If AI Can Code,Why Are Big‑Tech Founders Coding Again?

Zuckerberg moved his desk. Brin assembled a strike team. Vembu gave up the CEO title. Something deeper than nostalgia is driving the founders back to the terminal.

Something quietly paradoxical is happening inside the most powerful technology companies on earth. The people who stepped back from the keyboard years ago, who delegated engineering to armies of developers while they focused on strategy, fundraising, and board rooms, are rolling up their sleeves again. Not because the tools regressed. Because the tools became so capable that being away from them started to feel dangerous.

Mark Zuckerberg physically relocated his desk to sit inside Meta's AI lab. Sergey Brin surfaced at Google to lead what insiders call a "coding strike team." Sridhar Vembu handed the Zoho CEO role to someone else so that he could go back to writing and reviewing code. These are not symbolic acts. They are signals of something structural.

The Return of the Founder Engineer

Meta President Dina Powell McCormick let it slip publicly at the Semafor World Economy Summit. Zuckerberg, she said, had moved his desk to sit alongside Alexandr Wang and Nat Friedman, the co-leads of Meta's newly formed Superintelligence Labs, and was "coding all day long." The Financial Times separately reported that this amounts to five to ten hours of hands-on coding and code review every week. For a CEO of a company employing tens of thousands of engineers, that is a striking allocation of time.

"I think he feels so strongly that he has to understand it at that level to really think about how our model can be the strongest it can be."

DINA POWELL MCCORMICK, META PRESIDENT, AT SEMAFOR WORLD ECONOMY SUMMIT

Google's Sergey Brin presents a slightly different variation on the same theme. The Information reported that Google is assembling an elite coding group with Brin directly involved, with the explicit ambition of driving an "AI takeoff," a state in which AI systems can increasingly improve and code themselves. Brin had been largely absent from Google's day-to-day for years before this reengagement. The fact that he returned specifically for this effort says everything about where the centre of gravity in tech has shifted.

In India, Zoho's Sridhar Vembu made a structural choice rather than just a behavioural one. In January 2025 he stepped down as CEO to take the title of chief scientist, explicitly reorienting his role around writing and reviewing code. The software is not an afterthought at Zoho; it is the business. And Vembu, apparently, decided that proximity to it mattered more than the executive chair.

The Strange Logic of the AI Coding Era

On the surface this makes no sense. We are in the middle of a genuine acceleration in AI-assisted software development. Tools like Claude Code, Cursor, and GitHub Copilot have compressed the distance between intent and working software to a degree that would have been unthinkable three years ago. Anthropic has gone further still: Claude Code's creator Boris Cherny claimed the latest version was written entirely by Claude Code itself. A recursive loop of automation writing automation.

Dario Amodei, Anthropic's CEO, has argued publicly that coding as a discipline may eventually disappear altogether, replaced by systems that convert human intent directly into running software, with no programmer as an intermediary. If you believed that trajectory fully, the rational move for a senior executive would be to invest less attention in code, not more.

But the founders are not betting against Amodei's thesis. They are responding to an implication inside it.

CONTEXT

Anthropic's Claude Design is extending the same compression that Claude Code applied to engineering into the design layer, posing a direct challenge to platforms like Adobe and Figma. The stack being automated is broadening, not narrowing. Each layer that AI absorbs makes the remaining human layer more consequential.

When Prompts Are Architecture

Here is the thing about AI-generated code that doesn't appear in press releases about productivity gains: small differences in how you specify a problem can produce radically different outcomes. The same intent, expressed two slightly different ways, can yield production-quality output or subtly broken output that will take hours to diagnose. This is not a bug that will get patched away. It is intrinsic to how language models work.

A leader who cannot read the code that AI agents produce is flying blind. They can see the output, a feature, a demo, a deployment, but they cannot evaluate the quality of the substrate. And in the current moment, the substrate is moving fast enough that judgment gaps compound quickly. What the founders understand is that fluency with code is no longer just an engineering concern. It is a strategic information advantage.

The analogy that comes to mind is the transition in manufacturing from purely management-driven factories to the era of lean production, where executives who understood the factory floor had qualitatively better mental models of what was possible and what was fragile. The founders are trying to maintain that floor-level literacy during a transformation of the factory itself.

Compression and Its Consequences

A parallel trend is applying pressure from the other direction. AI is compressing software work, not eliminating it, but reducing the headcount required between an idea and its execution. Tasks that once required entire squads of engineers can now be handled by a single developer with the right set of tools and agents. Meta itself has set internal targets reflecting this: by mid-2026, 65% of engineers in core product groups are expected to be generating more than 75% of their code using AI.

The layoff data is stark. Layoff.fyi tracked over 81,000 employees let go by 97 tech firms in the early months of 2026. LinkedIn confirmed a 20% decline in overall tech hiring since 2022, though its chief legal officer was careful to attribute this to rising interest rates rather than AI displacement. The actual cause is likely a confluence of both forces, and separating them cleanly is probably impossible.

There is also a cost dimension that rarely gets discussed in the enthusiasm around vibe coding. AI-generated code is not free. Every token carries a price. At scale, teams generating enormous volumes of AI-assisted output will face economics that look different from what they modelled when they first adopted these tools. Senior leaders who have personally used these systems understand that constraint intuitively in a way that leaders who have only read about it do not.

Coding does not disappear. It moves up the stack. The future may need fewer programmers, but it will demand more leaders who can think like one.

Not Nostalgia. Strategy.

It would be easy to read these stories as sentiment: the old guard longing for the garage days, finding comfort in the familiar click of a keyboard after years of spreadsheets and shareholder calls. That reading is probably wrong, or at least incomplete.

What these founders are actually doing is repositioning themselves at the highest-leverage point in their companies during a period of maximum uncertainty. When the tools change this fast, the people who understand them at a technical level have better intuitions about where to place bets, which startups are actually as impressive as they sound, which internal projects have hidden brittleness, and which capabilities are genuinely transformative rather than well-marketed.

Routine coding tasks will decline. Large engineering organisations may shrink further. But the higher order skills, system design, judgment, the ability to translate loosely specified ideas into precise executable instructions, are becoming more valuable, not less. The founders returning to the terminal are not trying to compete with their engineers. They are trying to preserve their ability to lead them through a shift that has no obvious historical precedent.

What Changes, and What Doesn't

The companies that will navigate this transition best are probably the ones where the distance between the people making decisions and the technology those decisions depend on is smallest. That is a different kind of organisational advantage than the ones we typically discuss — not scale, not distribution, not brand. Epistemic proximity to the system being built.

For the rest of the industry, the layer below the hyperscalers, the mid-size software companies, the agencies and consultancies, the lesson is the same, even if the mechanisms are different. The era of managing software from a distance is ending. The managers who understand what their AI tools are actually producing, and why, will consistently outperform those who treat the output as a black box whose quality they cannot personally assess.

The founders already figured this out. They are coding again not because AI can't code. They are coding again precisely because it can.

SOURCES

• Business Insider

· Financial Times

· The Information

· Semafor World Economy Summit

· LiveMint Tech Talk

· layoff.fyi

· LinkedIn / TechCrunch