Here's what's on our plate today:

• 🧪 Cognition hits $47B in sales, with engineering work priced in human hours.

• 📰 China readies humanoid robots for combat; OpenAI admits a hidden wiki incident; US tech firms eye Australian data centers.

• 💬 Prompt of the Day: price your team's last sprint in engineer-hours.

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

Everyone's watching the valuation. The change worth watching is that Cognition started selling engineering by the hour.

• The unit changed first: Autocomplete can't be priced in hours because the person still does the job. Devin finishes the whole task, so there's something to bill.

• Then it went into a contract: Cognition estimates the engineer-hours each session saved, converts them to dollars, and hands back up to $10M in credits when the value falls short.

• The valuation didn't reprice it: $47B on more than $900M is a multiple near 52, roughly what investors paid in May. They bought more revenue, not a new category.

• The bill lands on juniors: Finished code still has to be checked by a senior. The drafts that taught judgment are the ones that stopped being written.

• What's at stake: No buyer has accepted the hour as the unit, and Cognition has never published what an hour is worth.

Cognition is putting a price on the hours AI can replace

At NASA's Langley laboratory in the 1950s, the word 'computer' was a job title held by a person. Hundreds of women worked out flight trajectories by hand and passed the results to the engineers who signed them off. When IBM's mainframes arrived at the end of that decade, the expectation inside the building was that the job would end. What happened instead is that Dorothy Vaughan taught herself and her staff to program the machine, which turned a department paid to produce arithmetic into one paid to check it.

Something similar has been happening in software since ChatGPT arrived, and it has been widely misread as the beginning of the end for programmers. The demand for engineers has split rather than collapsed, since the repetitive work once handed to the newest hire is exactly what a machine can now do. One company has taken that split further than anyone else and begun selling engineering work the way a contractor sells it. Cognition measures what its software produces in the hours a human engineer would have needed, then pays the customer back when those hours fail to show up. That change in what is being sold is a bigger event than the $47B valuation that briefly made the company famous.

The company that sells the hours

Most people who have used AI to write code have used something like autocomplete, which suggests the rest of a line while a person types. Cognition, founded in 2023 by the competitive programmer Scott Wu, built something with a different shape and called it Devin. Somebody gives Devin the sort of task a manager would hand to a junior engineer on a Monday morning, and the software goes off to do it all. It reads the existing code, makes the change, runs the tests, fixes what it broke, and returns a finished bundle of work for a human to accept or reject.

Selling that in hours is possible only because the unit of work changed first. An autocomplete tool cannot be priced in hours, since the person is still doing the job and the tool cannot say which part of it belonged to the machine. An agent finishes the task, which lets the vendor point at a completed piece of work and argue about what it would have cost in salary. Cognition is the first company at scale to make that argument in a contract, and the demand it is arguing into is real.

A valuation that grew with the sales

Evidence of that demand first appeared in the company's revenue, which roughly doubled in three months. Annualized revenue, meaning recent sales projected forward a year, stood at $492M at the end of May 2026 and above $900M by the end of August. On September 1, 2026, Bloomberg reported that Cognition was raising around $1B at a valuation of about $47B, citing people close to the deal. The company has said nothing publicly since. A year earlier, the same business was worth $10.2B, so the mark has quadrupled in twelve months.

The number is easy to misread as the market declaring that machines now do engineering work. What it mostly records is a bigger version of the same business, and the change worth watching happened in a document Cognition published three months before anyone reported the round.

The contract is the real change

In June 2026, Cognition began measuring its product in units of human work, which software companies do not normally do. It built an estimator that reviews every finished Devin session, decides whether the session produced anything useful, and calculates how many hours an engineer would have needed to achieve the same result. The company chose hours because lines of code say nothing about effort, while hours are already how organizations price engineering.

The commercial promise followed the same day, turning that measurement into money. Under a contract term it calls the AI Productivity Guarantee, Cognition converts the estimated hours into dollars at a standard global rate, then compares the total against what the customer actually spent over the year. Where the value falls short, the company hands back credits worth up to $10M. A vendor is now writing checks against a promise of delivered labor, which is a different business from renting seats to people who may or may not log in.

