NVIDIA’s Jensen Huang on Reasoning Models, Robotics, and Refuting the “AI Bubble” Narrative
In a Nutshell
NVIDIA CEO Jensen Huang refutes AI bubble and doomsday narratives, emphasizing 2025 breakthroughs in reasoning, grounding, robotics, and self-driving via scaling laws, plummeting costs (100x+ drops in tokens), and AI factories creating massive jobs in construction, energy, and maintenance—outpacing automation's task displacement by unlocking latent demand and productivity. Open source models (e.g., Qwen, DeepSeek) fuel innovation across a five-layer AI stack (energy to applications), enabling niches like coding (Cursor to $1B ARR), digital biology, and multi-embodiment robots, while US leads with pragmatic policies balancing China competition and export controls. 2026 predictions include reasoning cars/robots, protein/molecule design booms, and surging vertical AI firms (Open Evidence, Harvey), with grounded optimism: advance fast for safety, ignore extremes, and harness all energy sources for the AI industrial revolution.
These notes were generated by AI and may contain inaccuracies.
Host welcomes Jensen Huang. Exchange holiday greetings: Happy Hanukkah, Merry Christmas, Happy New Year, Happy Holidays.
Jensen reflects on 2025: Scaling laws and technology advancements expected. Pleased with improvements in grounding, reasoning, connection of models to search, and routers in front of models that route based on confidence to perform necessary research, improving quality and accuracy.
I'm hugely proud of that. I think the whole industry addressed one of the biggest skeptical responses of AI which is hallucination and generating gibberish.
Industry-wide leaps in reasoning and grounding across language, vision, robotics, self-driving cars. Examples: Open Evidence for medical information trusted by doctors; Harvey for legal, emerging as trusted tool for experts.
Token generation rate for inference, especially reasoning tokens, growing several exponentials. These tokens now profitable: Open Evidence at 90% gross margins; Cursor, Claude enterprise use, OpenAI enterprise all with great margins.
We're now generating tokens that are sufficiently good, so good in value that people are willing to pay good money for.
Geopolitics with China occupied much time. Spent more time traveling than ever, average elevation 17,000 ft. Importance of AI to nations, export controls, nuanced strategy for national security. Conversations on jobs, AI impact, energy, labor shortage—everything was AI.
Traditional AI community had doomsday narratives despite positives in healthcare, education, productivity. Technology shifts change important jobs but create more jobs overall.
Jensen grounds on three points in time: now, near future, distant future, plus counternarratives.
AI as software: Unlike pre-recorded software like Excel (written by hundreds of engineers, compiled, distributed static), AI takes context, generates every token for the first time every time, like human intelligence in conversation.
AI needs computers to generate tokens: AI factories producing tokens used worldwide. Also infrastructure affecting every application, company, industry, country—like energy, internet.
Three new industries/plants emerging in US at large scale:
- Chip plants (TSMC, SK Hynix building more).
- Supercomputer plants (e.g., Grace Blackwell rack is one GPU, unlike anything before).
- AI factories.
Enormous near-term jobs: construction workers, plumbers, electricians, technicians, network engineers. Electricians' paychecks doubling, traveling for work.
Geoffrey Hinton (5-7 years ago): AI would revolutionize radiology in 5 years, power all applications, radiologists no longer needed—advise against entering profession.
Now: 100% of radiology applications AI-powered, fully pervaded in 8 years. Yet number of radiologists increased.
Reason: Task (study scans) vs. purpose (diagnose disease, research). AI enables studying more scans deeply, requesting more, better diagnosis, more productive hospitals, more patients, more revenue, hiring more radiologists.
Jensen's task: typing; purpose: not typing. AI automates typing, makes him busier, able to do more work.
Example: Sister-in-law Erin leads nuclear medicine at Stanford (radiology). Doctors welcome tech, work 20 hours/day for more research, more patients.
Latent demand: Better healthcare far from peak. Productivity creates more demand, not layoffs. E.g., Nvidia more productive → explore more ideas, growth, profitability, hire more (met new hire class).
Productivity increases economy since world/problems not fully specified.
Robots address shortages: not enough factory workers (hard to retain); truck drivers short (people avoid travel, want to stay with family). Aging population worsens global labor gap.
Other shortages: accounting, nursing. AI fills gaps.
Automation creates jobs: Cars need mechanics. Robo-taxis (10 years to streets) require maintenance crews, hubs. Billion robots → largest repair industry.
AI industry booming jobs via infrastructure demand. Straight-line thinking (tools replace lawyers) misses latent demand sucking sound in economy.
