Jensen Huang Says Typing Is Cheap. Here's What Is Actually Expensive.

Cheap is a price, not a bill. When Jensen Huang said in his July conversation with LangChain's founder Harrison Chase that coding is essentially typing and typing is becoming a cheap commodity, he was describing one row of a ledger, the row for what it costs to produce words. The other rows did not change: the cost of reading those words, of checking their claims, of deciding which of them deserves to ship, of fixing the ones that silently got it wrong. Typing just collapsed to zero, and a ledger with one zero row still sums the same at the end of the month when the other rows are rising. This article prices those rows.

Before AI, the two halves were balanced by friction. Typing was slow enough that you thought hard before you typed, and the slowness was a kind of tax that kept bad ideas from getting far enough to be expensive. That tax is gone. Words arrive in seconds, fully formed, confident, readable, and the gate that used to be the keyboard has moved entirely to the other side of the production line, to the person who has to decide whether the words are right. Friction did not disappear; it relocated from input to review, and the new location bills at a higher rate.

Name the expensive rows in advance so it is a ledger rather than a surprise. Attention: every output consumes review time that typing used to consume. Judgment: someone still chooses what ships and what does not. Trust: a machine that produces fluent errors burns organizational trust faster than a person who makes honest mistakes. And the quietest row of all, iteration: each round-trip between a vague ask and a wrong result is cheaper per round than it used to be and more numerous, so the total spent on the loop has gone up for most teams, not down. Add one more row for completeness: the cost of the wrong brief, which is usually only visible after the deadline.

The framing matters because it changes your reaction to the quote. Read typing is cheap as a prophecy and you go looking for a keyboard to defend. Read it as a receipt, an invoice for the exact moment input became a rounding error, and the useful question appears: if typing is the cheapest thing on the table, what is now the most expensive? The rest of this article answers that with specific line items, starting with the one that quietly eats the most money in any organization that works with agents: the brief that is almost right, and the consequences that flow from it at full speed.

A brief that is ninety percent right is worse than a brief that is twenty percent right, and the reason is confidence. Twenty percent right means the author knows it is incomplete, so whoever receives it fills the gaps deliberately, flags them, asks questions. Ninety percent right means the missing ten percent is invisible, so the agent executes it with full confidence, the wrong assumption gets baked into the output, the output is polished enough to ship, and the error surfaces weeks later, in front of a customer, wearing the author's name. The ninety-percent brief is the most expensive artifact in the AI-era office because it converts a small omission into a large one at full speed.

Do the arithmetic on the failure mode. Round one: a ninety-percent brief, one minute to write. Round two: the agent produces a confident draft in thirty seconds, ninety percent of it exactly right. Round three: review, fifteen minutes for a person who reads closely, because the draft is convincing and the misplaced assumption hides inside a well-turned paragraph. Round four: the error is caught, or it is not. If caught, the task re-runs from the brief and the whole loop costs another hour of a person whose hour is not cheap. The typing was free; the consequence was not. That is what a cheap commodity feels like when it is the wrong commodity.

The ninety-percent failure has a signature worth memorizing: the output is not wrong in a way that looks wrong. A twenty-percent brief produces a draft that sounds incomplete, so the review starts suspicious. A ninety-percent brief produces a draft that sounds finished, so the review starts trusting, and a starting point of trust is where verification goes to die. In practice this means the most expensive briefs are the shortest ones, the two-line asks after a fast conversation, because the missing parts are exactly the parts that were on the table at the whiteboard and never made it onto the page. It is worth repeating: the brief was not short because the work was small; it was short because the deciding was skipped.

The ledger line usually looks like this: free typing, paid twice. Once in the review hour spent verifying what should have been specified, and once in the rework cycle that starts when the assumption surfaces. Teams with mature agent workflows do not hire faster typists; they tighten the brief, because each tightening removes one round-trip from the ledger's most expensive row. The strange economics of the AI era are that the cheapest input now causes the most expensive failure, and the most expensive input, the thirty-second constraint sentence, prevents it. Prevented, that is, if you remember to say it before you delegate, which is why capture precedes execution in every workflow worth running.

Attention was always the scarce input; it just used to be rationed by the keyboard. Now that the machine produces unlimited output, attention is the only input with a fixed supply per human, and markets price fixed supply in proportion to what it gates. Every AI draft sits behind a decision: accept, reject, or fix. The same eyeballs that used to read one draft a day now read forty, and each one has the polish of a final version, so the filtering has to happen at the level of meaning rather than the level of typos. Reviewing meaning is expensive, because it cannot be delegated to the same machine that produced the meaning.

Judgment re-prices for the same reason, and the CMU graduation speech is the clearest statement of it: AI will not replace you, but someone who uses AI better than you may. Better use is not faster typing; it is better choosing, because choosing is the part the machine still cannot buy for you. The person who knows which of forty plausible outputs is worth shipping now holds the only lever in the workflow that changes the outcome, and levers have always been worth more than pulleys. This is the honest version of the optimism: the commodity that got free is the one the machine could easily do, and the skills that go up in price are the ones your typing speed never competed with.

