NVIDIA's Jensen Huang Says Typing Is No Longer Worth Anything — So What Is?

Jensen Huang's one-line verdict, that coding is simply typing and that typing is becoming a cheap commodity, spread through the tech world because of its bluntness. Read literally, it sounds like an attack on keyboards, and much of the commentary took it exactly that way. But the sentence is an economic observation, not an ergonomic one. When software can mass-produce keystrokes on demand, the marginal cost of producing text falls toward zero, and anything that approaches zero stops being where value lives. His follow-up is the real headline: what is genuinely valuable is not the answer, but what questions you can ask. Value does not vanish when a skill gets commoditized. It relocates to the scarce input the machine cannot supply, which is knowing what you want before you tell it to produce anything at all.

The scarcity lens is why the claim has nothing to do with your fingers. Typing has been getting cheaper for years, through autocorrect, templates, and touchscreens. What changed is that the execution layer is now automated end to end. In his July conversation with LangChain founder Harrison Chase, Huang framed the same point directly: engineers will type less and increasingly build systems and direct AI. Executing via keyboard is being mass-produced by the machines, so what stays scarce is knowing what you want. The keystrokes were never the value itself; they were the toll gate that made the value slow to arrive. Machines now wave you through that gate, which is good news for speed and bad news for anyone whose edge was how quickly they typed.

For the individual, this creates an uncomfortable asymmetry. The value you used to create by typing is being duplicated by a model in seconds, while the value you create by deciding, discriminating, and framing is as scarce as it was a decade ago. That is not a judgment about the quality of your work; it is a statement about the clearing price of output. If everyone can produce a competent draft, the draft stops being the deliverable. The deliverable becomes the judgment that produced it, the reasons this draft is the right draft, and the risks the writer consciously accepted by choosing it.

So the useful reaction is not to mourn the keyboard or to boast about words per minute. It is to audit your own workday and ask, for each thing you produce, whether a stranger's AI could produce it just as well from the same inputs. The parts that pass that test are being repriced downward, and the parts that fail it are the ones worth more of your time. The rest of this article runs that audit in practice, then offers a small, concrete way to close the gap between saying what you want and sending something that says it exactly right.

Take a normal knowledge-worker day and list what actually leaves your desk: quick replies in chat, a meeting summary, a status update, the first draft of a document, and a handful of emails. Now price each one as if an AI could produce it. Chat replies are effectively free, since a model can match your tone and length in under a second. Meeting summaries are free too, and transcription has improved so far that capturing words is no longer the bottleneck. Status updates are free, because they are patterns over data that is already written down somewhere. That is four of the five categories already priced near zero, and none of them was ever the hard part of the job.

First drafts are where the price drop feels most personal. A document that used to consume an afternoon of typing now arrives as several options within a minute, and the author's work shifts to choosing among them, merging two, and cutting the weak third one. Emails that follow a known shape, scheduling, follow-ups, acknowledgments, were never valuable as typing; they were valuable only as proof that the conversation happened, and conversation has stopped needing a paper trail made by hand. Even code belongs on this list: boilerplate, glue, and the search-paste loop were the first to fall, because they are the most repetitive fraction of the work.

The audit is not an exercise in pessimism; it is a map of where your marginal hour buys more. If typing is no longer the expensive part, then the expensive part is everything upstream of the keystrokes: what the transcript actually meant, which of the five drafts fits this customer's situation, what the status update should leave out, and what the document is really trying to change. Those decisions cost exactly what they cost five years ago, and they are the reason your output is not interchangeable with the model's, no matter how similar the wording looks. The audit typically has a practical payoff too: three or four hours of every day land in the cheap pile and can be handed off immediately.

Once the five categories are repriced, the day rebalances itself. The sixty minutes you used to spend turning a thought into formatted text can be spent turning a half-formed intention into a crisp one. That is the mechanical meaning of the claim that what matters is the question rather than the answer: a good question is the compression of everything you know that the model does not, and writing it clearly is the most productive minute in any workflow. The work does not get easier; it simply moves to the part where your value was always hiding, and where nobody can outsource you to a cheaper typist.

Run the audit and four things survive intact. Intent survives, because the machine is excellent at producing a reasonable answer to whatever you asked and genuinely poor at producing the answer to what you should have asked instead. Taste survives, because taste is the ability to see that option three is the one your team would actually accept, and there is no shortcut for having read enough good examples to recognize it on sight. Judgment survives, because every real document is a set of trade-offs: what to include, what to omit, what to assert, and what to soften until the next round. None of those three can be bought as a service, and none of them becomes cheaper when the models get faster.

Context survives, and it is the big one. Your last three documents depend on facts that were never written down inside them: the meeting nobody recorded, the promise you made on a call, the constraint you were told but cannot put in an email. A model can write all four versions of a paragraph in parallel, but it cannot know that the customer's read on the deadline matters more than the finance team's, because that information never left your head, and no foundation model was trained on it. Context is the difference between a document that is correct and one that is right, which is also why editing upstream beats prompting downstream.

