Venture Capitalist Marc Andreessen: Software Will Devour the World at the Speed of Computing Power
Venture capitalist Marc Andreessen, writing for a16z after investing in Cognition, stated that software is shifting from the "speed of human writing" to the "speed of computing power" to devour the world.
He revisited his 2011 article "Software is Eating the World": at that time, technology accounted for 31.5% of the top ten companies by market value globally, now it is 94.4%; back then, about one in three hundred people on Earth wrote software, and every line had to be handwritten. He noted that three years ago, when Scott Wu's team began working on Devin, "coding agents" were almost non-existent, but now leading company engineers say agents write nearly all the code. Over the past year, Devin's share of code production at Cognition has risen from 13% to over 90%. He believes that when agents write 90% of the code, engineers can do ten times the work, transitioning from manual coding to being technical leads of agent fleets; Scott Wu describes this as engineers evolving from bricklayers to architects.
Client-side numbers tell the same story: Mercedes-Benz reduced the estimated eight-month COBOL migration to eight days; Rivian increased testing generation speed tenfold; at financial institutions like Itaú, Devin diagnoses and patches vulnerabilities in thousands of codebases, with about 70% of security vulnerabilities being automatically fixed. Cognition completed a $1 billion Series D round in May this year, with a valuation of $26 billion and an annual recurring revenue of approximately $492 million, with enterprise usage showing high month-on-month growth for several months; clients include Citigroup, Goldman Sachs, Santander, Dell, and the U.S. Army and Navy. a16z noted that they invested in Wu's Lunchclub seven years ago; Wu is a three-time gold medalist in the International Olympiad in Informatics and a world champion at age 17.
The buy side is venture capital that turns coding agents into the next layer of infrastructure, while the sell side is corporate engineering departments handing over manual coding hours to agent fleets. Funds flow into Cognition from the equity round with soaring valuations, then sold to banks, automotive companies, and government warehouses through positions and task packages. Beneficiaries are agent companies that can productize legacy migration and vulnerability patching; under pressure are outsourced and manual coding positions still priced by person-months—the demand is written as infinite, and the price war falls on those who can replicate "eight months into eight days" for unstructured new code.
Source: Public Information
ABAB AI Insight
Andreessen connects his 2011 industry judgment to the founders he has invested in: Lunchclub is social matching, Devin sells the speed of the same batch of "competitive programmers" to corporate legacy systems. With 90% of internal code self-written by the product, it serves to prove that the flywheel has closed—agents improve products, and products accelerate agents. The $26 billion valuation in the Series D round buys this recursion, not just another completion plugin. The cases of Mercedes and Itaú are clear-cut and verifiable migrations and patches, not products designed from scratch.
The capital path is a combination of competition medals and early fund relationships, plus access to bank-level warehouses. Nubank was once used as a model to open up Latin American bank procurement, and Itaú has connected 75% of its tech team to agents. Revenue has increased from about $37 million to approximately $492 million annualized, indicating that companies are willing to pay subscriptions for "fewer person-months". Compilers, open source, and cloud have not eliminated engineering jobs; Andreessen defends the next leap with Friedman-style infinite demand: cheaper software will create more software demand. What remains unresolved is that automatic patches still have a high failure rate and the risk of introducing new vulnerabilities in independent evaluations, with 70% of automatic fixes still requiring human review.
In contrast to GitHub Copilot's completions, Cursor's editor agents, and outsourced factories priced by person-months: Cognition bets on task-level autonomy rather than typing suggestions. The industry is in a transitional phase from "engineers writing code" to "engineers scheduling agents"; job titles remain, but the unit of measurement shifts from lines of code to tasks and repository coverage.
This is technological substitution. The mechanism is: when generation and verification loops can provide mergeable patches on legacy code, person-month costs are broken down into computing power and review; the concentration of market value in technology provides the payers, and the assumption of infinite demand provides the narrative. Whoever masters the scheduling power of agents in the repository will redefine software engineering pricing; the premise of bricklayers becoming architects is that the bricks of the building still allow machines to be laid one by one.
ABAB News · The Law of Cognition
- The speed at which software devours the world depends on whether the code is handwritten or generated by computing power.
- Tools make software cheaper each time, and software demand often increases together.
- When agents write 90% of the code, the scarcity of engineers shifts from typing to scheduling and acceptance.