Nine of eleven co-founders have left xAI. Two days before admitting the company was built wrong, Musk unveiled Digital Optimus — an AI agent designed to replace every white-collar worker. The company building the tool to automate office work cannot retain the people building it.
Five hundred fifty-five thousand GPUs. Eighteen billion dollars in hardware. Two gigawatts of power consuming a repurposed factory in Memphis, Tennessee. The largest AI training cluster on Earth. And on March 13, the man who built it admitted it was built wrong.
The Admission
Elon Musk posted a statement that restructured the narrative around one of the most well-funded companies in artificial intelligence: "xAI was not built right first time around, so is being rebuilt from the foundations up. Same thing happened with Tesla."
The admission came six weeks after xAI completed its merger with SpaceX — a deal that created a combined entity valued at one and a quarter trillion dollars. It came two days after the company announced Digital Optimus, the most explicit agent product any organization has described. And it came as nine of the company's eleven co-founders had departed.
The trigger was specific. After reviewing xAI's coding tools against competitors — Claude Code and Codex — Musk concluded that Grok was failing to keep pace. At the Abundance conference on March 12, he said it directly: "Grok is currently behind in coding." The next day, he announced the rebuild.
The Exodus
The departure list reads like a dissolution, not a transition.
Christian Szegedy — the Google scientist behind the Inception architecture — left in February 2025. Igor Babuschkin departed in August to found an AI safety venture firm. Greg Yang, who helped architect Grok's core systems, left in January 2026. Then the pace accelerated. Tony Wu and Jimmy Ba — specialists in reasoning and optimization — departed within twenty-four hours of each other in early February. Toby Pohlen, a former DeepMind researcher assigned to lead the company's most ambitious project, quit after sixteen days. Zihang Dai and Guodong Zhang followed in March.
Nine co-founders gone. Two remain.
Musk framed the departures as deliberate evolution — some people, he suggested, are "better suited for the early stages of a company and less suited for the later stages." The staffing data told a different story. In the same breath, he posted an apology to candidates xAI had previously declined or never interviewed, saying he and the company's head of talent were reviewing the entire interview history to reach back out to promising applicants.
The Vision
Two days before admitting the company was broken, Musk described what it was supposed to build.
Digital Optimus — internally called Macrohard, a deliberate shot at Microsoft — is a joint xAI-Tesla project to create an AI agent that performs any task a white-collar worker does on a computer. Accounting, human resources, data entry, administrative operations. Not an assistant. A replacement.
The architecture is a dual-process model borrowed from Daniel Kahneman. System 1 runs on Tesla's AI4 inference chip — processing the last five seconds of screen activity plus keyboard and mouse inputs, executing fast reactions at the perceptual level. System 2 runs on xAI's Grok, serving as what Musk described as "the master conductor with deep understanding of the world" — slow deliberation directing fast execution. The AI4 chip is priced at six hundred fifty dollars, minimizing dependence on Nvidia.
The deployment vision matched the ambition. Every Tesla equipped with an AI4 chip becomes a potential office worker when parked. Millions of dedicated units at Supercharger stations, drawing from seven gigawatts of available power. The user experience was targeted for September 2026 — six months from announcement.
Fortune reported the project was placed on hold within weeks. Pohlen's departure after sixteen days leading it triggered an exodus of approximately a dozen senior engineers. The most ambitious agent product any company has described was shelved before its lead engineer finished onboarding.
The Rebuild
Between the product announcement and the admission, xAI made a hire that captured the depth of the problem. Andrew Milich and Jason Ginsberg — co-heads of engineering and product at Cursor, one of the fastest-growing AI coding tools — joined xAI, reporting directly to Musk. Their mission: rebuild Grok's coding capabilities from the ground up.
The choice was revealing. Milich and Ginsberg did not come from a competitor Musk was trying to weaken. They came from the tool his own engineers were presumably using instead of Grok.
Meanwhile, auditors from Tesla and SpaceX were brought in to evaluate xAI's existing engineering staff. The rebuild extends beyond leadership into the ranks.
The Distance
The structural irony is precise. The company building the tool to replace all white-collar workers cannot retain its own.
xAI has two hundred fifty billion dollars in valuation, twenty billion in fresh capital from its Series E, and more GPUs than any competitor on Earth. It has two remaining co-founders. The Colossus cluster in Memphis — originally built in nineteen days, now expanding toward one million GPUs — represents unprecedented computational scale. The organizational capability to direct that scale is collapsing.
This is the singularity story that does not appear in pitch decks. The challenge is not compute, not capital, not even capability in isolation — it is whether any single organization can hold together under the pressure of building what comes next. Digital Optimus is the most honest description of the operational singularity anyone has offered: an AI agent that makes every office worker optional. The irony is that xAI is experiencing the displacement from the inside first. The talent it cannot retain is the talent the product is designed to make unnecessary.
Musk compared the rebuild to Tesla's production hell — the period when the company nearly went bankrupt before emerging with manufacturing discipline that no competitor has replicated. But Tesla's rebuild happened around a working product that needed to scale. xAI's rebuild is happening before the product works. The factory is built. The car is not.
The distance between Digital Optimus and xAI's execution is not technological. It is organizational. And that distance may be the most honest measure available of how far the operational singularity still has to travel.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)