Vercel CEO Guillermo Rauch: Software Output Next Year May Exceed Entire Computing History
Vercel CEO Guillermo Rauch stated that the software output next year is likely to exceed all that has been produced in the history of computing. Vercel took ten years to reach 1 billion deployments, and in the last ten months, it added another 1.4 billion. A large portion of this consists of one-time small software: HTML with JavaScript, reports, pricing tools, and slides, extending to applications, proxies, and platforms. Deployments have been compressed to about 1 second: uploading, binding domain names, and global propagation, while incorporating content delivery, firewalls, immutable versions, observability, and rollback.
The growth rate comes from proxies. By September 2025, the platform is expected to approach a cumulative 1 billion deployments; by mid-2026, over half of the deployments will be triggered by programming proxies, a significant increase from less than 3% six months prior, approximately a 17-fold increase. In February, deployments triggered by proxies rose from about 5% in June of the previous year to over 21%, with around 70% coming from Claude code. Customers using Claude make up only about 1% of users but contribute nearly 15% of deployments. AI gateway traffic increased from 20 trillion tokens per month to 200 trillion. There are over 6 million builds daily, completed in isolated micro virtual machines.
The company is repricing itself based on this usage. In September 2025, Series F raised $3 billion, post-money valuation around $9.3 billion; by February 2026, the annualized revenue under GAAP is approximately $340 million, an increase of about 86% year-on-year. Rauch mentioned that large models favor Vercel, and the infrastructure team is supporting automated Git pushes and command-line calls. Static small artifacts can skip builds, limited to about 10 files totaling 5 megabytes of HTML or Markdown.
The 1-second deployment targets artifacts, not arbitrary monolithic repositories. The platform continues to streamline build orchestration: end-to-end, it has previously reduced fixed overhead by up to about 7 seconds. Immutable deployments allow proxies to aggressively push preview addresses, while rollback and firewalls are used to handle "throwaway" traffic. The bottleneck in the software factory has shifted from writing code to whether publishing, domain names, and global caching can keep up with machine speed.
One-time software will shorten the average lifecycle. Many pages have no maintainers, generated only once and clicked once. The platform profits from build counts, preview bandwidth, and edge requests, not from each repository lasting ten years. Proxies deploy for proxies, transforming traffic from human submission rhythms to a continuous integration, unattended rhythm.
The market mechanism decouples deployment counts from human staffing. Buyers are proxies and ambient programming tools that turn natural language into clickable URLs; sellers are cloud platforms that must synchronize certificates, domain names, and edge nodes within one second. Beneficiaries are front-end clouds that charge based on builds and bandwidth; those under pressure are traditional PaaS that sell by seats and publish by the minute. Funding shifts from engineer hours to each Git push and each artifact propagation.
Source: Public Information
ABAB AI Insight
Rauch condenses the company's ten-year curve into a single capacity prediction. Humans took ten years to accumulate 1 billion releases, while machines added another 1.4 billion in ten months, indicating a change in submitters. Claude, with 1% of users, accounts for 15% of deployments, demonstrating toolchain lock-in: models repeatedly encountering Vercel commands in training and documentation default to pushing there. The valuation jumped from $3.25 billion to $9.3 billion, selling not the next.js framework, but the proxies' default launch wells.
The capital path transforms the edge network into the last mile of a software factory. Builds are sandboxed in micro virtual machines, with domain names and CDNs closing in seconds, while observability and rollback mitigate the accident costs of one-time pages. Revenue follows deployment counts and request counts, unrelated to "how long this application can last." Whoever achieves one-second artifact publishing can capture all overflow traffic from ambient programming tools.
Analogies should focus on GitHub and AWS Elastic Compute. GitHub turns submissions into default social actions; Lambda charges based on requests for execution. Vercel makes "having a globally clickable address" a default action. Heroku once positioned Git push as a product but failed due to enterprise complexity and pricing; this round of volume comes from machines, allowing unit prices to be diluted by massive one-time pages. The industry position shifts from front-end hosting to proxy infrastructure, with the phase characterized by capacity explosion and lifecycle contraction.
Structural judgments belong to technological substitution. What is being replaced is the software production function of "humans writing, humans clicking publish, humans monitoring." The mechanism is that as code generation costs approach zero, constraints shift to propagation and isolation; whoever can provide the same certificates and rollback for throwaway pages and serious applications in one second will capture nearly all new release volume. Historical total output is surpassed by machine submissions within a year, not as rhetoric, but as a change in the denominator.
ABAB News · Cognitive Laws
Once software can be submitted by machines, output is no longer tied to human hours.
Platforms selected by models will capture a disproportionate share of deployments.
Throwaway pages force publishing speed to become the infrastructure itself.