Context is exactly why a stranger's AI cannot reproduce your work even when it can reproduce your words. The gap between your draft and the model's draft is not vocabulary; it is everything the model was never told. That gap is also why the follow-up insight holds up so well: the scarce skill is asking well, because asking well is how you move context out of your head and into the prompt where it can actually change the output. A second, sharper question often replaces an afternoon of corrective edits, and that exchange rate is now the most valuable trade available in the day.

This reframes the discussion into something actionable rather than philosophical. The valuable work is not the performance of typing; it is the performance of deciding what the machine should type and of being able to say that decision out loud. People who name what they want clearly and early now move first, and when they move, they drag the team's default answer along with them, since every later draft gets written against their framing. Under the new pricing, clarity compounds faster than effort: one well-posed question at the start of a task is worth several rounds of polished output at the end.

Here is a test worth an hour, and it is the most honest instrument available for finding the part of your job that is still yours. Take your last three real documents: the email that mattered, the plan that got approved, the note you were proud of. Give a general assistant the same instructions and source material you worked from, and let it draft an answer on the same inputs. Then diff the two outputs side by side. The differences are rarely typos or style; they are the places where your document made a decision that the inputs did not contain, which is exactly where the thin line between duplication and expertise runs.

In most runs, the diff shows a short and repeatable list. Your version knows who the reader is and what that reader is worried about. Your version ranks the possibilities where the AI lists them, since a list is agnostic and a ranking is a bet. Your version contains a decision where the AI contains a summary of the debate. Your version carries a date, a name, or a number that lived in your head rather than in the brief. Those few passages are what you are paid for, and they have nothing whatsoever to do with typing, which is why they survive an audit that prices every keystroke at zero.

The test also shows the uncomfortable half of the truth: a large share of most documents is exactly what a competent AI would have produced from the same inputs. Admitting that is not surrender; it is accurate accounting, and the accounting is what lets you act. The generic share should be delegated, because hoarding it costs you the hours you need for the share that is genuinely yours. The people who outperform this repricing are simply the ones who delegate early and spend the reclaimed time on intent, on framing, and on the handful of sentences that decide the outcome. Competence at delegation is the new form of professional confidence.

Run this test once and the phrase 'typing is a commodity' stops being abstract. It becomes a personal inventory of what in your workday is duplication and what is decision. Most people are surprised by how much they find in the first pile and relieved at how clearly the second pile shows up the moment the first one is gone. From that inventory, the next step is not bigger software; it is reassigning your hours until the ratio matches the market price of each kind of output, which is the only durable way to keep a career that has already stopped paying for the typing.

If the scarce thing is knowing what you want, then the interface that helps you find out is the one that gets your thinking out fast and cleanly. That is the premise behind Voiskey, an expression-intelligence layer for Mac and Windows that turns speaking into text you can actually send. You speak; Voiskey understands intent, strips out fillers and verbal missteps, and hands back ready-to-send text, from a formal Gmail reply to a Cursor prompt with file paths included. The idea is to turn your typing into your asking: you stop being the typist and start being the person who says what the machine should do.

The name for this layer is expression intelligence, and it marks the difference between a verbatim transcript and a message that does what you meant. Transcription has become a commodity of its own. NVIDIA's open-source Nemotron 3.5 Transcribe handles 40 languages, diarizes up to eight speakers, and consumes a full hour of audio, supported by the open Granary 2 dataset; with Riva and NIM microservices it even runs fully offline on one GPU. Those are the ears and the throat, and they are effectively free. The layer between hearing and understanding what you wanted to say is still empty, which is exactly where Voiskey works.

The workflow is speak, polish, type, edit, and it matches the repriced workday exactly. You speak the messy first pass; the polish step surfaces the actual intent behind it; the cleaned text is typed straight into your application; and the edit step applies your judgment before anything leaves the building. The whole loop stays local by default: audio is processed on your machine, nothing is retained on servers, and the product holds SOC 2 Type II attestation. For teams, the same pipeline can run on the company's own GPU, mirroring the air-gapped posture that NVIDIA's Riva and NIM promise for speech infrastructure, so the fastest way of working does not have to be the least private way.

This is why the practical answer to 'so what is valuable?' now has an interface attached to it. The machines made typing cheap, and the people who benefit are the ones who can spend the reclaimed minutes discovering what they actually wanted to say. Turning your typing into your asking is not a slogan; it is the smallest upgrade available to a knowledge worker whose answers are already free and whose questions have finally become the job. The keyboard will still be there for corrections, but it no longer gets to be the bottleneck at the start of the work. The long journey from intention to output now has a faster first mile, and knowing what you want is the only toll that still matters.