Jensen Huang
Co-founder and CEO of Nvidia; architect of accelerated computing, the CUDA ecosystem, and Nvidia’s supply chain strategy.
Last updated: 2026-04-16
Overview
Jensen Huang founded Nvidia in 1993. His central thesis: general-purpose computing (CPU) has run its course for many workloads, and the future belongs to domain-specific accelerated computing. Nvidia built CUDA not as a GPU library but as a platform for programmable accelerated computing across every domain — particle physics, molecular dynamics, image generation, data processing, and AI.
His operating philosophy: “Do as much as needed, as little as possible.” Nvidia should own the irreducibly hard parts (architecture, CUDA, ecosystem) and partner with everyone else (supply chain, clouds, model makers, application developers). This is why Nvidia doesn’t become a cloud, doesn’t pick model winners, and enables the neocloud ecosystem instead of replacing it.
Mental Models
Electrons to Tokens
“The input is electrons, the output is tokens. In the middle is Nvidia.” Every component Nvidia doesn’t need to build, it partners and makes part of the ecosystem. The hard part — the transformation — isn’t commoditizable because it requires continuous co-design: architecture, numerics, networking (NVLink, Spectrum-X), system software, and algorithm invention. Blackwell was 50x Hopper; transistors alone contributed only ~75% improvement. The rest came from architecture and computer science.
Five-Layer AI Stack
AI is a five-layer cake:
- Energy — foundational; without energy, nothing else scales
- Chips — accelerated compute hardware
- Computing stack — CUDA ecosystem, libraries (cuBLAS, CUDA-X), frameworks
- Models — foundation models trained on the stack
- Applications — the layer that diffuses into society and benefits most
Key implication: the US must win at all five layers simultaneously. Export control policies that sacrifice the chip layer to protect the model layer are, in Jensen’s view, self-defeating. Conceding the second-largest market (China) strengthens Huawei and pushes global AI developers off the American tech stack.
Architecture vs. Transistors
Moore’s Law delivers ~25% per year in transistor scaling. But Hopper→Blackwell delivered 50x. The gap — ~2000% — came from algorithm co-design: MoE architectures, disaggregation, new attention mechanisms, hybrid SSMs. Great computer science is the lever, not chip density. This is why programmable architectures (CUDA) outcompete purpose-built ASICs/TPUs over time.
Key Positions
On TPU Competition
Nvidia built accelerated computing (not a tensor processing unit). TPUs are optimized for matrix multiplies; Nvidia runs everything. The CUDA flywheel: rich ecosystem → large install base → developers build on CUDA first → ecosystem compounds. Nvidia’s TCO is the best in the world; no TPU has shown up on InferenceMAX or MLPerf to challenge it.
Anthropic using TPUs is not a trend — it’s a unique historical artifact. Nvidia didn’t invest in Anthropic early (before the multi-billion VC was needed); Google and AWS did; they got the compute relationship as a result. Jensen’s stated miss: he didn’t internalize that VCs couldn’t fund foundation labs at that scale. He’s since invested in OpenAI (~10B).
On Supply Chain Moat
Nvidia made ~250B). The deeper moat is implicit: Jensen personally informs, inspires, and aligns CEO-level relationships at every supply chain company. They invest for Nvidia because they see the downstream demand is real. GTC is part of this — it’s as much supply chain education as developer conference.
Bottlenecks (CoWoS, HBM, EUV machines, plumbers) are solved by swarming them with capital and demand signal; none last more than 2-3 years. What worries Jensen more: energy policy and physical workers (electricians, plumbers) — the things that don’t scale with capital alone.
On China Export Controls
Jensen is strongly against broad AI chip export controls. His reasoning:
- China already has abundant compute (energy, chips at 7nm, Huawei at record revenue), most of the world’s AI researchers (50%), and the technical capability to build competitive models
- Compute advantage (1/10th US flops) can be compensated with scale and abundant cheap energy — AI scales with watt-hours, not just transistor count
- Restricting exports accelerated Huawei’s chip industry and forced China’s AI ecosystem onto non-American hardware
- 50% of AI developers in China programming on CUDA is an American strategic asset; losing that market means future AI models optimized for non-American stacks
- The American telecom industry was regulated out of the world market — Jensen considers this a cautionary tale
- Counter-argument he doesn’t accept: that compute is like enriched uranium. He calls AI “a chip” they can make themselves, whereas nuclear fission is in a different category
His preferred approach: dialogue between US and Chinese AI researchers on what AI should not be used for; ensure US labs get first access to each generation; invest in open source AI security ecosystem.
On Neocloud Strategy
Nvidia doesn’t become a cloud because “if we didn’t do it, somebody would show up.” Their job is the irreducibly hard parts. They invest in neoclouds (CoreWeave, Nscale, Nebius) not because they want to be in the financing business, but because enabling the cloud layer to exist grows the overall Nvidia ecosystem. Once a neocloud has its flywheel going, Nvidia steps back.
On Not Picking Winners
Nvidia invests in OpenAI, Anthropic, and as many foundation model companies as possible — without betting on any single one. Lesson from the graphics era: of 60 3D graphics companies, Nvidia survived (despite being “precisely wrong” architecturally at the time). Nobody could have predicted it. Enough humility to let winners emerge.
Roadmap Commitments
- Vera Rubin → Vera Rubin Ultra → Feynman → [unnamed] — one generation per year
- Groq being folded into CUDA ecosystem for high-ASP fast-response token segment
- Inference market segmenting: throughput tokens (cheap, slow) vs. premium tokens (fast, high ASP, software engineer productivity use case)
Connections
- agent-first-software — five-layer stack framing; 100–1000x agent volume aligns with Jensen’s “number of tool users is going to skyrocket” thesis
- dark-code — Jensen describes the AI cybersecurity ecosystem (thousands of AI agents watching one AI agent) as the correct response to AI offensive capabilities — same accountability problem Guo names
- agentic-engineering — “as little as possible” philosophy mirrors thin-harness-fat-skills; don’t build what partners can build
- auto-research — Jensen’s radiologist/software engineer point: Jevons paradox — AI expands demand for the skills it augments
- model-context-protocol — Jensen is skeptical of MCP; prefers CLIs; his CUDA philosophy is that you build ecosystems through programmability, not new protocol layers