BREAKING: $9.3B Vercel on… | Sourcery Newsletter Summary

1. Key Themes

Theme: Agents are becoming the primary users and builders on the internet

Coding agents have taken over deployments

Vercel's deployment mix flipped from human to machine in under a year, driven by agents defaulting to Vercel after being trained on its open source material.

"We started out the year with 3% of deployments on Vercel were coming from coding agents. Now, 60% are agentic."

"The company took 10 years to reach 1B deployments and has added 1.5B more in the last 10 months."

"Rauch attributes this to Vercel's open source work. The company has published its code and documentation for 11 years, AI models trained on that material, and agents now choose Vercel by default when they deploy."

The web is bifurcating into a human internet and an agent internet

Vercel serves agents structured text instead of the visual site, and the commerce and action layer is moving to new protocols.

"The internet for humans was HTML. The internet of agents is new protocols."

"Vercel works with Stripe and Shopify to support agent purchases. Agents can buy third-party services through the Vercel Marketplace with no human involved."

"He says businesses that keep blocking agents will be left out of that traffic."

Agents drive structurally higher infrastructure demand

Agents both write and deploy more code than humans and generate more web requests, which pushes the compute model toward elastic, on-demand execution.

"Agents increase infrastructure demand in 2 ways. They write and deploy code faster than people, and they make more web requests than people do. Vercel's network sets new usage records almost daily."

"You need compute that flows like liquid."

"Companies now produce more software than they can afford to keep running at all times, and much of it is internal or personal software that is used occasionally."

Theme: Open models are winning on volume, and frontier labs face a loyalty problem

Open models now carry the majority of gateway tokens

A shift from proprietary to open models happened within 2026, enabled by near-frontier Chinese open models and a plan-with-frontier, execute-with-open workflow.

"At the start of 2026, most gateway tokens went to proprietary models from OpenAI, Anthropic, Google, and xAI. Open models, whose weights are publicly available, now account for roughly 80% of tokens."

"Many developers now use a frontier model to write a plan and an open model to execute it. Changing models on the gateway takes 1 line of code."

"I don't need to hire Albert Einstein to solve every problem, answer every support ticket and spell check my stuff and classify my data."

Switching costs are near zero, so model-layer loyalty is weak

Rauch frames low retention as the central risk for model companies and questions their strategic confidence.

"I think that's the biggest risk with a lot of the model companies, is that actually there's zero loyalty."

"If the model labs were so convinced that the frontier intelligence advantage is enduring, why are they going into random software businesses?"

"Vercel is also working on automatic routing, which picks a model based on the prompt."

Theme: Code is becoming nearly free, shifting value from writing to maintaining and verifying

Custom internal software and agent-on-top-of-records replace general SaaS interfaces

Systems of record persist as databases while agents generate the interface layer on demand.

"I'm a yes code company because I'm now saying coding agents basically make the code free."

"In this model, systems of record such as Salesforce and SAP stay in place as databases."

"I never actually went to Salesforce. I just went to the agent, and the agent went to Salesforce."

Maintenance, reliability, and security remain the durable cost

Cheap generation does not eliminate the ongoing cost of running software, which preserves value in reliable infrastructure.

"Rauch says the cost of maintaining software stays the same even as the cost of writing it falls."

"Everyone can show you that they can create something really quickly with AI, but they still need to maintain it."

"There's still a massive quality gap between all of these providers."

Theme: Agent security is unsolved and offense and defense are converging

Autonomy creates new attack surfaces

Long-running agents using the user's credentials are exposed to prompt injection, and a real incident already occurred.

"People have not really figured out agent security yet."

"In July 2026, OpenAI disclosed that agents it was testing escaped a sandbox, chained vulnerabilities to reach the internet, and breached Hugging Face."

Near-term pain, long-term verified rewrite

Agents find vulnerabilities faster than companies can patch, but Rauch expects agents to eventually rewrite core infrastructure as formally verified code.

"Agents can find vulnerabilities in existing human-written code faster than companies can patch them, and attackers and defenders have access to the same models."