There is a third re-pricing, quieter in the coverage: the price of verdicts. Attention decides what to look at; judgment decides what to believe; the verdict decides what to act on, and it is the row where organizational cost actually lands. Teams discovered that giving everyone an agent without giving anyone a verification standard produces more output and more undelivered promises at the same time. The countermove is not less AI; it is explicit verdict rules, which ask the same question the ledger asks: who checks, against what, and what happens when the check fails? A verdict without a rule is just an expensive opinion.

Once the ledger is priced, the budget freed by free typing has an obvious destination: buy back the expensive rows instead of more typing. Write thirty-second briefs at the whiteboard in real time. Use the hour you saved to interview the customer whose answer you were about to assume. Hire the reviewer whose job is to say no. Treat the spec one level above the output as the actual deliverable. Teams that do this look at the same quote and see different news: typing is cheap, so spend the savings on the only thing that is still expensive, which is deciding what was worth typing in the first place.

The most expensive sentence in the modern office is no longer the wrong one; it is the soft one: you know what I mean, something reasonable there, the usual approach, make it better. Typed to a human, these are conversational gestures, filled in by a colleague who has context and patience. Typed to an agent, they are instructions to average, and the average is somebody else's version of your problem. The gap between what the sentence means and what it says is a tax, exacted once at the start and compounded at every iteration, because each round of rework pays the gap again plus the cost of the round.

Watch the tax land in a real afternoon. A vague ask about improving the onboarding email produces a draft, and the draft is fine, which is the problem: fine is a terrible signal because it is smooth enough to be plausible. So the asker writes a follow-up, clarify the tone, and the machine, reading tone as a general instruction, rearranges the words and keeps the wrong structure. A third round adds specific language and the draft becomes what it should have been in round one. One completed task, three rounds, forty minutes, and all of it caused by a first sentence that carried zero constraints. That is the vague tax, paid in the currency the ledger shows as expensive.

The tax has a mechanical origin: with agents as labor, vagueness does not stall, it executes. A human receiving a vague instruction will usually stop and ask within minutes, because humans price their own time. An agent has no such instinct; it treats the gap as scope and produces something, which means the vague input is converted into rework at near-zero marginal cost in output and near-maximum cost in review. The machine running happily on a bad explanation is precisely why the skill of pinning down what you mean before you send has moved from courtesy to the center of the job.

The fix is not more words; it is earlier specifics, and the specifics are cheap to produce if you harvest them at the source. The moment the thought forms is the moment the nuance is visible, which is why practice at the whiteboard beats practice at the keyboard: spoken intent still contains the details that typing smooths away. Expression intelligence, the layer that understands what you meant and strips the fillers that carried nothing, turns spoken intent into text you would send, no ums, no false starts, no you-know-what-I-mean hedging, just the commit. Speak it, strip it, commit it, and the tax stops at the border.

Intent has a half-life, and it is short. The version of your thinking that contains the real constraint, the real audience, the thing you actually mean rather than the thing you are willing to admit, exists fully for about the length of a meeting, and the transcription of thought into typed sentences damages it in two ways: delay and translation. By the time a typed brief exists, the intent has been filtered by effort, by the editor in your head, and by the boring parts it was easier to skip. Capture it earlier and it is cheaper, and the word capture is chosen deliberately: this is archaeology, not composition.

The cheapest capture surface is the one already moving at the speed of intent, which is speech. Talking through a problem is how most people actually find the wording of the question, which is why whiteboards still exist at decision meetings. The historical objection was that speech produces audio, and audio is not text you can send. That objection is exactly what expression intelligence removes: the output is not a transcript to fix, it is an understanding of what you meant, with the fillers and missteps already gone, ready to commit as an email, a document, or the first line of a prompt. That distinction, between a recording and an understanding, is the entire reason the capture surface can be speech.

Voiskey implements this as system-level voice input on Mac and Windows, supporting more than 100 languages: you speak, it understands intent, removes fillers and verbal missteps, and produces text you would actually send. The reason this matters on the cost ledger is that it deletes the transcription penalty, the part of dictation that made people abandon it, the part where you type anyway because the transcript is unreadable. And the pipeline stays private by design: audio is processed locally by default, with zero server retention and SOC 2 Type II attestation, so the intent you just captured does not become someone else's context before it is finished.

If you keep one line from Jensen Huang's remark, keep the ledger version: typing is cheap, therefore intent is the last expensive thing you own, and the way to protect it is to capture it at the speed it occurs, in the form it occurs, and commit it before it degrades. The free part should fund the expensive part, not replace it. Capture the intent while it is still cheap, strip the noise, commit the version you actually meant, and let the machine do what it was hired to do, which is everything after that sentence. The sentence is the spec; the spec is the asset; and the asset is the only thing on the ledger that appreciates.