Jensen Huang’s Year-End Case That Token Costs Can Fall a Billion-Fold Without Killing Jobs

Sarah Guo and Elad Gil open a New Year conversation with Jensen Huang after a year he describes as spent mostly in the air. His average elevation, he jokes, was about 17,000 feet: export controls, geopolitics, jobs, energy, and labor shortage all ran through NVIDIA. Scaling laws and raw model progress did not surprise him. What pleased him was grounding, reasoning, models wired to search, and routers that send a query to research when confidence is low. The industry, he says, finally answered the skeptic’s charge that AI hallucinates gibberish — across language, vision, robotics, and self-driving. What did surprise him a little is how fast inference tokens, especially reasoning tokens, are growing: several exponentials at once. And they are finally profitable. OpenEvidence, a medical-information company, is at 90% gross margins; Cursor’s margins are great; Claude’s are great; OpenAI’s enterprise business is great. Tokens are now good enough in value that people will pay real money for them. That, he says, is the year’s solid ground.

The jobs fight is where he wants to start, because a doomer streak in the AI community predated models that even worked. Gil notes the oddity: the people pushing the field were often the most pessimistic, and that mood captured a slice of the press. Huang’s near-term counter is infrastructure, not vibes. Excel was written by a few hundred engineers, compiled, and shipped as-is for years. AI generates every token in context, every time, which is why he calls the buildings AI factories. Three kinds of plant are going up across the United States at once: chip plants (TSMC, SK Hynix), supercomputer plants — a Grace Blackwell rack is one GPU — and the AI factories themselves. Construction workers, plumbers, electricians, technicians, and network engineers are in sudden demand. He is delighted electricians’ paychecks have doubled and that they are being paid to travel the way executives fly on business trips.

Then the Hinton story. Five, six, seven years ago, Huang recalls, Geoffrey Hinton said that within five years every radiology application would be AI-powered, radiologists would not be needed, and it would be the first profession to avoid. The first half landed: a hundred percent of radiology applications are now AI-powered. The second half did not. Headcount went up. A job has tasks and a purpose. The radiologist’s task is studying scans; the purpose is diagnosing disease, plus research. Better tools mean more scans, read more deeply, more patients, a more productive hospital, more revenue, and a reason to hire. Huang’s own task is typing; his purpose is not typing. AI that types for him has made him busier, because he can do more work.

Guo brings it home: her sister-in-law Erin leads nuclear medicine at Stanford. The doctors welcome the tools and work twenty-hour days to do more research and see more patients. Nobody, she says, thinks American or global healthcare has reached the mountaintop. Latent demand absorbs the productivity. Huang agrees: if NVIDIA gets more productive, it does not lay people off; it explores more ideas. Gil met a new hire class that morning; the company is hiring every week. Only if the world’s problems were already fully specified would productivity shrink the economy. They are not.

The same split applies to lawyers and waiters. Reading and writing contracts is a task; the purpose is resolving conflict and protecting the client. Taking an order is not a waiter’s job; giving a great experience is, even if an AI takes the order or delivers the food. NVIDIA’s software engineers all use Cursor. Coding is a task. The purpose is to solve known problems and find undiscovered ones. “Nothing would give me more joy,” Huang says, “than if none of them are coding at all. They’re just solving problems.” Guo’s year-ago prediction — that coding would be the first AI-native application category to a billion dollars of ARR — no longer sounds exotic. Developer tools were never a small niche once you priced the demand for software engineers.

Robotics, in this telling, fills gaps rather than emptying payrolls. The United States does not have enough factory workers; the economy is limited by that headcount. Truck driving is short because people do not want to sleep in a cab and miss their families for months. Accounting and nursing are short too, and aging makes the holes worse worldwide. Automation covers the shortage. People also forget the aftermarket: robotaxis took a decade to reach the street, and they already support maintenance crews and hubs. A billion robots, Huang says, would be the largest repair industry on the planet. Gil’s phrase is a “sucking sound” in the economy toward latent demand, while policymakers fixate on not shrinking what already exists.

Gil turns to China and open source: on closed models the United States still leads; on open benchmarks, Qwen, DeepSeek, and others now sit near the top. Huang wants a framework before a slogan. AI is automation of intelligence, optionally embodied with mechatronics. The stack is a five-layer cake: energy at the bottom; then chips; then infrastructure — land, power, shells, data centers, and the hardware-software orchestration; then models; then applications. Models are not chatbots. Biology, chemistry, physics, finance, and healthcare are information too. Molecules and proteins do not speak English. Applications range from OpenEvidence and Harvey and Cursor to full self-driving in a car and Figure in a humanoid body.

Frontier labs may keep closed models; that is their return-on-investment math. Without open pre-trained bases, though, startups, higher education, research, manufacturing, transportation, healthcare, and hundred-year industrial firms Huang works with would be suffocated. They need something pre-trained, some reasoning substrate, then they fine-tune into a domain. Open source already drives data centers and a large share of the world’s phones. He spends a lot of time telling policymakers: whatever you decide, do not damage that innovation flywheel — and do not forget biology.