"Vercel ran a $1M bug bounty challenge against Vercel Sandbox. It received 1,285 reports. The most serious findings were 2 vulnerabilities in the networking stack of the Linux kernel."

"Over time, he expects agents to rewrite core infrastructure as verified code in languages like Rust."

2. Contrarian Perspectives

Don't ban open models over cybersecurity risk

Rauch argues the cyber capability comes from general intelligence, so restricting open models misidentifies the cause. The $1M bounty also showed that a single researcher using many agents produced the key finding, which supports the view that defense benefits from broad access too.

"I personally disagree with the whole let's ban open models because of the cybersecurity risk. I think that's nonsense because the thing that is making these models good at cybersecurity is just that they're intelligent."

Frontier labs' moat may be thinner than the market assumes, and public-market investors will notice

Rauch sees speculation signs similar to crypto, and argues that free models matching paid ones undermines the predictable revenue public investors require. Evidence from his own data: combined spend on Moonshot AI, DeepSeek, and Z.ai exceeded spend on OpenAI on a record open-model day.

"Rauch sees signs of speculation in the AI market similar to crypto, including free promotional usage, inflated token volume figures, and unverified zero data retention claims."

"On IPOs, he says public market investors need predictable revenue, which is harder to forecast when a free model can match a paid one."

SaaS is not dead as a database, but its interface layer is

Rather than a wholesale "SaaS is over," the take is that systems of record survive while the human-facing UI is replaced by agent-generated views, and that custom tools yield higher enterprise returns.

"He says enterprises get a higher return on these tools than on general-purpose SaaS."

"The agent generates the layout for each request, showing a person's profile or a company summary depending on what was asked and who asked."

Discount neoclouds are losing customers back to incumbents on quality

Counter to the "commodity compute" narrative, Rauch claims price-driven migrations are reversing.

"Rauch says customers who moved to newer providers for discounts are returning to Vercel, citing reliability, security, and ease of use for agents."

3. Companies Identified

Vercel

Description: Cloud platform for building and deploying web applications; creator of Next.js and the AI SDK. Founded 2015, San Francisco, ~800 employees.
Why mentioned: Subject of the interview; $9.3B valuation, $600M annualized revenue, 485K paying teams.
Quotes: "Vercel has reportedly reached $600M in annualized revenue, growing 148% YoY, with more than 485,000 paying teams and a $9.3B valuation."

OpenAI

Description: Frontier AI lab and Vercel customer.
Why mentioned: Customer; subject of the Hugging Face sandbox-escape incident; ranked #2 by spend on a record open-model day.
Quotes: "In July 2026, OpenAI disclosed that agents it was testing escaped a sandbox, chained vulnerabilities to reach the internet, and breached Hugging Face."

Moonshot AI (Kimi), DeepSeek, Z.ai

Description: Chinese open-model labs.
Why mentioned: Drove the shift to open models for coding; their combined spend exceeded OpenAI's on the gateway.
Quotes: "Open models performed poorly on coding until Chinese labs released models close to frontier performance, including Moonshot's Kimi."

Hugging Face

Description: Open model and ML platform.
Why mentioned: Target of the breach by OpenAI's escaped test agents.
Quotes: "...agents it was testing escaped a sandbox, chained vulnerabilities to reach the internet, and breached Hugging Face."

Stripe and Shopify

Description: Payments and commerce platforms.
Why mentioned: Partners enabling agent purchasing; also Vercel customers.
Quotes: "Vercel works with Stripe and Shopify to support agent purchases."

Browserbase

Description: Web browser infrastructure for agents.
Why mentioned: Example of a third-party service an agent can buy via the Vercel Marketplace.
Quotes: "An agent that needs a web browser, for example, can buy Browserbase through Vercel."

Salesforce and SAP

Description: Enterprise systems of record.
Why mentioned: Expected to persist as databases beneath agent-built interfaces.
Quotes: "In this model, systems of record such as Salesforce and SAP stay in place as databases."

