Bildnachweis: Image via Source article. Used under fair use for news commentary. · source
Can anyone actually dethrone Nvidia in AI chips before the end of the decade? If you follow this space, you have probably asked yourself this question more than once. The company's grip on data center AI hardware has seemed almost absurdly secure for years now, and a recent analyst forecast from DA Davidson suggests that grip is not loosening anytime soon.
Gil Luria, DA Davidson's head of technology research, argued this week that Nvidia's profit margins (we are talking 70% or higher) are "relatively safe through the year 2030" because hyperscalers simply have no viable alternatives for the chips powering their AI data centers. The claim appeared on Bloomberg Surveillance and was subsequently reported by Bloomberg.
It is a bold forecast. And while I am generally skeptical of sweeping multi-year predictions in semiconductor markets (I know I am being picky here, but this industry moves fast), Luria's argument rests on a structural reality that is worth taking seriously.
To be precise, the argument is not that Nvidia makes the best chips. It is that the entire AI training ecosystem has been built around CUDA, Nvidia's proprietary software platform, and that switching costs are so high that even well-funded alternatives struggle to gain traction.
This is not a new observation. Researchers have been writing about CUDA lock-in for years. What is notable about Luria's framing is the timeline: he is essentially saying that even with all the investment pouring into alternative architectures (AMD's ROCm, Intel's oneAPI, various custom silicon efforts from Google, Amazon, and Microsoft), none of them will meaningfully erode Nvidia's position before 2030.
Verwandte Beiträge
More in AI Models
A $30 billion cloud computing agreement isn't really about cloud computing. It's about something weirder.
Sarah Williams · 8 hours ago · 4 min
A technology analyst argues hyperscalers have no real alternatives to Nvidia's data center chips, but the reasoning deserves closer examination.
Aisha Patel · 8 hours ago · 7 min
Morgan Stanley's CIO thinks the semiconductor rally is built on a misunderstanding, and honestly, I'm starting to wonder if she's right.
Sarah Williams · 12 hours ago · 4 min
Wealth managers are staring down an AI threat, and it's not humanoid robots — it's chatbots that clients actually trust.
The key points of his argument, as I understand them:
Hyperscalers are locked into CUDA-based software stacks that took years to develop
Rewriting those stacks for alternative hardware is expensive, time-consuming, and risky
Even where alternatives exist (Google's TPUs, Amazon's Trainium), they are optimised for specific workloads and do not offer general-purpose replacement
Nvidia continues to iterate faster than competitors, maintaining its performance lead
The sheer capital expenditure already committed to Nvidia-based infrastructure creates path dependency
This is, actually, the research shows, a reasonable structural argument. But it has some significant gaps.
Luria is correct that switching costs are real and substantial. Anyone who has worked on porting a large ML codebase from CUDA to an alternative framework knows this is not a weekend project. The tooling, the debugging infrastructure, the optimised libraries: all of it has been built around Nvidia's ecosystem for over a decade.
He is also correct that hyperscalers' custom silicon efforts, while impressive, remain narrowly scoped. Google's TPUs are excellent for certain training and inference workloads, but they are not drop-in replacements for H100s or whatever Nvidia ships next. Amazon's Trainium chips are, by most accounts, still catching up on software maturity.
And the margin argument has historical support. Nvidia has maintained gross margins above 60% for years, even as it has faced nominal competition. The company's pricing power is, to put it bluntly, remarkable.
Here is where I start hedging. Forecasting that margins are "safe through 2030" requires assuming that the current competitive landscape persists for four more years. That is a long time in AI hardware.
It is worth noting that we do not actually know what AMD's MI400 series will look like, or whether Intel's Falcon Shores (assuming it ships on schedule, which, well) will close the performance gap. We also do not know how aggressively hyperscalers will push their own silicon once they have had a few more years to mature their software stacks.
The sample size for "AI chip transitions" is small. We have basically one data point: the shift from CPUs to GPUs for ML training in the early 2010s. That transition took longer than anyone expected, which supports Luria's argument. But it also eventually happened, which suggests that "safe through 2030" might be overstating things.
I would also note that Luria's forecast is, by its nature, based on limited data. Analyst predictions about semiconductor market structure five years out have a mixed track record. (Remember when ARM was supposed to take over the data center by 2020? I remember.)
What I find more interesting than the margin forecast itself is what it implies about the AI hardware market's structure. If Luria is right, we are looking at a market where a single company maintains near-monopoly pricing power on the most strategically important technology of the decade.
That seems... unstable? Not in the sense that Nvidia's position will collapse, but in the sense that it creates enormous incentives for well-funded actors to find alternatives. Hyperscalers are spending tens of billions annually on AI infrastructure. Even a 10% reduction in chip costs would be worth billions to them.
The question is whether those incentives translate into actual competitive pressure before 2030. Luria is betting they do not. I am less certain.
If I were evaluating this forecast more rigorously, I would want to know:
First, what is the actual utilisation rate of custom silicon at hyperscalers today? Google, Amazon, and Microsoft all claim their chips are in production, but we have limited visibility into what percentage of their AI workloads actually run on proprietary hardware versus Nvidia GPUs.
Second, how are training frameworks evolving? PyTorch and JAX have both made progress on hardware abstraction. If the software layer becomes more portable, the switching costs Luria emphasises could decline faster than expected.
Third, what happens when (not if) China develops competitive AI accelerators that cannot use CUDA? The geopolitical dimension of this market is, in a way, underappreciated in most analyst coverage.
None of this means Luria is wrong. His core argument about CUDA lock-in and switching costs is sound. But "safe through 2030" is a strong claim, and the methodology behind it remains unclear. DA Davidson has not, to my knowledge, published the detailed modelling underlying this forecast.
For now, I would frame it this way: Nvidia's near-term dominance is probably secure. The 2028-2030 window is genuinely uncertain. And anyone claiming to know exactly how this plays out is, to be precise, overconfident.
