Jensen Huang — Will Nvidia's moat persist?

Key points

• Nvidia's mental model: input is electrons, output is tokens. The transformation is 'hard, artful science' that won't commoditise. Do as little as possible but own the hardest 5% — and ecosystem-partner the other 95% across upstream supply chain + downstream operators + model makers + app developers.
• On the 'AI will commoditise software' fear: opposite direction. Number of agents explodes → tool usage explodes → Synopsys, Cadence, Excel-style instances skyrocket. Software is limited today by number of engineers; tomorrow every engineer has N agents driving the existing tools.
• Nvidia's moat is the flywheel, not CUDA alone: (1) programmability unlocks algorithmic invention (new attention, hybrid SSM, diffusion-autoregressive fuses); (2) install base of hundreds of millions of GPUs; (3) presence in every cloud + on-prem; (4) best perf/TCO + perf/watt, which = tokens/watt = revenue per data-centre-gigawatt.
• On TPU / Trainium: 'Anthropic is a unique instance and not a trend.' TPU growth is 100% Anthropic. Trainium growth is 100% Anthropic. Without Anthropic, the ASIC narrative collapses. OpenAI's AMD/Titan is a bet-hedge, not a Nvidia displacement.
• Jensen's biggest strategic regret: 'I didn't deeply internalise that Anthropic had no other options — a VC would never put $5-10B into an AI lab.' Anthropic needed supplier-as-investor; Google and AWS stepped up; Nvidia wasn't in a position to. Hence $30B reported in OpenAI, $10B in Anthropic now — won't make the same mistake again.
• Hopper → Blackwell delivered 50x energy efficiency. Moore's Law gives 25%/year. The rest is algorithmic + system co-design: MoE, disaggregated parallelism, offload into NVLink fabric, Spectrum-X, CUDA kernel work. Programmability is what makes this possible — fixed-function ASICs can't get these leaps.
• TSMC N3 share: AI is 60% of N3 this year, 86% next year (per Semianalysis). Jensen's answer to 'how do you 2x from here?': all physical bottlenecks (CoAS, HBM, EUV) are 2-3 year problems. The industry swarms them when the demand signal is clear. Real long-term constraint is energy policy and permitting.
• On doomers: 'we're short of radiologists.' Ten years of warnings that AI would end radiology produced a shortage, not a surplus. Same will happen with software engineers. Don't let the doom narrative talk talented people out of the field.
• On ASIC-margin-as-savings: 'ASIC margins are actually very good.' Broadcom's take on a custom chip eats much of the notional Nvidia-margin savings. Pair that with 2x kernel optimisation gains from Nvidia-provided engineers inside labs, and the TCO math flips back.
• On supply chain: Nvidia spends huge time educating its ecosystem about what's coming. Most keynotes contain a 'torturous edge-education' section so upstream (memory, foundry, optics, power) and downstream (clouds, system builders) can reason about scale the way Jensen does and make their own multi-year capex commitments.

Notable quotes

Themes

• Why the transformation from electrons to tokens won't commoditise
• Nvidia's moat is a flywheel, not CUDA alone
• Anthropic is a one-off, not the start of the ASIC revolt
• Energy policy as the only permanent AI bottleneck