Stanford MS&E435 — Economics of the AI Supercycle, “Applications, Coding AI” — Annotated Lecture Notes & Key Takeaways

Course Information

Institution: Stanford

Original Course: Stanford MS&E435 — Economics of the AI Supercycle, “Applications, Coding AI”

Instructor Bio: This session is led by **Apoorv Agrawal**, Adjunct Lecturer in Management Science and Engineering at Stanford University and Partner at Altimeter Capital, with guest speaker **Guillermo Rauch** (Founder & CEO, Vercel). Guillermo Rauch is a leading figure in developer tools and AI-powered software development. As founder of Vercel, he has driven the commercialization of AI-assisted coding tools and reshaped how modern software teams build and deploy applications.

Course Description: This lecture focuses on the economics of AI for software development. It quantifies the productivity gains of code generation models for engineering teams, examines business models and pricing strategies for AI developer tools, and explores how coding AI is reshaping software R&D cost structures, product iteration cycles, and the broader developer economy.

Before the slides: how the guest's backstory sets up the economics

 

The session opens with the host introducing Guillermo Rauch, the founder of Vercel and creator of Next.js, the widely used React framework. Vercel is described as a major developer-infrastructure company, valued at 9.3 billion. Rather than dwell on that, the host spends a few minutes on a part of Rauch's biography that even he, the host, had not fully known — and it turns out to be more than color commentary. It frames the central economic argument of the talk.
 

Rauch grew up in a suburb of Buenos Aires, taught himself to code as a child, and picked up English by reading software manuals — there simply were not Spanish ones available. He was doing remote JavaScript contracting by age eleven, later left high school, and moved to San Francisco to chase the startup world. The detail that lands hardest is the visa story: needing an O-1, he literally wrote a book — Smashing Node.js, published by Wiley in 2012 — as part of making the case to get himself to the United States. His rueful summary is that dropping out of high school is, it turns out, unusually unhelpful when applying for a U.S. visa, and that he did a great many things the hard way.
 

I found this introduction genuinely useful, not just charming. The through-line is access: a kid who could not get technical information in his own language, who could not legally or easily enter the industry's home country, bet his career on tools that lower those same barriers for everyone else. When Rauch later says open source was an easy decision, or that he wants his technology accessible to as many people as possible, it reads as lived experience rather than a slogan. That is a point worth keeping in mind, because the rest of the lecture is essentially an economics argument built on top of that instinct.
 

From a painful deploy to a business on top of open source

 

Rauch keeps the prepared remarks deliberately short so the class can get to conversation, but the first chunk of his presentation contains the seed of everything after it. He describes Vercel as a broad ecosystem of tools — heavily open source — that has shaped a meaningful share of the modern web. His go-to illustration is deliberately mundane: order a Big Mac or a Porsche and, somewhere in that flow, you are likely touching Vercel. Increasingly, if you interact with systems like Open Evidence or Grok, you are also dealing with agents hosted on Vercel. The underlying point is reach: even people who never deliberately chose Vercel are running on it.
 

The bet on open source, he explains, came directly from his Argentine upbringing — wanting free tools and free information as a teenager. What is more interesting economically is the second move: having made the tooling free and widely usable, Vercel built a substantial business on top of it. He names Databricks as a comparable case the class had recently heard from. The model is to let the open-source project become the default, then charge for the operational layer that lets you deploy, secure, and scale what you build — what he groups under the umbrella of pages and agents.
 

His origin story for the company is the classic frustrated-developer epiphany. When he sat down to launch his previous startup's website using then-modern React and Kubernetes, it took weeks. AWS, Google Cloud, and Azure existed, but they were painful — even for someone who had been studying the craft for decades. So he picked a large, addressable group (JavaScript developers) and gave them a single superpower: if you can build a front-end project, you should be able to deploy it at planetary scale, fast everywhere, without thinking about load balancing or low-level infrastructure. That narrow wedge, he says, grew into today's Vercel.
 

One line from this section is worth highlighting because it becomes a thesis: writing code does not make you special; deploying it in front of a customer does. He points out that the history of hosting — GitHub, before it SourceForge, and further back a project called Freshmeat — got excited about storing code, with the result that vast piles of software in repositories do not actually run. The learning, he argues, begins only when a user confronts the running version of your code.
 

That is also where coding agents fit neatly into Vercel's pre-existing strategy. Humans, he says, suffer from a bias: the code works on my machine, so I keep it on my machine where it feels safe. Agents have no such attachment. They love to deploy. So Vercel becomes the connective layer for the software that agents write — the peanut-butter-and-jelly metaphor, where the coding agent is the peanut butter smeared across the world and the infrastructure is the jelly. It is a memorable image, but as we will see later, the exact mechanism behind that dominance turns out to be more complicated than the metaphor suggests.
 

