Three years ago, Elon Musk assembled twelve researchers to build the AI lab he said the world actually needed. He called it xAI, positioned Grok as the no-holds-barred alternative to ChatGPT and Claude, and poured billions into Colossus, the Memphis supercluster that briefly held the title of world's largest AI training facility. Last week, he posted on X that the whole thing "was not built right first time around."
The quote is blunt by anyone's standard. For an AI lab that just closed a merger valuing it at $250 billion, it's extraordinary.
The Exact Words Musk Used
"xAI was not built right first time around, so is being rebuilt from the foundations up."
He followed that with an apology of sorts: "Many talented people over the past few years were declined an offer or even an interview [at xAI]. My apologies."
Both posts landed on March 12, 2026, the same day The Information reported that xAI had hired two of Cursor's most senior engineers to run the rebuild of Grok's coding product. Musk said he expects xAI to "catch up and exceed our competitors" in coding by mid-2026.
That is an ambitious timeline. Right now, by most measures, Grok's coding tools are not close.
What Broke the Coding Team
The immediate trigger for the restructuring was xAI's performance against Anthropic's Claude Code and OpenAI's Codex. According to reporting by the Financial Times, Musk became increasingly frustrated that Grok Code Fast 1 — xAI's dedicated coding model — was prioritizing speed over reasoning depth and falling behind both rivals on the benchmarks that matter to professional developers.
The frustration found specific targets. Co-founder and engineer Guodong Zhang, who led image generation and had been responsible for parts of the coding product, announced his departure on March 12 with a post saying: "Last day at xAI. Wild journey past three years but excited about next chapter." Zihang Dai left the same week. Both, according to people familiar with the situation, departed after being held accountable for the coding team's underperformance and relieved of their primary duties by Musk.
That brought the total co-founder count down to two. Out of twelve.
The Co-Founder Exodus: A Complete Picture
The departures did not happen overnight. They started quietly in late 2025 and accelerated into a flood by February 2026.
| Co-Founder | Area of Work | Departure |
|---|---|---|
| Igor Babuschkin | Core research, early Grok | Early 2026 |
| Kyle Kosic | Infrastructure | Early 2026 |
| Christian Szegedy | Research | Early 2026 |
| Greg Yang | Theoretical research | Jan 2026 (health) |
| Jimmy Ba | Model training (Grok) | Feb 10, 2026 |
| Tony Wu | Research | Feb 10, 2026 |
| Toby Pohlen | Engineering | Feb 27, 2026 |
| Zihang Dai | Coding product | Mar 2026 |
| Guodong Zhang | Image generation / coding | Mar 12, 2026 |
| Additional founder | — | 2025–2026 |
Remaining: Manuel Kroiss and Ross Nordeen.
Jimmy Ba's exit drew particular attention. A University of Toronto professor whose research was foundational to Grok's architecture, Ba departed amid reported tensions over Musk's demands for faster model performance gains. Tony Wu left the same day. The week of their departure, Fortune ran the headline: "Half of xAI's founding team has left."
That was February. Since then, four more founders have gone.
The causes cited across reporting are consistent: burnout from Musk's "extremely hardcore" working demands, Tesla and SpaceX executives sent in to audit teams and identify underperformers, and an environment where being blamed for competitive shortfalls became grounds for exit. Several engineers left after receiving offers from Anthropic and OpenAI, companies where the model quality bar and the organizational stability were higher.
The Two People Musk Hired to Fix It
Andrew Milich and Jason Ginsberg are not AI researchers. They are product and engineering executives who took a raw coding tool at Cursor and turned it into the most-used AI development environment among professional engineering teams, reaching $2 billion in annualized revenue.
We covered Cursor's $2B revenue milestone in detail here.
Milich and Ginsberg co-founded Skiff, a privacy-focused document collaboration platform, and served as its CEO and CTO respectively. When Notion acquired Skiff in 2024, Milich moved to lead Notion Mail. Both then joined Cursor in June 2025, where they shared the role of Head of Engineering and Product — running the full product operation that drove the company's growth.
They are now at xAI and SpaceX, both reporting directly to Musk.
Their background is notable for what it is and what it is not. Milich came to Cursor from Notion after the Skiff acquisition. Before Skiff, he built collaborative document tools. Ginsberg spent time at Apple before co-founding Skiff, later served as a Sequoia Capital Scout while running the company, and holds a Stanford MS in Computer Science. Neither is a model researcher — they are builders of developer-facing products that people actually want to use every day. xAI's problem with its coding product is not purely a model problem. Grok Code Fast 1 achieves 70.8% on SWE-Bench Verified. The architecture works. What Cursor built was not a better model — it was a better product around a model. Musk appears to have reached the same conclusion.
The Third Hire: Devendra Singh Chaplot
Announced on March 14, 2026, the day after the Cursor hires became public, Devendra Singh Chaplot is joining xAI and SpaceX to work directly on Grok model training.
Chaplot's credentials are striking. He completed a PhD in Machine Learning at Carnegie Mellon University, where he focused on autonomous navigation. He then co-founded Mistral AI, contributing directly to the training of Mistral 7B, Mixtral 8x7B, and Mistral Large — the models that put European AI on the map. He led Mistral's multimodal team for Pixtral 12B. Most recently, he was a founding member of Thinking Machines Lab, the AI startup built by Mira Murati, OpenAI's former CTO.
Where Milich and Ginsberg are tasked with rebuilding the product, Chaplot appears to be tasked with the underlying model work — particularly in areas that connect to robotics and multimodal reasoning, both priorities Musk has flagged for xAI's next phase.
