NVIDIA's Jensen Huang: Typing Is a Cheap Commodity. Here's What's Actually Valuable.

Calling typing a cheap commodity is an insult only if you think commodities are insults. Wheat is not bad; it is just abundant, and its price reflects the fact that almost anyone can grow it. The AI era is producing text the way farms produce grain: reliably, at scale, and at a marginal cost that keeps falling every quarter. When Jensen Huang said that coding is essentially typing and that typing is becoming a cheap commodity, he was not sneering at the people who type for a living. He was pointing at the wheat and telling everyone to stop betting their careers on its price. That message lands differently for a founder than for a typist, and the founders who read it correctly tend to win the next two quarters.

The mechanism is marginal cost, and it is worth stating precisely. The first page a model writes costs you a prompt; the ten-thousandth costs the same prompt, which is the definition of a commodity input. Once the marginal cost of producing output hits zero, the market stops rewarding production and starts rewarding everything that is not production, the same logic that made hand-assembled calculators worthless and chip design valuable. In commodity economics this is standard: when one input falls in price, the value does not vanish, it migrates to the bottleneck that remains. The wheat gets cheaper, and the mill that knows which grain to buy gets more valuable with every price drop.

This is why the quote is economic instead of personal. Huang was describing a structural shift that applies to everyone at once, not a critique of one profession's effort or a jab at keyboard users. His follow-up line did the arithmetic for us: what is really valuable is not the answer, but what questions you can ask. Answers are the wheat now, mass-produced by AI at negligible cost, and questions are the scarce input, because a question is a compressed bundle of judgments, priorities, and context that no model generates on its own. Every genuinely useful advance in tooling moves the premium further up that chain.

For a founder or a strategist, the lesson is to stop competing in the wheat. Nobody builds a commodity strategy around grain prices and expects a premium, and nobody should build a career around keystroke speed in an era when the machine produces all the keystrokes. The positioning question of this cycle is where the bottleneck sits after the machines take production, and the answer is upstream: in intent, in taste, and in the ability to decide what the machine's typing should be spent on. The rest of this article maps that migration in economic terms and ends with a self-test you can run on your own workday to see where you actually stand.

There is an odd asymmetry in the way AI panned out: answers became free long before questions did. Producing an answer is a search problem, and search problems scale beautifully with compute: more parameters, more data, more tokens, better answers. Producing a question is a compression problem, because a good question is everything you know about a situation folded into a few sentences, including the parts you did not know were relevant until you said them out loud. The industry spent years making answers better, and the result is that answers are now the cheapest thing in the entire pipeline, while the question still has to be carried into the machine by a person.

NVIDIA's own release pattern shows the asymmetry from the supply side. On October 28, the company open-sourced Nemotron 3.5 Transcribe: 40 languages, speaker diarization for up to eight speakers, an hour of audio at a time, plus the open dataset Granary 2 for training and evaluation. TechCrunch described the moment as voice AI in vogue. With the Riva speech library and NIM microservices, the full pipeline runs completely offline on a single GPU, per NVIDIA's own technical documentation. Speech capture, the raw answer-making side of voice, is open, cheap, and local, and the price curve is only heading one way, which is toward zero.

But here is the gap nobody closed. The transcript is verbatim, so it is full of ums, you knows, restarts, and abandoned sentences, because that is how people actually speak when they are thinking. The machine heard you and wrote down every word you said, and it still has no idea which part was the point. Capturing the speech was easy; capturing the intent was not. That single distinction is the line between the commodity and the value, and it is the line most of the industry has stopped at, on the speech side and the text side alike, which is why every polished transcript still needs a human to explain what actually happened.

The asymmetry explains a lot of the frustration knowledge workers report with modern tools. They now have access to excellent answers and mediocre understanding, and the tools keep improving the part that was never the problem. The question, which is the part the human actually brings, still has to be fired into the machine at the right angle, and no faster model fixes a vague angle. The answer got free first; the question is still waiting for its infrastructure, and the teams building that infrastructure are betting on the correct side of the whole price shift rather than trying to out-type the machine.

If the machine can type everything, the valuable job is deciding what it types. That decision is the scarce complement: the input that the commodity cannot function without, and the one that becomes worth more precisely because the commodity is cheap. In a company, the scarce complement is not the person who produces the output; it is the person whose intent the output serves. That person decides the customer, the constraint, the deadline, and the definition of done, and no keyboard shortcut can fill the role. Watch any working agent system and the shape is the same: cheap labor at the bottom, expensive intention at the top.

The complement is priced by the market in ways a team can observe directly. Ask a team which person's absence would actually stop the work, and the answer is rarely the fastest typist. It is the person who knows which problem is real, which is why every agentic rollout inside a company raises the same organizational question: who signs the intention, not who clicks the button. GTC 2026 made the framing official, declaring that every company needs an agentic strategy and that this is the new computer. That turns every prompt into a work order, and every work order needs an owner with context, authority, and a name attached to the outcome.