Lawyer task (read/write contracts) vs. purpose (resolve conflict, protect client).
Chinese open source rising: Highest-scoring models on benchmarks (e.g., Qwen, DeepSeek). US leads closed models. Jensen long-time open source proponent.
AI = automation of intelligence, combinable with mechatronics for embodiment.
Five-layer stack:
- Energy.
- Chips.
- Infrastructure (hardware/software: data centers, Lambda, Power, Shell; orchestration software).
- AI models (system of models; understands diverse info: biological, chemical, physical, financial, healthcare, all modalities. Not just human language—biology/molecules/proteins don't understand English).
- Applications (industry-specific: Open Evidence, Harvey, Cursor, full self-driving, Figure embodied in mechanical human).
Frontier models closed-source (labs' business choice for ROI/scaling).
Open source essential: Startups, industries (manufacturing, transportation, healthcare), 100-year-old industrial/healthcare companies would be suffocated without pre-trained reasoning models to fine-tune for domains.
Open source drives data centers, telecom (Telefunken? context suggests telecom), Android/devices, industrial apps, higher ed, education, research, startups.
Whatever we decide to do with policies do not damage that innovation flywheel.
Educating policymakers: Don't forget open source, biology.
Counternarrative: Monolithic vertical player/one model to rule all, can't share crown jewel. Instead: Need diversity of AI applications. Sovereign advantage in whole stack.
Someday God AI, but on biblical/galactic scales—not soon. No company/researcher near it (supreme understanding of human/genome/molecular/protein/amino acid/physics languages).
AI like next computer industry—every company/industry/nation needs it. Don't wait for God AI; world needs to advance now.
Monolithic God AI unhelpful, extreme. If so, stop everything—even governments/policies useless.
Think practically, use common sense.
2025 theme: Extremes in public AI narratives (no jobs, God AI solves all, no open source) illogical, not happening soon.
Doom/doomer/end-of-world/science fiction narratives hurtful to people, industry, society, governments. Many in government unfamiliar with tech, influenced by PhDs/CEOs.
PhDs and CEOs describing end-of-the-world scenarios and dystopian futures to governments prompts questions about their purpose and intentions. Are they seeking regulations to suffocate startups? This raises concerns of regulatory capture to prevent new competitors. Companies should not advocate for regulations on other companies and industries. Their intentions are deeply conflicted and not entirely in society's best interest; CEOs advocate for themselves.
Focus on current reality and technology trajectory. In one year, aggressive AI investment advanced grounding, reasoning, and research, improving functionality for good. Two years ago, calls to slow the industry down were rejected, leading to progress without catastrophe.
The first part of AI safety is performance as advertised, like a car working 99.999% of the time. Significant industry investment over two to three years enhanced AI functionality. Much work remains over the next 10 years to make AI work as advertised.
Investors see companies building synthetic data generation for more grounded, diverse, less biased, safer AI, plus AI for cybersecurity.
Marginal cost of AI dropping makes it safer, not more dangerous: one AI monitored by millions of AIs, with AIs monitoring each other. AI operates surrounded by monitoring agents, like more police if safety costs less.
In 2024, cost of GPT-4 equivalent models dropped over 100x per million tokens.
Andre Karpathy estimated first ChatGPT cost versus now: trainable on a PC for tens of thousands or less in a weekend via open-source project. What was said to cost billions on supercomputers now costs nothing.
NVIDIA improves architecture and performance yearly. First GPT trained on Volta, then A100, without Hopper initially; now Hopper and Blackwell for 1.5 years. Each generation improves architecture, increases transistors and capacity, delivering 5-10x yearly. Compounded, surpasses Moore's Law (2x every 1.5 years: 10x in 5 years, 100x in 10). For AI, 100,000-1M x in 10 years from hardware alone.
Tokonomics: Token generation costs could drop a billionfold in 10 years. Training costs reduce less aggressively (2-3x yearly), but $100M or $500M today becomes 10x less next year.
Counterargument: Scale up 10-100x yearly to offset. But scale increases by 10x while computational burden does not, due to compounded hardware, algorithms, models, and shared learning.
DeepSeek paper: most important for Silicon Valley researchers recently; frontier open work benefiting American startups, labs, infrastructure. Greatest contribution to American AI last year.
American AI benefits from global researchers, including Chinese natives; ideas come worldwide.
Narratives scaring others from pre-training exaggerate costs. Due to model, algorithm, compute innovations, AI costs decrease >10x yearly. One year or six months behind keeps you close.