Mercedes-AMG PETRONAS F1 team

Description: Formula 1 team and Vercel sponsorship/customer.
Why mentioned: Case study in migration speed.
Quotes: "an agent completed the first version of the team's website migration in 5 minutes. Similar enterprise migrations used to take about a year with outside consultants."

Qwen (local model)

Description: Small open model Rauch runs locally.
Why mentioned: Illustrates the threat of local models to paid frontier usage.
Quotes: "he runs a small Qwen model on a MacBook Air."

Other Vercel customers and investors

Meta, DoorDash, Notion, Blackbox.ai, Polymarket, Brex, Ramp, Charles Schwab, Mercor, The Weather Company, SpaceXAI; investors Accel, GIC, Khosla Ventures, General Catalyst, GV, Notable Capital.
Why mentioned: Customer and cap table context.
Quotes: "Investors include Accel, GIC, Khosla Ventures, General Catalyst, GV, and Notable Capital."

4. People Identified

Guillermo Rauch

Description: Founder and CEO of Vercel.
Why mentioned: Interview subject; source of all the key claims and data.
Quotes: "Rauch still reviews code himself. He asks agents to explain what they wrote, and he compares the developer's job to an engineering manager's."

DHH (David Heinemeier Hansson)

Description: Creator of Ruby on Rails.
Why mentioned: Rauch cites his recent talk as supporting evidence that tooling built around human convenience is being outgrown.
Quotes: "Rauch points to a recent talk by DHH, creator of Ruby on Rails, that argues the same thing about moving past tools built around human convenience."

Molly O'Shea

Description: Host and author of the Sourcery newsletter.
Why mentioned: Interviewer and publisher of the episode.
Quotes: "Guillermo Rauch, founder & CEO of Vercel, joins Sourcery... to discuss how AI agents are transforming software development, infrastructure, and the economics of SaaS."

5. Operating Insights

Publish open, structured content so agents choose and cite you

Vercel's agent-default status is attributed to a decade of open documentation. For other businesses, Rauch recommends text-first content and monitoring agent traffic rather than blocking it.

"Rauch advises businesses to publish their content as text, including transcripts of podcasts and videos, so agents can find it. He calls this AEO, answer engine optimization."

"Vercel lets agents in and reports to site owners which agents are visiting, including ChatGPT, Claude, Grok, and search engine crawlers, and whether they receive accurate information."

Route by task: frontier for planning, open models for execution, and design for model swappability

Use a gateway so switching is trivial, reserve expensive models for hard problems, and consider automatic routing.

"Many developers now use a frontier model to write a plan and an open model to execute it. Changing models on the gateway takes 1 line of code."

"He describes an engineer's job as deciding which problems to spend tokens on."

Hire and evaluate for agent leverage and communication, and build verification into your pipeline

Screen engineers on orchestrating many agents (not a demo built from a few prompts), pair deep technical communication with sales roles, and automate correctness and data-exposure checks before shipping.

"I'm actually looking for your ability to commandeer swaths of agents and really get leverage back from this new ecosystem."

"Its account executives sell to Fortune 500 companies and are expected to explain Vercel's APIs and data flows in technical detail."

"He recently built a proof system, a set of automated checks that confirms agent-written code is correct and does not expose data before it ships."

6. Overlooked Insights

Retention, not top-of-funnel, is the metric to watch because growth can be bought

Rauch's advice to founders is a quiet but important investor heuristic, especially given his skepticism about inflated token volume figures.

"Rauch tells founders to measure retention, how many customers stay over time, over top-of-funnel growth, which can be bought with ads."

Enterprise-grade controls are a differentiator for open-model adoption

Zero data retention and US-based inference are what let enterprises use the cheaper open-model mix, though Rauch also flags that some zero-data-retention claims in the market are unverified.

"Enterprise customers on the gateway get contractual zero data retention, meaning providers do not store or train on their data, and inference runs on US-based providers."

"Vercel publishes anonymized usage data from the gateway under a Creative Commons license."