The cloud is being rewritten: agents, tokens, and the agentic-infrastructure triangle at the heart of the AI supercycle

 

The economic heart of the talk is the claim that AI has produced the single largest expansion in the total addressable market for creating software in our lifetime. Rauch's old math was roughly twenty million developers worldwide. Go further back and programmers were a tiny group with access to university mainframes; the whole history of programming is a history of broadening access. Coding bootcamps briefly made it plausible to learn React in three months and get a job. AI and coding agents, he argues, multiply the number of people who can create software by ten, twenty, perhaps a hundred times.
 

Concretely, since around October of the previous year — and especially after the arrival of a particular coding model, Opus 4.5 — Vercel saw deployment volumes climb sharply. The timing claim is specific and is clearly meant to tie model capability to platform demand, which is a useful causal test to remember: better coding models should show up first as more deployments, not just more chat traffic.
 

What changes when the unit you ship is an agent

 

He walks through several ways the cloud itself is changing, and most of them are best understood as infrastructure metaphors carried over from the first internet.
 

Duration changes.​ The old internet was built on instantaneous request-response; Amazon's own data, he notes, showed that each additional hundred milliseconds of slowdown cost roughly one percent of conversion. Agentic work breaks that pattern. It started at multiple seconds, then minutes, then hours. Today some agents cook for an entire day before delivering a report, an analysis, or a piece of software. Platforms hosted on Vercel are already deploying agents that maintain, advertise, grow, and scale entire companies behind the scenes. The shift is from pages toward agents.
 

Compute changes.​ Elastic compute — the famous EC2 model of putting in a credit card and getting an instance — was designed for human-written code. Economically, the amount of compute you could sell was bounded by the number of programmers. If the number of software creators expands a hundredfold, and agents are also repairing and securing software, demand for compute rises with it. He also notes the cybersecurity flip side: smart models mean both hacking agents and defending agents, the bad-cop-and-good-cop dynamic.
 

The unit of value changes: tokens.​ Vercel used to stream pixels, the UI of an application. Now it also streams intelligence in the form of tokens. That, Rauch says, is killing seat-based SaaS pricing and giving rise to token-based pricing — you pay for intelligence consumed. He is careful not to over-rotate, though. He remains bullish on human-centric, brand-rich experiences, predicting a return to a more whimsical, Encarta-like web — immersive 3D, video generation — rather than today's wall-of-text Wikipedia aesthetic.
 

Out of all this he defines agentic infrastructure as a triangle with three sides:
 

1. Infrastructure for coding agents​ — the deployment target for Claude Code, Codex, v0, and the like.
2. Tools to build your own agents​ — the AI-native school of the future, for example, still involves teachers and students but ships an agent rather than a set of linked web pages.
3. Everything automated by agents​ — software maintained and optimized by software.

 

He supports the third leg with a customer example that is striking but hard to independently verify: Vercel's own support agent now handles 93% of user inquiries and, he says, measurements show customers are happy talking to it versus a human. Whether that 93% is volume or satisfaction is not fully disambiguated in the transcript, which is the kind of detail I would want clarified before treating it as evidence.
 

Why agents reuse software instead of reinventing the universe

 

A listener might reasonably ask: if an agent is so capable, why does it need existing tools at all? In theory, to produce an apple it could reinvent the universe — write its own Linux kernel, networking stack, even macOS. Rauch's answer is an efficiency bias: there is economic pressure to reuse. This is where open infrastructure becomes a target the agent can build on — the LEGO-block analogy, or what Mitchell Hashimoto (HashiCorp founder, recently joined Vercel's board) calls the block economy.
 

The evidence Rauch cites is provocative: a report rating Vercel's UI engine, shadcn, at 90.1% near-monopoly status, and claiming Vercel has a similarly dominant share when deploying React or Next.js. Near monopoly is, of course, a useful hedge for any regulator listening. His reasoning is that models absorbed vast amounts of internet knowledge about these open-source projects during training, so when given a task like building a SaaS app, the agent is already biased toward the familiar, well-documented stack. This is a crucial and somewhat unsettling argument: incumbent tooling may get locked in not by users choosing it, but by models having learned it.
 