2025: Age of research alongside scaling. Six months behind (100k vs 200k cluster) matters in symmetric competition, but frontier labs pursue diverse progress.
Market growth enables model verticals: better coder, more accessible consumer product. Niche leaps possible without excelling everywhere; no need to boil the ocean.
Pre-training prepares for real training (post-training); now just training. Compute scaling translates directly to intelligence. Data needs are small (verifiable results); algorithmic, compute-intensive.
AI labs and startups will specialize like humans: focus energy for superhuman performance in niches. Micro-niches on open models will excel.
Prediction: Coding first application-level business to $1B ARR as AI-native app. Challenges monolithic single-model view and idea that developer tools don't scale big. Software engineering demand surges.
NVIDIA uses Cursor pervasively; every engineer uses it. Hiring incredible numbers; Monday is "come to work at NVIDIA day."
Purpose of software engineer: solve known/new problems. Coding is one task. If purpose is only coding, AI replaces; but engineers solve/explore undiscovered problems. Ideal: no coding, all problem-solving.
Framework: Purpose vs Task. Waiter purpose: great experience, not just taking orders/delivering food; reshape jobs.
Dedicated chips (CNN, transformer, NP) promised NVIDIA's end but failed. Transformers evolve: attention mechanisms, context (diffusion vs autoregressive), hybrids like Neurom (hybrid SSM).
Architecture changes rapidly; flexible/programmable enables adaptation. Moore's Law over (10% transistor gain in years); hundreds-x yearly from algorithms. Universal architecture wins.
Programmability protects large installed base: algorithms (Flash Attention, SSM, diffusion, autoregression, CNN, LSTM) run everywhere.
Drives costs down; NVIDIA's NVLink 72 lowest-cost token generation machine by enormous margin. Inference hard, especially foresees.
Falling costs open new applications/verticals, e.g., coding (Cursor, Cognition).
Several industries get ChatGPT moment due to multimodality, long context, synthetic data.
Digital biology: Protein understanding advanced; multi-protein next. LaPrina model (open) for multi-protein understanding, representation learning, generation. Protein generation advancing. Molecule design (e.g., Chai). Chemical understanding/generation, protein-chemical confirmation.
Biology/chemistry/material science need real-world data/experimentation infrastructure. Synthetic data key due to data sparsity vs language. World foundation models, protein foundation models, cell foundation models enable data flywheel.
Reasoning breakthroughs: Cars become reasoning cars (beyond perception/planning), thinking constantly, breaking down novel situations into known ones.
Addresses out-of-distribution; enables more than taught. With generative AI, multimodal vision-language-action models, reasoning: big breakthroughs in humanoid robots, multi-embodiment robots.
Self-driving eras:
- Era 1: Smart sensors, human-engineered algorithms, HD mapping (digital rails like Disneyland).
- Era 2: Perception, world model, planning modules (perception first deep learning, then propagated; brittle, limited to taught behaviors).
- Era 3: End-to-end models.
- Era 4: End-to-end + reasoning.
Started self-driving too soon; if begun 3 years ago, same place. NVIDIA self-driving stack #1 safety rated worldwide (last week); #2 Tesla. Two American companies lead.
Robotics more optimistic: foundational tech advanced. Humanoid challenges (mechatronics, 300lb weight, falling on kids), but tech applies to pick-and-place robots.
Perceived winners in self-driving are incumbents like Waymo and Tesla, with high safety ratings from NVIDIA. These companies invested significant capital over a long time, managing supply chain and hardware complexity. In robotics, Tesla with Optimus will be a major leader, but everything that moves will be robotic—a vast space beyond just humanoid forms.
AI will be multi-embodiment, like humans who embody different tools: sitting in a car, picking up a tennis racket, or using chopsticks. AIs become general purpose, adapting to one-arm pick-and-place, two arms, six arms, caterpillar shapes, excavators, tractors, etc. Construction workers embody excavators or tractors similarly.
Expect many software companies serving various verticals, plus vertical solution providers that ground the tech for perfection. Consumer AI delights at 90% success; industrial/physical AI demands 99.99999% reliability—failures cause 100% dissatisfaction. Core tech reaches 99%, vertical providers like Caterpillar push to 99.999%.
Full verticalization speeds iteration in immature markets due to control, but general-purpose tech lacks R&D scale without open source. Expect a surge in vertical AI opportunities over the next five years; excitement will be in verticalization (e.g., Open Evidence, Harvey, Cursor—horizontal but vertical-like).
AI is gonna get so good that all these wrapper companies are going to be obsolete—it misses the big point.
