Living question · AI infrastructure
Is CUDA still what protects Nvidia?
The current read
The software lock-in is weakening. The systems and supply lock-in is not. Sonya Huang and Dylan Patel both describe the CUDA moat as at least partly disentangled, but Patel relocates the advantage rather than retiring it — to hardware-software co-design and to supply chains. Jensen's own case, made about eleven weeks earlier, never rested on programmability alone.
What changed · between 15 April and 30 June 2026
Entering the April conversation, the framing is that per-chip switching cost — having to rewrite kernels — is what stops anyone selling against Nvidia.
Sonya Huang is the one who dates the change, to the last three to six months. Dylan Patel gives the mechanism without attaching that window: models got good enough at coding that the software layer commoditizes. Neither claim lands cleanly on Jensen, whose April case was already broader than programmability — workload breadth beyond TPUs, the ecosystem and install base, presence in every cloud, and pinch points he can press on TSMC and ASML.
The moment that moved it
Dylan Patel
SemiAnalysis, on Sequoia Capital —
“Why Hardware-Software Co-Design Is AI's Real 100x”
…certainly the CUDA moat and software moat is at least partially, uh- … disentangled because, you know, models are just great at coding, and all software gets commoditized in that case. I do think there is some level of, like, open source and, you know, what people call the CUDA moat is not actually anything to do with CUDA…
Word-timestamped · speaker-attributed · verbatim
The strongest case that the moat holds
“…there are just, you know, a whole bunch of applications that we can address that you can't do so with TPUs. Because NVIDIA's built CUDA as a fantastic tensor processing unit as well, but it does, you know, it does every, every lifecycle of data processing and computing and AI…”
Jensen Huang · Dwarkesh Patel · 15 Apr 2026 · 19:23
“CUDA, CUDA is, um, uh, it's a, a rich ecosystem, and so if you wanna build on any computer first, building on CUDA first is incredibly smart, and because the ecosystem is so rich, uh, we support every framework…”
Jensen Huang · Dwarkesh Patel · 15 Apr 2026 · 25:58
“…if I can convince TSMC, AS- ASML will be convinced. And so that's, that… You know, we have to think about the critical, critical pinch points…”
Jensen Huang · Dwarkesh Patel · 15 Apr 2026 · 14:47 — the supply-side version of the moat, stated plainly
“…supply chains matter, what technology you can bring in matters, and more and more as the industry gets bigger, supply chain diversification happens.”
Dylan Patel · Sequoia Capital · 30 Jun 2026 · 46:52 — the same source as the counterpoint below, relocating the advantage rather than retiring it
The strongest counterpoint
“But I remember for a long time thinking, you know, one, the programmability of NVIDIA and then just CUDA as, as such a big moat. It seems to me that narrative has kind of changed, at least in my mind for the last three to six months. Like, model companies no longer care about if we have to write custom kernels for, you know, this other chip, so be it.”
Sonya Huang · Sequoia Capital · 30 Jun 2026 · 35:50
“…what people call the CUDA moat is not actually anything to do with CUDA…”
Dylan Patel · Sequoia Capital · 30 Jun 2026 · 36:50 — later in the moment above
What would change this answer
Watching: whether memory, not compute, becomes the binding constraint. If it does, the moat question stops being about software at all and becomes a question about who has DRAM and HBM supply — where Nvidia's position is different and weaker.
“No, well, one, DRAM is the most important bottleneck.”
Gavin Baker · All-In Podcast · 27 Jun 2026 · 1:03:36 — the constraint asserted flatly, and ranked first
“…memory is the bottleneck. I mean, it's true, but like…”
Dylan Patel · Sequoia Capital · 30 Jun 2026 · 42:24 — three days later, waved off as a shallow take. The disagreement is live, which is why this has not moved the read
Method & sources
Quotations are verbatim from this engine's word-level transcripts, with omissions marked and nothing reordered — the hesitations are left in. Speaker names are resolved against a voice database and corrected by hand where it has no match. Selection and comparison are edited, not generated. Every timestamp above links to the second it was said. When a read changes, the earlier framing stays visible rather than being overwritten.
- Dwarkesh Patel — Jensen Huang: Will Nvidia's moat persist? · 15 Apr 2026
- Sequoia Capital — Why Hardware-Software Co-Design Is AI's Real 100x, Dylan Patel · 30 Jun 2026
- Bg2 Pod — SpaceX IPO, AI Capex Update & Market Check · 11 Jun 2026 — reviewed, not cited in this read
- All-In Podcast — China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter · 27 Jun 2026
This dossier is a prototype assembled from already-processed episodes to show the format. No reads have been published yet, so there is no correction history to show.