: the resources, processes, and priorities of a company willing to fund a compounding platform through the long years when it read as a cost center. This is the locked heart of our platform work. Platform plays typically require resources, processes, and priorities most established companies don't have, and that misfit is one of the most significant and invisible failure modes. CUDA is the rare case of a company that had the fit. Intel and Microsoft, for all their resources, did not have the priorities. That, more than any chip, is the moat.

The moat's live risk: the system of use is climbing the stack

None of this makes CUDA permanent, and the honest version of the story has to say so, because the same dynamic that built the moat is now working against it.

CUDA's great strength was riding an evolving system of use up through graphics, HPC, and deep learning. But the explosive AI software growth that CUDA enabled is now pulling ecosystem investment upward, into higher layers of the stack that are deliberately indifferent to what runs beneath them. Framework compilers and intermediate layers, such as PyTorch's torch.compile and the Triton kernel language it leans on, increasingly let developers write once at a high level and target whatever hardware is underneath. The stated aim of much of that work is precisely to make the layer where CUDA lives interchangeable.

The transition is far from complete. Today a great deal of that higher-level code still compiles down through Nvidia's own lower layers, and "hardware-agnostic" remains more aspiration than fact in production. But the direction of investment is the signal that matters. A platform's defensibility rests on complementors continuing to build at its layer. When the ecosystem's energy migrates to a layer above yours, one designed to treat you as a swappable component, the compounding that protected you starts, slowly, to work for whoever owns the new layer instead. This is the envelopment-and-abstraction risk that competitive monitoring on a platform is supposed to catch: not a rival chip, but a higher abstraction that quietly commoditizes you from above.

That is the same hinge on which platform incumbents have turned before, and it's the cautionary case we take up in next week's post. BlackBerry mistook durable-looking signals, a loyal enterprise base and years of continued subscriber growth, for a durable position, while the system of use moved to a smartphone platform it didn't own. Which is to say: being the CUDA of your era is not a finish line. It's a position you have to keep re-winning as the system of use climbs.

What's actually transferable

The unsatisfying truth is that "build a CUDA" is not advice anyone can follow. Nvidia's position came from a rare alignment of an under-defended system of use, a standalone-product ignition path, and an organization willing to fund the platform for a decade, and the fact that two far larger incumbents couldn't replicate it is the proof of how rare that alignment is.

What is transferable is the diagnostic underneath it. Before betting on a platform, run the honest disqualifier test we lay out in evaluating a platform strategy: is this a genuinely compounding platform, a real PIVA, or a reuse play carrying compounding-platform expectations it will never meet? And run it alongside the paired Company Fit question that decided the whole Nvidia, Intel, and Microsoft story: do your resources, processes, and priorities actually fit a platform play, or only a product one? The most expensive platform mistakes come from answering the first question with optimism and never asking the second at all.

That paired question, is it really a PIVA, and are we the company that can run one, is the work we do, drawing on decades of practitioner experience in corporate innovation, new business, and platform strategy. Turning it into a repeatable discipline is what our consulting and Growth Forge® Software are built to support.

BRI Associates helps companies grow by drawing on decades of practitioner experience in corporate innovation and new business development — practitioners, not pundits or academics — through direct consulting, training workshops, and Growth Forge® Software, built for the unique requirements of corporate innovation and growth organizations.