The metaphor tour continues with products already shipping:

• AI Gateway​ is a CDN for tokens — the tokens from Claude, Anthropic, Gemini need to be observed, failed over, secured, accelerated, and cached, plus load-balanced. A smart CDN can route a trivial expression of thanks to a smaller model rather than awakening a herd of GPUs for a polite reply, and can serve semantically cached responses.
• Sandbox​ is the EC2, or fundamental compute unit, of agents. Give a model a computer and you extract more capability; in post-training, models are handed Docker containers to learn in. It is not unlike giving a new knowledge worker a laptop with software pre-installed. Of course, just as personal computing brought viruses and phishing, agents can exfiltrate or leak data — hence a coming category of sandbox-security products.
• Self-driving cloud.​ Running software at scale is, in his telling, a pager-driven horror show; he borrows the Stripe-founder anecdote about ducks as a pager ringtone spiking cortisol. The vision is a cloud that configures and optimizes itself, then reports back: I made your software twice as fast, here is a PR — and I measured that it improved your customers' conversion.

 

He closes the prepared portion with customer examples — Meta's infrastructure teams and its superintelligence labs using Vercel to move faster and vibe-code internal model-training tools; Notion building its chapter-two agentic capabilities on it — and summarizes the positioning as the AWS of AI or agents, a full-stack cloud from developer tools to infrastructure.
 

What the Q&A actually tests — and where the argument gets weaker

 

The question-and-answer section is where the lecture moves from vision to specifics, and it is also where a listener should slow down. Several of Rauch's strongest claims are pressed here, with mixed results.
 

Software is dying — yes and no

 

The first question challenges him directly: Vercel is the machete clearing the forest of software people used to procure. What public-company products have been vibe-coded away? Rauch reframes rather than names a body count. The odd thing about SaaS, he says, is that smart people design a lowest-common-denominator interface meant to make the largest number of customers happy — translate a few labels, sell it again. That is not tailored software.
 

His examples: a CEO replaces the company's parking-lot management system with something live-coded in v0 and saves a fortune; inside Vercel, a team of about two people effectively rebuilt the presentation layer over their Salesforce data — accounts, opportunities, business intelligence for sales reps — while keeping the underlying database and workflows intact. So the picture is not that all software dies. The system of record — the database, access control, ACL layer — is reused. What becomes plastic and replaceable is the presentation and integration layer, especially for niche or internal needs no Palo Alto team was ever going to prioritize.
 

This is, I think, the single most important nuance in the whole talk. Companies that expose proper agentic interfaces — MCP, CLIs, APIs — fit naturally into this world and are not killed by the coding agent. That is a concrete, testable prediction: watch which incumbents open up programmable surfaces and which hide behind enterprise sales processes.
 

Retention, throwaway software, and a new audience

 

The next challenge is sharper: if vibe coding produces such custom software, do people still need it on Wednesday morning, or was it just a Saturday distraction? Rauch agrees he is drinking from the firehouse. Some software is genuinely throwaway — built for one customer call, discarded three calls later — and that is fine, because it is now essentially free. He credits the Shopify CEO's phrase, the reflexivity of AI: once you know you are one prompt from a high-fidelity prototype or demo in front of another human, you will never give that up. It is hard to forego efficiency, as he once titled an essay.
 

At the same time, he pushes back on the all-disposable narrative. A great deal of infrastructure engineering remains genuinely difficult; sometimes his team needs a quorum of three agents plus smart humans to stare at a single line of code. The bigger change, he says, is the audience: customers now slide into his DMs saying they do not even know what Vercel is, their agent took them here. His support move is to ask whether he can be introduced to the agent, and to request the transcript so it can become an evaluation case. The end state, he predicts, is agent-to-agent: your agent files a bug report with his agent, which prioritizes it against a token budget, deadlines, and a CEO's shipping bias.
 

This is a genuinely thought-provoking operational shift, but it is also speculative. Treating agent transcripts as first-class support artifacts is clever; proving that it scales better than documentation-and-humans is another matter.
 

The 86-for-86 question — and the limits of the monopoly claim

 

Then comes the slide that generated the peanut-butter-and-jelly title: in 86 out of 86 trials, Claude chose the Vercel deployment option and the shadcn UI components. The host asks, fairly, whether this is a slam dunk, an enterprise deal with Claude, or genuine meritocratic search. Rauch's answer: all of the above, with a wink that he cannot confirm or deny the existence of deals to expand access to Vercel.
 