Guo voices the counter: one vertically integrated champion, one model that does it all, a crown jewel that must not go abroad. Huang’s reply is dry. Someday there will be God AI, on biblical or galactic scales. Jumping there from today is not a plan. No company or researcher is close to an intelligence that is supreme at human language, genomes, molecules, proteins, amino acids, and physics at once. Meanwhile every company, industry, and nation needs computers next week. AI, at the simple level, is the next computer industry. A monolithic nation-state God AI is too extreme to be useful; if you believed it, you might as well stop writing policy, because God AI would walk around it.

Gil lists the year’s extremes: nobody will work, God AI will solve everything, open source is too dangerous even as it already powers industry. Huang calls the damage real. Respected PhDs and CEOs have walked into governments with end-of-the-world, science-fiction stories. Many officials are less fluent in the technology. The question he wants asked is purpose and intention: why describe a dystopia to people who write rules that can suffocate startups? Gil asks whether that is regulatory capture. Huang will not mind-read, but as practice he does not think companies should lobby governments to regulate other companies. Their interests are conflicted by definition.

Two years ago the talk was slowing the industry down. Instead it invested, and grounding, reasoning, and research got better. The end did not come. The first safety of a product, he says, is that it performs as advertised — a car that works 99.999% of the time, not a lecture about someone using it as a missile. Making AI work as advertised takes technology, and the next ten years are mostly that work. Guo’s world of synthetic-data and cybersecurity startups is part of it. The cheap-AI-is-dangerous story runs the wrong way: if marginal cost collapses, one AI is monitored by millions of others. Agents surround agents, the way cheaper public safety puts a police officer on every corner.

Gil quotes a friend: doomers sound smart at dinner parties; optimists drive humanity forward. Huang will not call every warning irrelevant or every optimist naive. He wants balance. When 90% of the public message is doom, people get scared out of the investments that make AI safer, more functional, more useful. Safety and security are technologies. His car is safer than a car from fifty years ago because of technology, not because someone paused the industry.

Guo’s team clocked more than a 100× drop in 2024 in the cost of a million GPT-4-class tokens, even as some model companies talk up capital moats. Karpathy’s estimate, she notes, is that the first ChatGPT can now be reproduced for tens of thousands of dollars, or less — he has an open-source project you can run in a weekend on a PC. Huang needles the last clause: not quite a PC; think a Spark-class machine. Still, three years from “billions and a supercomputer” to a weekend is the tokenomics story. Early GPT trained on Volta, then Ampere; the first breakthroughs missed Hopper. Blackwell has been in the field about a year and a half, with Rubin around the corner. Architecture plus transistors plus capacity delivering 5–10× of compute every year is not unusual.

Moore’s Law was 2× every eighteen months: 10× in five years, 100× in ten. AI hardware over ten years, he says, is probably 100,000× to a million× — before algorithms and model architecture. A billion-fold drop in token-generation cost over that span would not surprise him. Training cost falls a bit more slowly, maybe 2–3× a year, which is still “next year it is ten times less” when someone quotes a $100 million or $500 million training bill. The objection is that labs just scale 10× to eat the savings. Mixture-of-experts is why the compute burden does not rise one-for-one with parameter count. NVIDIA stays programmable because CNN chips and then transformer chips were each going to kill it, while attention, diffusion versus autoregressive, and hybrid SSMs — Nemotron is one — keep mutating. After Moore’s Law, transistor gains are maybe 10% over a couple of years; the hundreds of × live in algorithms. NVLink 72, he says, is now the world’s cheapest token-generation machine by a lot, because MoE inference is hard.

DeepSeek may have been the single most important paper Silicon Valley researchers read in years — the only thing that felt frontier and open. American startups, labs, and infrastructure companies learned from it. He calls it, out loud, perhaps the greatest contribution to American AI last year, even if saying so makes people shudder. Researchers in the United States already come from everywhere; not every idea has to be born here. Cost narratives that say nobody else should pre-train are, in his view, about scaring people out of the market. Fall a year, even six months, behind, and you can still stay close. Ilya’s line that the field is back in an age of research is half right: scaling continues, and models will specialize — better coder, better consumer product — so a niche leap no longer requires boiling the ocean. Pre-training is not over; it is preparation for training. The data for that later stage can be small, maybe just verifiable results. Compute then maps more directly to intelligence.

Asked what gets a ChatGPT moment in 2026, Huang starts with digital biology. Multimodality, very long context, and synthetic data — biology is data-sparse compared with human language — should unlock protein understanding that is already moving fast, then protein generation, then chemistry and protein–chemical conformation. NVIDIA recently open-sourced a model for multi-protein understanding, representation learning, and generation. The real break is a foundation model for proteins and one for cells; once those exist, the data flywheel starts. Guo adds that the inputs are physical: wet-lab data, new experimental infrastructure, not only GPUs.

