"The only remaining job for programmers in the future will be signing off on budgets for AI." The founder of Next.js: There will definitely be trillion-dollar one-person companies in the future.

I went on a business trip and returned to the company, only to find that my 12-person technical team, behind my back, had uprooted the entire underlying architecture from AWS and migrated it entirely to Vercel.

This is a real experience that Dara Ladjevardian, founder of Delphi, had not long ago. The team's reason for ditching AWS is extremely pragmatic: with autonomous coding agents and modern development infrastructure, they no longer need to pay high salaries to employ a full-time infrastructure engineer.

Later, in an exclusive Delphi podcast hosted by him, Dara and Guillermo Rauch, who is the founder of Vercel and the creator of Next.js and Socket.io, laid bare the full engineering and business cards behind this boom of Autonomous Coding Agents:

Back in the day, Instagram, with only 13 people, leveraged AWS S3 to outperform Facebook which had thousands of engineers; and on the Agentic infrastructure, companies with only 1 employee but a market capitalization of 1 trillion USD will definitely be born in the future.

The software industry is undergoing drastic reshuffling: traditional drag-and-drop low-code has been completely dead, since code generation itself has become free; the next core purchaser of software is not human beings at all, but agents that call APIs millions of times every day. When the marginal cost of writing code, fixing bugs, and performance optimization drops to zero in all aspects, the only remaining job for human engineers will be "capital allocation".

💡 Quick Overview of Key Points:

• Single-person trillion-dollar companies are bound to emerge: Instagram, with only 13 people, defeated industry giants head-on with the help of cloud storage back then; with programming agents bringing a thousand-fold boost to human efficiency leverage, the era of running a super software company with a single person is becoming a reality.
• Low-code tools are dying out across the board: Drag-and-drop design was originally a compromise for "people who cannot write code", but today, code not only has zero marginal cost, but is also the most efficient medium that unifies the expression of video, 3D, mathematics and logic, so redundant abstraction layers are no longer necessary.
• The core customers of software are shifting from humans to agents: Software used to serve programmers, then non-technical personnel (such as v0), and now directly serves agents. Any tool that does not provide clean APIs, CLIs and MCP (Model Context Protocol) will be quickly eliminated, just like those that refused to access the World Wide Web back in the day.
• The ultimate role of human programmers is "capital allocator": Specific development, testing and performance tuning are completely taken over by agents, and the only task for humans is to allocate tokens, computing power and budget for agents according to business goals, and make the decision to stop when marginal returns diminish.
• Trust is an outcome, not an input: Do not try to build a brand by posting on Twitter, holding Meetups and doing PR. Just like "abs are made in the kitchen", long-term trust from developers only comes from boring basic skills such as system security, tireless bug fixing, minimalist design and clear documentation.
• Set up a "Scream" channel in Slack to face up to nasty problems: The vast majority of founders are avoiding the worst parts of their products. Vercel has a dedicated internal "Scream" channel that fully receives internal error reports and user negative reviews, which is monitored by the CEO in person, forcing the team to face up to system defects.

The full text of the interview is as follows:

Abandoning AWS behind the CEO's back: Small teams no longer need full-time operation and maintenance

Host: Guillermo, I'm very glad to have you here.

Six months ago, Delphi was rewriting its underlying codebase. I happened to be away from the office on a business trip, and before I left, AWS was still our only infrastructure provider.

When I came back from the trip, the team had migrated the entire underlying system to Vercel behind my back. To be honest, shifting from a mature giant to another platform, which was not well-known for general underlying infrastructure at that time, made me very nervous. But it turned out that the team's decision was extremely wise, and this product saved us unimaginable time and capital.

I would like to ask first: as a latecomer, how did you initially decide to enter the market to compete with cloud giants? And how did you fight this battle?

Guillermo Rauch: I always follow one principle for product building: it must come from real personal interest and first-hand experience.

In the early days, I worked at a company that deployed software to traditional physical data centers, and later the concept of "cloud" began to gain popularity. As an engineer with more than ten years of experience, I even felt stupid when I first started using the cloud. To get a product running, I often had to spend a whole month grappling with all kinds of infrastructure, network and configuration tools.

Every time I finished a tedious task, I would habitually review it: What did I learn from this operation? Can I solidify it into a tool or infrastructure to make the same task much easier next time?

In many dimensions, the cloud is one of the greatest inventions of mankind, providing unlimited computing capacity. But for the vast majority of developers around the world, this set of underlying infrastructure is still too complex and too anti-human.

