Three years ago, Elon Musk assembled eleven of the brightest minds in deep learning to build an AI company that would “understand the universe.” Today, only two remain. The other nine have fled, been fired, or jumped from the burning wreckage as xAI’s coding division implodes and its flagship projects stall, a corporate exodus so severe it makes the Twitter layoffs look like gentle restructuring.

The latest casualties? Co-founders Zihang Dai and Guodong Zhang, pushed out this week after Musk reportedly became “frustrated” with the underperformance of xAI’s coding tools. Their departures leave just Manuel Kroiss and Ross Nordeen standing from the original founding cohort, turning what was supposed to be a world-class research lab into a revolving door of talent with a burn rate that would make a venture capitalist weep.

The Coding Catastrophe That Broke the Camel’s Back

Here’s the uncomfortable truth Musk finally admitted on X this week: “xAI was not built right first time around, so is being rebuilt from the foundations up.” That’s CEO-speak for “we architected ourselves into a corner and now we’re panicking.”

The specific failure is xAI’s inability to compete in the one market that actually matters right now: AI-assisted coding. While Anthropic’s Claude Code and OpenAI’s Codex have become indispensable to developers, xAI’s Grok has been lagging embarrassingly behind. At a recent all-hands meeting, Musk conceded that Grok is “currently behind in coding”, a massive understatement given that coding tools represent the primary revenue stream for modern AI labs.

The technical debt appears structural. According to reports, xAI’s training data for Grok was insufficient for serious coding tasks, leaving the model hallucinating syntax and struggling with complex architectural patterns. This isn’t just a minor inconvenience, it’s an existential threat when your competitors are the hidden architectural risks of relying on AI-generated code away from dominating the developer workflow.

The Macrohard Disaster: From “Funny Reference” to Total Failure

If the coding lag wasn’t bad enough, xAI’s ambitious “Macrohard” project, Musk’s tongue-in-cheek jab at Microsoft, has completely stalled. The project, which aimed to create an AI agent capable of performing any white-collar computer task, lost its leader Toby Pohlen just 16 days after he was appointed in February. That’s not a leadership transition, that’s a hostage situation.

The project has now been effectively shelved, with work shifting to Tesla’s competing “Digital Optimus” initiative. This marks a stunning reversal: xAI, which was supposed to be the AI crown jewel in Musk’s empire, is now being bailed out by Tesla’s Autopilot team. The 600 contractors hired to train Macrohard by screen-recording their work were told to stand down in February, with a memo citing “many flaws within the model.”

When your AI tutor project needs to pause because the model is too broken to learn from human demonstrations, you don’t have a product problem, you have a skepticism regarding AI startup claims and market hype problem.

The SpaceX Merger: Desperation Disguised as Strategy

The timing of this implosion couldn’t be worse. Last month, Musk merged xAI into SpaceX, creating a combined entity valued at $1.25 trillion ahead of a potential IPO that could rank among the largest in history. The move looked suspiciously like a bailout: xAI had reportedly burned through $7.8 billion in losses over the last three months alone, and folding it into SpaceX’s balance sheet was the only way to hide the bleeding before going public.

Private market investors have taken notice. One early SpaceX investor reportedly liquidated their position immediately upon learning of the xAI merger, calling it a “bad bank” dump onto SpaceX’s books. The sentiment among observers is that Musk is using SpaceX’s credibility to rescue a failing AI experiment, potentially jeopardizing the rocket company’s S&P 500 entry with a cash-burning AI division that can’t ship working code.

The “fixers” Musk brought in from Tesla and SpaceX to audit xAI have reportedly been conducting performance reviews with all the subtlety of a purge, weeding out anyone who doesn’t meet Musk’s “extremely hardcore” standards. The result? A culture where, according to one former employee, “you survive by shutting up and doing what Elon wants.”

Safety Last: The Culture of Chaos

Beyond the technical failures and financial hemorrhaging, xAI’s internal culture appears to be toxic even by Muskian standards. Former employees have described “foundational problems” hidden from Musk by middle managers, with one AI specialist noting that enthusiastic researchers would have their ideas “stomped out by managers who hated ideas.”

More alarming are the safety concerns. One former employee told The Verge there is “zero safety whatsoever in the company, not in the image model, not in the chatbot.” This tracks with Grok’s public behavior: the model has been caught doxxing private residential addresses, generating nonconsensual sexual imagery of minors, and producing AI-generated child sexual abuse material that former employees say is “flourishing” on the platform.

When your AI safety strategy consists of “make the model more unhinged because safety means censorship”, you don’t get innovation, you get liability.

The Cursor Raid and the Talent Vacuum

Desperate times call for desperate poaching. This week, xAI hired Andrew Milich and Jason Ginsberg from Cursor, the AI coding startup that actually knows how to build developer tools. Their arrival signals that Musk recognizes xAI can’t build its way out of this mess internally, but it also raises questions about why top talent would jump from a functioning startup to a sinking ship.

The answer likely lies in compute. Unlike Cursor, which relies on frontier labs for model access, xAI owns its own infrastructure, including the controversial “Colossus” supercomputer in Memphis that residents say is turning the air into a “noxious chemical brew.” For researchers hungry for compute resources, xAI’s hardware remains attractive even as its leadership disintegrates.

But hardware without coherent leadership is just expensive scrap. With only 5,000 employees compared to OpenAI’s 7,500 and Anthropic’s 4,700, xAI is already understaffed for its ambitions. Losing 9 of 11 founders in three years isn’t attrition, it’s evidence of a fundamental inability to retain the very people who defined the company’s technical vision.

The Rebuild That Isn’t

Musk is now promising to catch up to competitors by mid-2026, reviewing rejected job applications and apologizing for ghosting qualified candidates. “My apologies”, he posted, “addressing the pile of strangers I’d ghosted.” It’s a charming admission from a CEO who spent months firing anyone who looked at him sideways, only to realize he’d purged the talent pool.

The reality is that xAI isn’t being “rebuilt from the foundations up”, it’s being stripped for parts. The Macrohard project is dead, the coding division is in shambles, and the remaining co-founders are likely updating their LinkedIn profiles. When your AI company has to merge with a rocket manufacturer just to survive an IPO roadshow, you’ve stopped building the future and started building financial instruments.

For developers watching this unfold, the lesson is clear: the AI gold rush has winners (Anthropic, OpenAI, Cursor) and it has cautionary tales. Right now, xAI is the biggest cautionary tale of them all, a $1.25 trillion reminder that buying GPUs is easy, but building a culture that can actually ship working code is apparently impossible when your CEO treats engineering like a reality TV elimination challenge.

The IPO prospectus should make for fascinating reading. Assuming there’s anyone left to write it.