Non-experts undervalue domain complexity (surgeons, accountants, tax experts, bus boys). Technology often addresses tasks, not purpose—empathize with work depth.
Narrative: Will energy suffice for AI? Trump’s “drill, baby, drill” reversed anti-energy sentiment, preventing loss of this industrial revolution to others. Still power-constrained; without energy, no new industry or prosperity to address domestic/social issues.
Need every energy form: natural gas, grid power, behind-the-meter, nuclear. Wind/solar insufficient—take everything. Natural gas key for next decade. AI demand drives climate innovation: new battery companies, solar concentrators, sustainable energy with real willpower.
Doomers sound smart at dinner parties; optimists drive humanity forward.
Doomer messages shape policy, but demand creates jobs/industries. AI is biggest driver for sustainable energy ever. Balance narratives: doomers raise sensible points, optimists aren't naive—grounded optimism needed. 90% doom messaging scares investment in safer, productive AI; technology ensures safety/security (e.g., modern cars safer than 50 years ago).
Optimistic on improved US-China relations under Trump: grounded, common-sense view—adversary yet partner. Decoupling naive; economies highly coupled. Both nations invest in independence to avoid emotional dependence. Need nuanced strategy for productive relationship—most important bilateral tie for century.
Pleased with export controls grounded in national security: China makes own chips, relies on Huawei for military (as US avoids their tech). Supports technology leadership, prosperity: world's mightiest military backed by mightiest economy—wealth creates jobs, taxes, funds defense.
China sorting attitudes on US tech. Historical: Great Firewall blocked US competition while US jobs expatriated to China. Economic imbalances noted, but first principles: China's internet boomed Intel/AMD CPUs, Micron/SK Hynix/Samsung DRAM—second-largest market for US tech.
Don't isolate layers: China's internet hurt Google-like firms but prospered hardware. China contributes most to open source globally, benefiting US startups. View full lifecycle/stack: China's internet generated enormous US prosperity, not just at app layer.
Shift from general-purpose to accelerated computing as Moore's Law ended—CPUs insufficiently productive/deflationary. NVIDIA invented accelerated computing for graphics, data processing (SQL), molecular dynamics, quantum chemistry, classical ML (XGBoost, recommenders, feature engineering, ETL)—essential pre-generative AI.
Without chatbots (OpenAI, Anthropic, Gemini), NVIDIA still multi-hundred-billion-dollar company due to computing foundation shift. AI > chatbots.
OpenAI revenues: 100% capacity-utilized; double capacity doubles revenue, 10x capacity 10x revenue—they need factories for tokens like NVIDIA needs TSMC wafers/DRAM.
NVIDIA AV business nearing $10 billion (world models, robotaxis globally). Other AI: digital biology, financial services (quants shifting from classical ML/human features to AI—needs supercomputers; fastest-growing segment, billions). Robotics, digital biology: billions each.
Global GDP $100 trillion; 2% annual R&D = $2 trillion. Drug companies shift wet labs to supercomputers—$2 trillion now AI-powered, driving infrastructure.
Computing demand insane: no startup/researcher has enough capacity—global multi-company/industry shortage. $12B OpenAI revenue + hundreds of billions infrastructure refutes bubble simplisticly.
MIT study on useless enterprise AI ignores change management, reorgs, tooling integration (enterprise planning 1-year cycles). Innovation at 30-40K startups (OpenEvidence, Cursor), not slowest-adopting enterprises. Fast growth: Harvey, Sierra (enterprise ARR); end-user adoption in conservative fields (healthcare: Open Evidence for grounded research—doctors want answers, not research).
Abridge: physicians get answers easily (task vs. purpose). Jobs impossible without AI: doctors/radiologists tracking accelerating biomedical knowledge.
No more manual search/research—load papers into ChatGPT, make it learn/summarize/interact for answers/smarter insights.
AI as multi-layer cake: diverse across industries/modalities/applications/information. America wins AI across domains, not one company—open source key.
Energy is a helpful framework because factories need energy. Without energy, no factories; without factories, no AI. Better frameworks for understanding AI will make narratives more common sense, pragmatic, and balanced. Advancing technology quickly is one of the best ways to keep people safe, and the industry is doing that.
No one wants to drive a car from the first decade of cars.
ABS is a good thing. Lane keeping is a good thing. FSD is a good thing. People will be excited about the third or fourth year of AI.
The industry made tremendous strides this year across all mentioned technologies. Scaling laws remain intact: more compute leads to more intelligence. Innovations in one sector diffuse and spread across all sectors quickly.
The next five years will be extraordinary, and next year will be incredible.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