Before the industry generally started talking about AI, Vercel was working on one thing: can we turn it into autonomous driving? Now people call it "Agentic infrastructure", but a few years ago, we internally called it "autonomous driving infrastructure". Its positioning is to make cloud deployment as natural as driving an autonomous car.

So I never thought about snatching those old solutions that I left behind head-on. What I want to build is the next-generation product.

Host: So during the decision-making process, you didn't even feel afraid of competition from the giants?

Guillermo Rauch: Not really. The evolution of the technology world is always completed through the replacement of generation-based platforms.

The previous generation of technology usually does not die completely overnight, it just gradually loses the core attention of the industry. Just like everyone uses Slack widely today, but people still send emails occasionally; you can even see fax machines in the corner of the office once in a while.

I try to think about what is the technology for the future. Whenever a new wave of innovation emerges, what Vercel wants to do is to become the best base to carry it, rather than helping enterprises to move 20-year-old outdated applications pointlessly.

The current boom of programming agents has proved this point. Many large enterprises around the world access Vercel because their internal engineers have begun to fully use development agents, and these agents require extremely lightweight and instantly responsive infrastructure.

For startups, it is far more effective to take on the upcoming wave in advance than to turn back and fight to the death with the old wave that is losing momentum.

One person plus AI agent is fully capable of supporting a trillion-dollar company

Host: This really changes the rules of the game. We only have 12 people, and it is very difficult for a startup team to recruit a top full-time operation and maintenance engineer, but now we don't need this position at all.

Guillermo Rauch: My goal is to enable those companies that are "extremely large in scale but extremely small in team size".

When I first came to Silicon Valley to start a business, I witnessed the rise of Instagram in a co-working space. The whole team had only 13 people, but it directly impacted Facebook, the giant with thousands of engineers.

The key infrastructure that supported them at that time was the newly launched Amazon S3. They didn't need to deploy redundant racks in the computer room, configure disk arrays, or set up power supply equipment, they only needed to call one line of API to store hundreds of millions of photos.

I firmly believe that companies with only 1 employee but a market capitalization of 1 trillion USD will definitely be born in the future. This is not to deny the significance of teamwork, but today an ordinary person plus a team of programming agents can already complete the work that used to require hundreds or even thousands of people. This human efficiency leverage is being infinitely enlarged.

Host: When promoting your infrastructure agents, are you focusing on serving those small teams that have reaped the benefits, or trying to convince large enterprises to migrate out of AWS completely?

Guillermo Rauch: There is a very interesting "Trickle-up effect" here.

Just like middle-aged people always want to stay young, traditional large enterprises are also staring at those cutting-edge 12-person small teams, and they are eager to have that kind of delivery efficiency. By serving the most radical innovative teams around the world, we have naturally attracted the attention of large customers instead.

Sometimes this transmission is realized through mergers and acquisitions - large companies acquire startups built on Vercel, and the technology stack penetrates in accordingly; sometimes the architecture of large enterprises themselves is forced to face modernization transformation.

For example, many years ago, I strongly advocated "Headless Architecture" - completely decoupling front-end display from underlying business APIs. Many traditional enterprises did not understand it at that time. But in today's agent era, everyone is actively embracing this architecture: software must expose clean APIs, command-line CLIs and MCP protocols. Only in this way can agents freely call on it and generate brand new interactive interfaces.

The popularization of code agents has made the world's development logic quickly move closer to the architecture we initially set.

The fastest way to explain Vercel's value is to "calculate money for customers"

Host: It sounds like your early positioning was very broad, and you finally achieved business growth by focusing on vertical fields?

Guillermo Rauch: Yes. When we first entered the market, we even naively publicized to the outside world: "We are just like a complete AWS with a better experience, or Google Cloud with a better experience."

This kind of broad statement often yields little results. Startups must not be mediocre in everything, the world only pays for the most specialized products.

So I narrowed the front line, and focused my energy on the most penetrable scenario - e-commerce.

For e-commerce, the business logic is extremely simple and straightforward: Time is money. Every improvement in page loading speed will be accompanied by a synchronous rise in checkout conversion rate. When I can help merchants improve their front-end performance by 200%, they can see real revenue growth on their financial statements the next day.

This set of logic completely breaks down technical communication barriers. When I face the CEO or CIO of a traditional enterprise, I no longer start from the underlying technology, but directly calculate the return on investment (ROI) for them.

