Some thoughts on why agentic AI will reinforce Nvidia's moat rather than weaken it.
While many argue agentic AI poses a structural risk to CUDA, I believe the opposite: semiconductor ecosystems like Nvidia's compound over time, and this only makes it stronger.
Point 1 - Not all AI software is created equal: Frontier models generate better code for Nvidia because the installed base and public corpus for CUDA are the deepest. You can test this yourself: even Gemini writes better kernels for CUDA than for TPUs. It's true agents can also write software for challenger ecosystems, but their starting capability is strongest on the dominant platform. Nvidia enjoys a tremendous data advantage. In addition to the vast public corpus, the company itself holds a rich proprietary dataset of every stack optimization, kernel, and improvement it has ever made. As long as AI model capabilities remain a direct function of training data volume and quality, this creates powerful compounding of Nvidia's competitive advantage.
Point 2 - The installed-base advantage will be compounded by agents: Nvidia's leading installed base means far more developers are writing, optimizing, and iterating on CUDA than any competitive platform. This is arguably the single biggest reason its ecosystem has stayed dominant for so long (Jensen has alluded to this repeatedly). The same broader dynamic applies powerfully to agents: Nvidia holds the largest installed base of both developers and deployed high-performance compute, which should result in far more agents being trained and deployed on its infrastructure than on any competing platforms. I also think his comment to Dwarkesh is underappreciated: Nvidia's architecture is so thoroughly 'wrung out' by its ecosystem that its systems are simply the most reliable.
Point 3 - Porting workloads effectively is becoming harder, not easier: The notion that agents will easily port code across platforms misses the point. The minimum viable port is now easier, but doing it effectively (ie maintaining full performance and capabilities) is becoming much harder for two reasons:
First, many recent frontier gains come from low-level stack innovations that are inherently hardware-specific: While high-level frameworks like PyTorch and JAX dominate mindshare, the biggest leaps now often happen lower down (e.g., using Nvidia’s PTX for a novel attention algorithm). These optimizations are tightly coupled to each platform’s memory hierarchy, execution model, compiler, networking, and runtime. They rarely port cleanly, and performance usually drops when they do.
Second, as architectures scale beyond the chip into full datacenter fabrics, porting complexity explodes. As we go to 10 trillion and eventually 100 trillion parameter models, the workload itself must be spread across the entire datacenter. The bigger the system, the greater the differentiation between platforms at the most fundamental level.
So while agents compound the performance of Nvidia's own systems, much of that will remain unique to its platform because the translation is more physically challenging, or in some cases impossible.
Point 4 - Nvidia can accelerate with AI faster than its competitors: Nvidia has the most engineers, the biggest/broadest/best performing stack, unmatched scale, and the deepest understanding of its own architecture. It is already applying AI aggressively and just rolled out Codex internally to 10,000+ engineers (with more to come). It also has more deployed compute than anyone else, so more agents can work on Nvidia's own stack. My view: they are already moving with blazing speed and the pace at which agents can catapult Nvidia forward will likely exceed what challengers can achieve.
Point 5 - CUDA's existing substrate lets agents do higher-value work: Leading systems benefit from extreme hardware-software co-design, where libraries must communicate precisely across the entire stack to control data flow and algorithms. Nvidia has spent decades building those libraries inside CUDA, which remains central to its historical leadership. This lets CUDA agents skip the scaffolding and focus immediately on higher-value optimizations, work that often doesn't yet exist in challenger ecosystems.
Point 6 - Hardware leadership is the backstop: It’s worth remembering that so much of this still comes down to physical architectures, and that Nvidia remains the undisputed leader here. Even Broadcom has publicly acknowledged that Nvidia is setting the pace for the entire AI landscape. The head of CUDA spends half his time with the hardware engineers for a reason. Nvidia enjoys a serious hardware advantage in what have become deeply interconnected rack-scale systems: extreme co-design, a generally unique physical architecture, and a deep well of proprietary technology. Google’s recent pivot with its TPU 8i network topology to Boardfly is a clear acknowledgment that Nvidia was correct all along. Meanwhile, Nvidia continues to charge ahead on hardware innovation.
The larger and deeper your physical architecture, the more beneficial agents become to your unique platform, and the more defensible it is against competitive agents trying to copycat it. In a true-AGI world where software could theoretically write perfect kernels for every stack, hardware leadership is the ultimate protection. Nvidia has it. Hardware remains a major differentiating factor.
Point 7 - In production, the ceiling matters more than the floor: Nvidia has improved single-GPU inference by more than 1,000x over the last decade through full-stack co-design. Agentic AI will accelerate this loop and push the frontier even further. Nvidia typically ships new algorithms as reliable, optimized, production-ready workloads faster than anyone else. In a market increasingly priced on tokens per watt per dollar (which maximizes revenue per GW), that deployment edge translates directly into superior price-to-value.
For frontier labs operating early in a multi-trillion-dollar GenAI TAM, every engineering hour (or token) spent porting to a secondary stack is an hour not spent improving model quality, efficiency, latency, or utilization on the dominant platform. That is a risky bet. As long as the world remains competitive and the opportunity remains enormous, the infrastructure ROI for frontier labs is clearly in favor of leaning into Nvidia rather than away.
TLDR; Semiconductor ecosystems compound and Nvidia’s will continue to do so in the agentic era. Agents write the best code for CUDA, more agents will build on CUDA, and CUDA provides the richest foundation to build on. Nvidia is also better positioned than anyone to use AI to accelerate its own stack. Collectively, this should raise Nvidia’s ceiling faster than challengers can raise the floor. Nvidia’s hardware leadership and the trend toward datacenter-scale workloads and architectures also provide protection against competitive agents while enhancing the value of its own.All in, I believe these dynamics create an environment where agentic AI acts as a tailwind rather than a headwind for $NVDA, and I suspect that's exactly how it will play out.