So the money question for an AI-era team is not which model; it is who decides. The role is a new kind of ownership, and it has a naming problem: it is not the manager, because the manager reviews; it is not the individual contributor, because the contributor executes; it is the person who can state the outcome well enough that the machine's typing becomes recognizably theirs, and who accepts the consequences when the outcome is wrong. That role is the job market's new bottleneck, and the salary premium will follow it, which is exactly how every re-pricing before this one behaved in previous cycles.

For founders, the implication is allocation, not tooling. If the scarce complement is the person who decides, then the investment that compounds is in that person's time, context, and judgment, not in more seats or faster GPUs. The company that buys everyone a model and still lets vague intentions reach it will pay the commodity price for commodity output, while the company that invests in crisp intentions gets premium output at commodity cost. That is the only durable margin available in this cycle, and it is decided long before any code is generated, in the quality of the ask. Tools will keep getting cheaper; a sharp ask is the one asset nobody can undercut.

A working definition helps more than a slogan. Something is actually valuable if it passes two tests at once: a competent AI cannot yet produce it from the same inputs, and it changes an outcome that someone cares about. Typing fails the first test immediately, since the machine types better and faster. A meeting summary fails the second, because the meeting is over either way and nobody's decision was changed by the summary. What survives is the class of work we label intent, taste, judgment, and context, which is why those four words keep coming back in every credible analysis of what AI does to the value of work.

Four criteria make the definition operational. It is valuable if it requires context nobody wrote down. It is valuable if it makes a trade-off between two defensible options. It is valuable if it commits to a position that can be wrong. It is valuable if it takes responsibility for a result after the generation is over. Typing satisfies none of these; asking a sharper question satisfies all four, which is exactly the distribution Jensen Huang was describing when he said the valuable thing is not the answer but the question. The four criteria also double as a hiring rubric for the roles the machines cannot fill.

The self-test is quick, and it should hurt a little. First: of your last ten deliverables, what fraction could a stranger's AI have produced from the same meeting notes and inputs? Most people land between sixty and eighty percent, and the first honest answer is usually lower than their instinct. Second: what did you decide that the inputs did not say? If the answer is nothing, the deliverable was a commodity even if it took you four hours. Third: what did you reuse that you could have generated instead? The reuses you cling to out of habit are the hours the market is now pricing at zero, and they are the easiest hours to give away.

Answer the three questions once a quarter and the trend line tells you what a company is actually selling. Falling decision content with rising hours is the signature of a team quietly getting replaced despite working harder. Rising decision content with falling hours is the signature of a team that understood the repricing and started spending its time upstream. The definition is not about feeling valuable; it is about being able to name the fraction of your work that the market still has to pay a human to do, and then arranging the week around that fraction. The self-test takes ten minutes, twice a year, and it is cheap insurance against the most expensive mistake an organization can make this decade: mistaking activity for the bottleneck.

Between the free commodity and your scarce intent sits a rung that most of the stack has not built yet. NVIDIA built the ears and the throat: open transcription with Nemotron 3.5 Transcribe, offline pipelines with Riva and NIM microservices, and now Nemotron 3 VoiceChat, which can listen and talk in real time. But verbatim speech is not expression. The missing rung is the layer that hears you speak, understands what you meant, and produces text worth sending: no fillers, no restarts, no false starts, and the intent intact. That is the difference between being heard and being understood, and it is a different engineering problem entirely.

That layer is expression intelligence, and it is the terrain Voiskey works on. Voiskey is a system-level voice input layer for Mac and Windows: you speak, it understands the intent behind the words, strips the fillers and verbal missteps, and outputs clean, ready-to-send text, whether the destination is a formal Gmail reply or a Cursor prompt carrying file paths for a code change. It supports more than 100 languages, because intent does not localize the way dialects do, and the moat is not the microphone on your desk; the moat is the understanding. Capture is the commodity; expression is the product.

Privacy is the second half of the rung, because the fastest way to think out loud is only viable if the thinking stays yours. Voiskey processes audio locally by default, retains nothing on servers, and holds SOC 2 Type II attestation; for enterprises, Voiskey Team brings the whole pipeline onto the company's own GPU, matching the air-gapped posture that NVIDIA's NIM microservices are designed to enable. Local compute was the third road Jensen paved when he put a personal agent, RTX Spark, on every desktop, and it is what makes expression a private act instead of a cloud demo. The pipeline is not just faster when it is local; it is more honest.

With the rung in place, the workflow lines up end to end: you speak the first pass, the system understands the intent, and the polished message hands off to whatever agent you are directing, whether that agent answers an email, drafts a ticket, or edits a codebase. The wheat stays free, and the mill becomes your expression. NVIDIA makes the machine hear you; expression intelligence makes the world understand you. That is the practical difference between owning a commodity and owning the bottleneck, and it is the difference between a transcript and a message that does what you meant. A transcript tells someone what you said; a message tells them what you meant, and only one of them is worth sending.