Many listed companies clearly stated in their public financial reports: "By upgrading the underlying system to Vercel, our transaction conversion rate has directly brought revenue increments." We have broken through the public perception in the e-commerce track, and then completed concentric circle expansion to other fields.

This is very similar to Tesla's business plan back then: first make an ultimate Roadster to sell to a very small number of people, polish the product strength to the extreme, and then use the existing advantages to penetrate the mass market layer by layer.

Trust is like abs: it is the product of the kitchen, not the result of scrolling Twitter

Host: Many developers have extremely high loyalty and faith in Vercel, and the outside world even thinks you have a natural personal charisma. How do you view the establishment of brand trust?

Guillermo Rauch: In the post-AGI era, trust will be the core business barrier of enterprises. But trust is by no means an input, but the final output.

Many founders always think that "building a brand" means posting more on Twitter, holding several offline gatherings, asking PR to write articles, or making a good-looking official website.

This is like training abs in the gym. Many people think that they can get abs by doing crunches crazily, but the real common sense is: abs are made in the kitchen. The core variables that determine your figure are nutrition, calorie gap, sleep and long-term daily self-discipline, not just doing a few crunches.

Running a business is exactly the same. Developers' trust in a platform does not come from how many Meetups you hold, but from solid basic skills at the underlying level: whether the system is safe enough, whether bugs can be fixed in seconds, whether the documentation is thorough enough, and whether the interface operation is absolutely minimalist.

Users' trust in you is established subconsciously: maybe it's a high-quality long technical article you wrote, maybe it's the experience of second-level rollback when you encounter an unexpected downtime in the production environment. Five to seven tiny positive experiences add up, and finally condense into instinctive business trust.

Face up to nastiness and defective products: set up a "Scream" channel in Slack

Host: Many peers are emphasizing design sense now, but Vercel introduced the aesthetic standards of consumer-grade products into developer tools very early. In the startup period when progress had to be rushed, what supported you to make this heavy investment?

Guillermo Rauch: Back then, even writing a good README in the GitHub repository, designing a high-precision logo for open source projects, or even making an independent official website with animation demonstration for tools was regarded as a deviant and crazy idea in the industry. But today these have become the standard configuration of the whole industry.

But the process of pursuing quality is extremely tormenting. It requires you to maintain a physiological level of pickiness and dissatisfaction with the products you deliver.

Many founders, when running a company, do not really face up to the worst and most dilapidated problems in the system. They are like people who sit in the gym scrolling through their phones, seemingly spending enough time there, but actually not working on the core pain points at all.

How long has it been since you went through those hundreds of negative review questionnaires? Did you dare to take the initiative to contact those paying customers who angrily cancelled their accounts because the system crashed?

To fight against this human weakness, we built a dedicated Slack channel inside Vercel, named "Scream about internal developer experience".

I set it as the highest priority reminder. Any internal employee or external developer who encounters a nasty, anti-human system experience, the error message will be fully synchronized to this channel. I force myself and the whole team to stare at the most harsh feedback every day, to face up to the defects of the system.

Low-code is dead: because code is completely free now, and everything is code

Host: Vercel initially served hardcore programmers, then expanded to non-technical personnel through v0, and now it has ushered in a brand new third type of customers - AI agents. How does this reshape your product logic?

Guillermo Rauch: The technical threshold has been decreasing, but we never do this to accommodate those who do not want to pursue the ultimate.

Agents are breaking the so-called boundary between "technical personnel" and "non-technical personnel", and everyone will eventually converge into one role: Builder.

In terms of internal software procurement, we recently decisively got rid of several traditional SaaS suppliers. Many Low-code/No-code tools are dying at an accelerated pace, because the original intention of this type of tools is to provide a set of visual drag-and-drop components for people who cannot write code.

But today's general premise has completely changed: not only has the marginal cost of code dropped to zero, but everything is code.

You can generate 3D scenes with code, render videos with code, and describe complex mathematics and business logic with code. Code is an all-mighty carrier. With programming agents, the workflow that originally needed to be pieced together by 4 to 5 dedicated low-code tools can now be directly written into code and deployed to Vercel to run in seconds.

The core customers of software have changed:

Tools that do not access MCP and agents will be eliminated immediately

Host: Now that agents are becoming mainstream users, how does this change the interaction mode between software and the outside world?

Guillermo Rauch: Any tool that refuses to provide interfaces for agents should be decisively abandoned.

If a software does not provide clear machine APIs, no command-line interfaces, and does not support MCP (Model Context Protocol