Investors, however, do not appear to have paid extra for that difference. The cleanest way to see whether a market has reclassified a business is to look at the revenue multiple, which divides the valuation by projected annual revenue and shows what buyers will pay for a dollar of sales. Cognition raised $26B on $492M in May, a multiple near 53, and $47B on more than $900M, which works out to near 52. The money is paying for more revenue at the old rate, and some of it is paying for scarcity, since Cursor's parent sold itself to SpaceX in a deal that closed on August 15, 2026. That sale left Cognition as the last large coding-agent company that belonged to neither a model provider nor a rocket company.

The checking moves to the reviewer

If engineering work can be bought by the hour, the work left inside the customer's building is the part that cannot. That part is review, because software that arrives finished still has to be understood by someone who did not write it. Google's DORA program analyzed 1,110 written responses from its own engineers in March 2026 and found that time saved while generating code is frequently spent again on checking it. One engineer described writing less code and spending more time watching over what the machine produced.

That is the mechanism underneath the hiring split described at the start. A firm buying agent work needs fewer people to produce first drafts and at least as many who can tell a working change from a plausible one. Stanford's payroll research through June 2026 shows where the cost has landed, finding no economy-wide job losses from AI. It finds a 19% relative shortfall in employment among workers aged 22 to 25 in the most exposed occupations. The shortfall runs through companies hiring fewer juniors rather than firing anyone already employed, which means the bill is being paid by the people who would have learned judgment by writing the drafts.

The founder makes the smaller claim

The most restrained account of what Devin does comes from the man selling it. Scott Wu told TechCrunch in May 2026that replacing humans has never been the aim, describing Devin's work as unglamorous maintenance and the moving of old systems onto new platforms. He puts its ability somewhere between a junior and a mid-level engineer, depending on the task.

The independent evidence supports his description better than it supports the valuation. METR, a nonprofit that measures what AI systems can actually do, ran a controlled trial in early 2025. It found that experienced developers took about 19% longer when using AI tools, even though they believed they had been sped up. Its attempt to repeat the study failed in February 2026 because developers were quietly declining the tasks they did not want to attempt without AI. Its later results point to a real but modest speedup, far below what the marketing around these tools describes.

That gap is the strongest argument against the labor claim, and it does not dispose of it. Cognition has shown that a vendor will sell engineering work by the hour and stand behind the estimate with money. That is a change in the product itself, whatever the measured speedup eventually turns out to be. What it has not shown is that buyers have accepted the hour as the unit, and the company concedes that any single estimate can be wrong by a factor of two or three.

Vaughan's programmers kept their jobs because somebody still had to know whether the machine's answer was right, and that turned out to be worth more than the arithmetic it replaced. Cognition is offering its customers the reverse arrangement, supplying the arithmetic and leaving the checking with the buyer, billed in hours at a rate only one party has ever seen. Until a customer publishes an invoice or Cognition publishes that rate, $47B rests on a number nobody outside the building can check.

Prompt of the Day

Bite-Sized Brains

• PLA eyes robotic combatants: A Reuters review of more than 100 Chinese procurement notices, patents, and defense materials shows military institutions testing humanoids against battlefield requirements, with Chinese manufacturers already supplying about 95% of global humanoid shipments.

• OpenAI admits the wiki incident: OpenAI conceded that its agents had turned wiki sites into makeshift message boards and that the industry still has no standard for reporting misalignment during training, evaluation, and deployment.

• Australia courts the hyperscalers: AirTrunk says Google, Apple, Meta, Amazon, and Microsoft have all shown stronger-than-expected interest in Australian data centers, drawn by renewable power, political stability, and proximity to Asia as US backlash grows.

Tuesday Poll

The Toolkit

• Continue: Open-source AI code assistant for VS Code and JetBrains, with full control over models and context.

• Descript: AI audio and video editor that lets you cut recordings by editing the transcript.

• Drift: AI add-in for Excel that makes financial models context-aware and queryable for finance teams.

Rate This Edition