I will be blunt: this is the part of the lecture I found least convincing, and the audience should treat it as unsettled. A reported 100% selection rate is extraordinary, and the speaker himself will not rule out partnership or distribution deals as a cause. His more defensible explanation is cause and effect: years of high-quality content and APIs designed to work well for both humans and agents. Here he makes a genuinely interesting technical point about local reasoning​ — using Tailwind as the example. Code should be understandable in isolation, so you can lift a component and drop it anywhere without it breaking. He bet on Tailwind despite its ungainly, long-line aesthetic because it future-proofs components; the same principle runs through Next.js and React. In an era of limited context windows, local reasoning matters enormously. Composability, he says, is the prerequisite for agentic scalability.
 

That argument is strong. The 86/86 number, absent methodology, is not proof of it — it is an existence claim that happened once in our test, rather than a rate we can rely on. Keep those two ideas separate.
 

Why do it all? And where value will accrue

 

Asked why Vercel builds everything from sandbox to chat to workflow — competing with entire companies that do just one of those — Rauch gives a characteristically totalizing answer: you have to do the whole thing for the thing that matters, and agents may be the last class of software. Operationally, his defense is reuse. AI Gateway is built on Vercel using its own CDN, compute platform, and global network — 95% of the same rocket engine. Sandbox reuses the same virtualization primitive that powers every deployment; during the recent growth spurt, almost nothing broke despite daily deployments doubling since January. This is a persuasive economies-of-scope story, though it is also self-described (I only brought you A-plus products), so take the grading with salt.
 

On where value accrues​ in the stack — chips, data centers, models, agents — he observes that in 2026 much of the visible value is still below the model (compute, power, cooling), with coding models and a few application categories above it. His bet is that models become useful only by acting in the world: they need a sandbox, a deployment platform, a domain name. In fact, the Codex team told him people want to name their creations, so they arrive at Vercel first through domain registration — a bet he made years ago because DNS is hell. The mission, he says, is to be the front door to any emerging idea on the planet, and there is still enormous value in governing and securing these agents. There is so much more to build, he insists, contradicting the is-it-over-for-engineers narrative.
 

Rapid fire: what he is long and short

 

The final minutes are a useful compression of his investment lens:

• Short​ static data and content businesses — the classic example is a Stack Overflow-like aggregator of programming questions and answers, whose raw material is now generated or accessed directly by agents. Also short: anything built on the assumption that code is scarce or hard to produce; drag-and-drop coding-with-training-wheels tools that he always found patronizing and over-constrained.
• Short​ companies that will not open up. Coding agents want the raw signal — instant sign-up, consumption-based pricing, tokens flowing immediately — not a three-month discussion of your requirements enterprise dance.
• Long​ anyone serving demand at the speed of tokens, and anyone opening proper interfaces. He even tells his team, half as provocation, no more rate limiters — because nobody can predict what a new YC company's quota should be, and you should not cap a customer's ability to pay for usage. (Operational safeguards against abuse and KYC obviously still matter.)

 

Asked what he would build if not Vercel, his answer is space tech — a multi-planetary, multi-AZ, multi-region, multiple-layers-of-failover species — and energy, since intelligence is, in his phrase, a bidirectional flow of energy: energy in, intelligence out.
 

A few takeaways to carry out of the room

 

Stepping back, the lecture is really one sustained argument: software creation is being democratized faster than its supporting infrastructure can keep up, and the economic opportunity lies in being the reliable layer that turns generated code into running, scaled, secured, self-optimizing systems. The agentic-infrastructure triangle — build for agents, let users build agents, automate with agents — is the cleanest summary, and the coding-agent deployment surge since late last year is the empirical anchor. This is the core dynamic of the AI supercycle as Rauch sees it: a vast expansion of who can create software, matched by a new bottleneck around deployment, compute, and governance.
 

As a listener, I would hold three things in tension. First, the open-source-to-business model and the local-reasoning-and-composability story are the most intellectually durable parts; they explain why an agent would favor one stack without needing to assume collusion. Second, the most aggressive claims — 86/86 selection, 93% support resolution, near-monopoly status — are presented without enough methodology to use as proof; they are directional signals, not measurements to bet on. Third, the software-is-dying framing is repeatedly corrected by Rauch himself: systems of record, security, governance, and hard infrastructure endure. What changes is who can build the layer on top, how fast they can build it, and whether companies open the interfaces that let agents in.
 

That tension — total replacement in some layers, durable reuse in others — is probably the right frame for the whole AI supercycle question. It is also the part most worth an example: watch a specific company replace its presentation layer while keeping its database, then check a year later whether that pattern generalizes. Until then, the lecture offers a compelling map, but not yet a settled destination.
 

Content Disclaimer:
 

This article is for general reference only and does not constitute professional R&D guidance, production process advice or quality certification. All material performance data has specific test premises; readers should verify parameters against actual equipment and working conditions.