The second bet is reasoning on wheels. Cars become reasoning cars, not only perception-and-planning cars. Faced with a scene they have never seen, they decompose it into familiar pieces and construct a path. Out-of-distribution tails are a reasoning problem. Self-driving, he says, had four eras: smart sensors and digital rails (Mobileye, early Waymo, heavy maps); modular perception, world model, and planning, which is brittle and only does what it was taught; end-to-end models; and end-to-end plus reasoning. The industry started too soon. Begin three years ago and you would be in the same place. NVIDIA’s stack just took a global number-one safety rating, with Tesla second — two American companies on top, he notes with pride. Robotics can jump the lost decade because the foundation tech is already paid for.

Humanoids still have mechatronic problems: a 300-pound machine falling next to a child is not a rounding error. The same core tech goes into a pick-and-place arm. Gil asks whether incumbents win again, as in robotaxis. Huang will not give Tesla’s Optimus the whole map. Everything that moves will be robotic, and every AI will be multi-embodiment — the way a person embodies a car, a tennis racket, or chopsticks. One arm, two arms, six arms, an excavator, a bulldozer. Software companies can serve many verticals; each vertical still needs a solution provider to grind the last nines. Consumer AI at 90% delights. Industrial AI at 90% is 100% dissatisfaction, because people only see the failing 10%. Core tech might reach 99%; a Caterpillar takes it to 99.9999%. His five-year call is verticalization — OpenEvidence, Harvey, Cursor as a “horizontal vertical.” The “wrapper companies die when God AI arrives” line, he says, is what you say when you have never been a surgeon, a tax expert, or a busboy. Technology hits the task. It does not hit the purpose.

Energy is the bottom of the cake. Huang defends the first-week “drill, baby, drill” turn in the Trump administration: without a sentiment shift toward energy growth, the United States would have handed this industrial revolution to someone else. The country has been power-starved for a decade. Wind and solar will be taken; they will not be enough. Nuclear is coming. For the next decade, natural gas is probably the only path that meets 2027–28 cluster demand. Guo’s causal claim is that AI electricity demand is now the strongest pull for climate tech — batteries, concentrators, SMRs. History may record AI as the largest driver of sustainable energy yet. Demand, Huang says, beats a doomer policy memo.

His 2026 geopolitical hope is a more constructive U.S.–China relationship. The administration’s common-sense frame, as he hears it, is adversary and partner at once. Decoupling, for philosophy or national security, is not common sense once you look at how coupled the countries are. Independence is healthy so the relationship is not too emotional; a productive tie between the two most important nations is the century’s most important relationship. Export policy, in his telling, should rest on national security, technology leadership, and prosperity: China’s military already has Huawei and ample silicon, just as the U.S. military is too smart to run on Chinese parts. The mightiest military is funded by the mightiest economy. On the historical grievance — the Great Firewall, jobs gone east — he zooms out to the stack. China’s internet boom was a boom for Intel, AMD, Micron, SK Hynix, and Samsung. It is the second-largest internet market for American technology. Maybe it hurt Google and Uber. Do not erase every other layer. No country, he says, contributes more to open source than China, and American startups ate that work.

Then the question Gil’s investor friends will not forgive him for skipping: is there an AI bubble? Huang splits computing from chatbots. NVIDIA invented accelerated computing for graphics, SQL, molecular dynamics, quantum chemistry, classical ML, XGBoost, recommenders, feature engineering. Moore’s Law ending makes the shift from general-purpose CPUs inevitable. If ChatGPT, Claude, and Gemini had never existed, NVIDIA would still be a multi-hundred-billion-dollar company. The bubble test that jumps straight to OpenAI revenue is capacity-blind: twice the factories, twice the revenue; ten times, ten times — the same logic as NVIDIA needing TSMC wafers. AI is also an AV business coming up on $10 billion, quants leaving human feature engineering for supercomputers, digital biology, robotics. Simple math: ~$100 trillion world GDP, about 2% a year in R&D, $2 trillion. Five years ago the largest drug companies spent that in wet labs. Today they build supercomputers. That $2 trillion is moving. Every startup and every university lab he meets is dying for capacity. A $12 billion revenue number set against hundreds of billions of infrastructure, he says, is too simplistic a proof of a bubble.

The MIT study that most enterprise AI deployments were not useful gets the same treatment. Guo: did they do change management, a reorg, workflow integration? Enterprise planning cycles are a year; six months is early. Huang would not look for innovation inside the enterprise at all — the slowest adopter. Talk to the 30,000 or 40,000 startups. Talk to OpenEvidence and Cursor. Harvey and Sierra are already at multi-hundred-million enterprise ARR. Healthcare is the most conservative industry, and it moved because doctors are terrified of the wrong answer; grounded tools that deliver answers, not search, win. “Nobody wants to do research. They want answers. Nobody wants to do search. They want answers.” He no longer reads arXiv the old way. He dumps papers into ChatGPT, makes it learn them, summarizes, then interrogates. Keep people safe, he closes, by advancing the technology quickly. Nobody wants a car from the first decade of cars. ABS, lane keeping, FSD are good things. Scaling laws still map compute to intelligence, and a trick in one sector diffuses fast. The next five years, he thinks, will be extraordinary.