At Modal, an engineer who changes documentation or makes a basic fix doesn't wait for a colleague to approve it. "Why do you need a human to look at every piece of code?" Erik Bernhardsson, Modal's founder and CEO, asked on the Latent Space podcast, in an episode published October 22, 2026. Modal still does code reviews, he said, but treats them as optional.
That is one of several operating choices Bernhardsson described. Modal is an AI infrastructure company. It rents out computing power and storage, including scarce graphics processors (GPUs), so customers can run AI models, execute code in isolated "sandboxes" and sometimes train models without managing servers themselves. The publisher's episode notes say Modal reports passing $1 billion in annualized revenue run rate and has raised a $750 million Series D round at a $16 billion valuation. Bernhardsson said the company has grown about fourfold on an annualized basis over the past 12 months. Over the conversation he explained how Modal handles engineers who work with customers, product management, managers and code review, and what that pace of growth costs.
Code review as an option, tests as the safeguard
In a conventional software team, nearly every change waits for a second engineer to read and approve it before it ships. Bernhardsson thinks that habit often serves a different purpose. Code review, in his view, is in many cases people trying to "escape the accountability" of putting things into production. If a team has that culture, he said, it is a sign it needs to invest much more in continuous integration and continuous deployment: the automated systems that test each change and push it out.
He still asks for reviews himself. When he writes code, he said, he sometimes asks the authors of that part of the system whether his change fits how they think about it, and checks that he isn't fundamentally changing the architecture. But for documentation and basic changes, Modal tells staff not to wait, "especially in the era of AI."
That freedom depends on strong internal controls around testing, he said, and an infrastructure company needs more of them than most. Reliability is often measured in "nines" of uptime, such as 99.9% versus 99.99%, and Bernhardsson said Modal needs "one more nine than the average startup." Getting there means rate limits everywhere, systems decoupled from one another, no single point of failure and "extremely good test coverage."
AI has helped with that last part. A few years ago, Bernhardsson said, he considered test-writing one of the major things AI agents lacked, because good tests require stepping outside what you are building and playing devil's advocate. Now, he said, agents are "actually quite good at writing tests."
Engineers sent to customers, with an exit plan
Modal has expanded its forward-deployed engineers, or FDEs, engineers who work directly on customer projects. Bernhardsson said the company started the year with about four and now has "like, 24 or something like that."
Much of their work is tuning the performance of customers' large language models. He described FDEs sitting down with customers to train draft models and adjust parameters in serving engines, including SGLang, an open-source engine for serving language models. A draft model is a small, fast model used in a technique called speculative decoding: it proposes upcoming words, and the larger model checks them, which can speed up responses. The SGLang team's SpecForge documentation describes what training one involves: generating training data with the target model, training the draft on that model's internal signals, then exporting it to serve with SGLang.
FDEs also help customers work around problems, migrate code from other platforms during active deals, and look for customers who are over-provisioning, for example requesting more processor capacity or memory than they need. In those cases, Bernhardsson said, the message is roughly: you're spending a lot of money on Modal right now, and "we can actually help you save money."
He said Modal never intends to become "a service-heavy business." He listed three goals for the FDE team: help close deals, support customers, and, most important, bridge gaps in the product that should later become real features. "My question is always, like, long-term, how do we turn this into products so that we don't need an FDE for this?" he said.
Why there's no product-management team
Asked whether Modal hires product managers, Bernhardsson said no. He doesn't see FDEs as stand-in product managers either, though they cover the customer-facing part of the job. He split product work into three pieces: the customer-facing side; an internal, strategic side that decides which building blocks Modal should invest in next; and a middle layer that looks across 10 or 20 customers asking for slightly different things and works out the common pattern.
Some people at Modal do product work. Former developer-relations staff and engineers have moved into product roles after showing they could think strategically about what to build. "I'm actually not, like, anti-product management," he said. His objection is that many companies hired product managers "too eagerly," and he said you can see the result in products where eager managers came up with "widget-y things and just ship them."
Infrastructure makes the job harder, in his account. It requires deciding what the foundational building blocks are and where the company wants to be in one to two years, and avoiding "35 features that are, like, kind of similar looking" but each slightly different. He said that kind of thinking is very rare, and for that reason Modal hasn't aggressively tried to build a product-management team yet.
Managers who go back to coding
Bernhardsson, who earlier built Spotify's music recommendation system and was chief technology officer at the mortgage company Better.com, said team structures have changed a lot. He isn't convinced generative AI caused the change. Ten or 20 years ago, he said, everyone wanted to go into management because that brought more money, recognition and prestige. He sees a swing back toward individual contributors (ICs), the engineers who build rather than manage, which he called "very, very positive" for the field.
That shift began before generative AI, in his view, but AI has greatly sped it up by making code cheaper and easier to write. At Modal, he said, managers and ICs carry no difference in prestige; everyone is a "member of technical staff," and many former managers have become ICs because they want to build again. He said he still writes a lot of code, as do cofounder and CTO Akshat Bubna and nearly everyone on the technical team, and many people outside it write code too. That was the background to the code-review question: if non-technical staff commit code, who checks it?
Trying every coding tool
Modal doesn't standardize on one AI coding assistant. Bernhardsson said early use leaned on Cursor, Claude and OpenAI's Codex, and now many people use open-source tools and other setups. He said he has no strong preference and wants people to experiment with both frontier and open-source models. Because Modal now runs inference for many open-source models at large scale, the company is also encouraging staff to use its own internal model endpoints for agent-driven development, using its own product from the inside.
What fourfold growth costs
Bernhardsson said Modal recently crossed 200 people and will probably reach 250 by the end of the year. At another point he spoke of the company as already having 250 people. Doubling a team every year is not very hard, he said, but he called growth of about 4x "an unsustainable growth rate on a relative basis."
He named the costs. Culture drifts. "Islands of different working styles" form. People start very speculative projects; he said he sometimes asks an engineer what they are working on and finds something "kind of, like, fringe," which makes him ask whether that is the right use of their time. New hires also take time to get up to speed.
Size changes how he leads, too. At 10, 20 or 30 people, he said, the culture is implicit: everyone can read him and tell when he is annoyed or happy. At 250, people don't see that, and new employees may be nervous about talking to the CEO. He called the Dunbar number, the idea that people can keep only about 150 stable relationships, "more sort of fake than anything," at least the story behind it, but said something real still happens around 150 to 200 people, when laying the foundation of the culture becomes very important.
He expects growth to continue at perhaps 2.5 to 3 times, which he said makes about 700 people "or something like that" likely next year. Most new hires will be engineers, though Modal is also ramping up sales. Systems engineers who know Linux, file systems, networking and large-scale systems remain the core. Because Modal has moved into serving language models, it also wants people who work on GPU performance, squeezing the most output from each processor. Bernhardsson said that work still turns out to be mostly a systems problem, including managing models' short-term working memory (the KV cache) and routing requests.
Modal is also starting a research team for post-training, distillation and better quantization, techniques that refine, shrink or compress models after their initial training. Asked whether that work is for customers or for Modal itself, he said both, and that it follows the FDE pattern: hands-on work with customers first, then internal tools and eventually fully self-serve products. "I think we'll sort of arguably fail in a certain way if every time a customer needs to optimize a model, we have to go and work with them," he said.