Nvidia's most valuable product may not be a chip at all — its CUDA software platform has created switching costs so high that even faster, cheaper rivals struggle to break in.
Back
Nvidia's CUDA software moat drives 75% gross margins and $82B revenue
Nvidia's most valuable product may not be a chip at all — its CUDA software platform has created switching costs so high that even faster, cheaper rivals struggle to break in.
Nvidia's revenue nearly doubled to $82 billion in its latest quarter from $44 billion a year earlier, while gross margin expanded to 75% from 61%, according to company filings. The numbers reflect a business that dominates artificial intelligence computing, but the driver extends beyond hardware specifications.
"Developers have spent two decades building AI tools on CUDA, and they rarely want to rebuild from scratch," said Stacy Rasgon, senior analyst at Bernstein. "The switching costs are enormous, even when a competing chip offers better specs."
The CUDA platform, launched in 2007, now supports more than 5 million developers worldwide, according to Nvidia's published estimates. Its network effect operates as a self-reinforcing cycle: more developers using CUDA makes the ecosystem more valuable, which makes Nvidia's hardware more attractive, which draws in more developers. The result is a competitive barrier that no rival has yet breached, even as AMD, Amazon, Google, and Microsoft pour billions into alternative chips.
The stakes for investors extend beyond any single product cycle. Nvidia's software moat means the company competes on ecosystems, not just chip specifications — a dynamic that has historically allowed Apple, Microsoft, and Amazon to sustain dominant positions for decades. If CUDA's network effect holds, Nvidia's revenue growth and margin expansion could persist through multiple hardware generations.
Why CUDA's network effect is hard to replicate
The switching cost problem for Nvidia's rivals is not about performance — it is about installed base. A company that has spent years optimizing AI models on CUDA faces retraining engineers, rewriting software, and testing applications before moving to a competing platform. That friction exists regardless of whether the alternative chip is faster or cheaper.
AMD has made the most credible push at the edge of Nvidia's territory. OpenAI signed a 6-gigawatt agreement with AMD in October 2025, starting with a 1-gigawatt MI450 deployment in the second half of 2026. Meta followed in February with its own 6-gigawatt AMD deal, worth more than $100 billion according to the Wall Street Journal, including warrants that could give Meta up to 10% of AMD if milestones are met.
Yet even these commitments do not threaten Nvidia's core business in the near term. AMD's Instinct GPUs will serve specific hyperscaler workloads, but the broader market of enterprise AI buyers — banks, hospitals, manufacturers — remains locked into CUDA. Those customers are not building custom silicon or rewriting software stacks. They are buying Nvidia's integrated hardware-software package.
The ASIC threat is real but distant
The bigger long-term risk comes from custom silicon. Amazon's Trainium, Google's TPU, Microsoft's Maia, and Meta's MTIA are designed to serve one company's specific workloads very well, bypassing Nvidia's general-purpose advantage. Broadcom, the quiet partner behind much of this shift, announced a 10-gigawatt custom accelerator agreement with OpenAI in October 2025, with deployments beginning in the second half of 2026.
Broadcom, Apollo, and Blackstone also launched a $35 billion AI XPV Platform this month, targeting more than 20 gigawatts of compute capacity through 2028, including Anthropic's previously announced expansion of more than 1 gigawatt. ASIC shipments are growing at triple the rate of merchant GPUs, according to industry data cited by analysts.
But custom silicon takes years to develop and deploy. Nvidia's GB300 NVL72 rack-scale system, which beat its own GB200 platform by 45% on DeepSeek R1 inference tests in the latest MLPerf benchmarks, shows the company is still improving from a position of strength. The 130-terabyte-per-second NVLink fabric across the 72-GPU rack creates a system-level advantage that discrete chips cannot match.
For investors, the calculus is straightforward. Nvidia shares trade at roughly 35 times forward earnings, a premium that reflects the market's confidence in its ecosystem moat. The risk is that hyperscaler custom silicon programs eventually erode that advantage, but that scenario remains years away. In the meantime, CUDA's network effect continues to generate the kind of recurring revenue and margin expansion that hardware-only chipmakers cannot replicate.
This article is for informational purposes only and does not constitute investment advice.
