Jensen Huang – TPU competition, why we should sell chips to China, & Nvidia’s supply chain moat

What It Covers

Jensen Huang explains why NVIDIA functions as the "electrons to tokens" transformation layer, how $250B in supply chain commitments create a structural moat, why TPU competition is overstated, and why restricting chip exports to China damages American technology leadership across all five layers of the AI stack rather than protecting it.

Key Questions Answered

• •Supply Chain Moat via CEO Alignment: NVIDIA's $250B in upstream purchase commitments work because Huang personally briefs CEOs of foundries, memory makers, and packaging firms on market size projections, convincing them to invest capacity. Suppliers commit because NVIDIA's downstream demand is large enough to absorb supply. This flywheel — downstream demand justifying upstream investment — is what competitors cannot replicate without equivalent market reach and revenue velocity.
• •Bottleneck Resolution Timeline: Every hardware bottleneck in AI compute — CoWoS packaging, HBM memory, EUV machines, logic capacity — resolves within two to three years once a clear demand signal exists. NVIDIA pre-fetches bottlenecks years in advance, investing in silicon photonics ecosystems with Lumentum and Coherent, licensing patents openly to suppliers, and funding capacity expansion. The genuine long-lead constraint is energy infrastructure and skilled trades like electricians and plumbers, not semiconductor manufacturing.
• •Architecture Efficiency Outpaces Moore's Law: Moore's Law delivers roughly 25% annual transistor improvement, but NVIDIA achieved 50x energy efficiency gains from Hopper to Blackwell through co-design across processors, NVLink fabric, networking, libraries, and algorithms simultaneously. Techniques like Mixture of Experts, disaggregated inference, and new attention mechanisms each contribute 10x gains independently. This means architectural innovation, not raw lithography, is the primary lever for compute scaling.
• •CUDA Moat Is Install Base, Not Lock-In: CUDA's defensibility comes from hundreds of millions of deployed GPUs across every major cloud — A10, A100, H100, H200, L-series — meaning any framework or model built on CUDA runs everywhere. NVIDIA contributes heavily to Triton's backend and supports every inference framework including vLLM and SGLang. Developers choose CUDA first because the install base guarantees their software reaches the widest possible fleet, not because alternatives are technically blocked.
• •TPU Competition Is Concentrated, Not Broad: Huang argues that virtually all TPU and Trainium revenue growth traces back to a single customer: Anthropic, whose compute relationship with Google and AWS originated from early multi-billion dollar equity investments NVIDIA was not positioned to match at the time. Without Anthropic, neither TPU nor Trainium shows meaningful external adoption. NVIDIA's TCO benchmark InferenceMax remains unchallenged by any competing accelerator, and NVIDIA's share of external cloud workloads continues growing.
• •Tool Software Will Expand, Not Collapse: Contrary to market expectations that AI commoditizes software, Huang predicts the number of agent instances using tools like Synopsys design compilers, floor planners, and EDA tools will increase exponentially. Today's constraint is that agents are not yet proficient enough to operate these tools reliably. As agent capability improves, each software tool license effectively multiplies across thousands of AI instances, turning per-seat tools into per-agent tools and expanding total addressable markets.
• •China Export Controls Accelerate Huawei Adoption: Restricting NVIDIA chip sales to China — which represents roughly 40% of the global technology market — does not eliminate Chinese AI compute capacity because China manufactures 60% of mainstream chips, has abundant energy, and employs approximately 50% of the world's AI researchers. Huawei posted its largest revenue year on record following restrictions. The practical effect is forcing Chinese AI development onto non-American hardware stacks, reducing the global developer base building on CUDA and weakening American technology standards diffusion.