The Context: SpaceX, Tesla, and a $250B Admission
The timing of Musk's admission carries extra weight because of what happened six weeks earlier.
On February 2, 2026, SpaceX acquired xAI in an all-stock deal. xAI was valued at $250 billion. SpaceX was valued at $1 trillion. The combined entity, at $1.25 trillion, was described as the largest merger in history. The stated rationale involved orbital data centers and Starlink's infrastructure — but the financial reality is that the merger also gave xAI investors a path to liquidity through SpaceX's anticipated IPO.
Tesla had already invested $2 billion in xAI's Series E round in January, taking preferred stock. After the SpaceX-xAI merger closed, Tesla received FTC clearance on March 11 to convert that $2 billion stake into a minority SpaceX equity position.
Two days later, Musk said xAI was not built right.
The sequence matters. Tesla shareholders had just placed a $2 billion bet on xAI's trajectory — and Musk's public acknowledgment of structural failure came before the ink on Tesla's converted stake had dried. Electrek, which follows Tesla closely, ran the story under the headline: "Musk admits xAI 'not built right' — weeks after Tesla invested $2 billion."
Is This Chaos or Strategy?
Both, probably — but which reading you land on depends on which part of the company you're looking at.
At Tesla, early leadership turned over repeatedly before the Model 3 production ramp. At SpaceX, the first three Falcon 1 launches failed before the fourth succeeded. Musk's pattern is not to stabilize organizations and then accelerate — it is to tear them down to first principles when he decides the foundation is wrong, then rebuild at speed.
The co-founder exodus, from that view, is not necessarily evidence of a company failing. It may be evidence of a company being remade. The researchers who built Grok's early architecture were exactly the people needed to get to Grok 3 and Grok 4. Whether they were the right people to build a developer product that competes with Claude Code on a daily-use basis is a separate question.
But there are genuine warning signs that this reading is too generous.
The departures of Jimmy Ba and Tony Wu, researchers central to Grok's model quality, are harder to frame as strategic pruning. Losing Mistral-pedigree talent and CMU ML researchers who left for burnout reasons removes capabilities that product executives cannot replace. Milich and Ginsberg know how to build tools developers love. They cannot personally close a gap on SWE-Bench.
And Anthropic's response to xAI's competitive position adds a detail worth noting: in early January 2026, Anthropic cut off xAI's access to Claude models after discovering xAI staff had been using Claude through the Cursor IDE — apparently to inform their own development work. The block was cited under Section D.4 of Anthropic's Commercial Terms of Service, which prohibits using Claude to build competing products. That xAI engineers were using Claude to do their jobs is itself a signal about the competitive gap.
What This Means for Developers Who Chose Grok
For developers who have built workflows around Grok Code Fast 1, the near-term picture is uncertain.
Grok Code Fast 1 is fast — processing at 92 tokens per second with a 256,000-token context window, it outperforms Claude Sonnet on raw response latency by a meaningful margin. Developers who prioritize interactive speed over peak accuracy have a reasonable case for staying. The model is not bad. The benchmarks are competitive at the frontier speed tier.
But the rebuild announcement signals that xAI itself regards the current product as a false start. The architecture is being reconsidered. The team running it has been replaced. The mid-2026 target for "catching up and exceeding competitors" implies that parity does not exist yet.
For a developer choosing between Claude Code and Grok right now, the honest framing is this: Anthropic has a stable research team, a product that senior engineers have publicly endorsed as their daily driver, and no leadership crisis in progress. xAI has a bolder promise, a faster model, and a CEO who just admitted the company was built wrong.
That is not a case for abandoning Grok. It may be a case for waiting to see what Milich and Ginsberg actually ship.
The Bottom Line
Elon Musk has publicly admitted structural failure at a company he founded and valued at $250 billion. Ten of its twelve original co-founders are gone. The last two co-founders who remain watched Musk bring in engineers from the outside to rebuild the core product from scratch.
The Cursor hires are credible. Milich and Ginsberg built something developers actually used every day, at scale, at a company that went from zero to $2B in revenue. If anyone can build a Grok coding product that competes with Claude Code on user experience, they can.
But the gap between a credible hire and a shipped product that closes a benchmark deficit is exactly the kind of gap that takes time to cross. Musk has set mid-2026 as the deadline. The co-founders who built the original foundation are gone. The developers who depend on Grok will find out if that bet pays off.
Sources
- TechCrunch — 'Not built right the first time': Musk's xAI is starting over again, again (March 13, 2026)
- CNBC — Elon Musk says xAI must be 'rebuilt' as co-founder exodus continues (March 13, 2026)
- Bloomberg — Musk Pledges to Rebuild xAI as Another Co-Founder Departs (March 13, 2026)
- Futurism — Elon Musk Says He's Epically Screwed Up at xAI, Is Rebuilding "From the Foundations" (March 2026)
- Inc. — Elon Musk Admits xAI Wasn't Built Right and Is Rebuilding the Company From Scratch (March 2026)
- CNBC — SpaceX-xAI merger valued at $1.25 trillion, largest merger of all time (February 3, 2026)
- CNBC — Musk's xAI loses second co-founder in two days as Jimmy Ba departs (February 10, 2026)
- Bloomberg — xAI Co-Founder Toby Pohlen Is Latest Executive to Depart (February 27, 2026)
- Zee News — IIT Bombay alum Devendra Chaplot joins xAI & SpaceX to build 'superintelligence' (March 2026)
- Electrek — Musk admits xAI 'not built right' — weeks after Tesla invested $2 billion (March 13, 2026)
- Winbuzzer — Anthropic Blocks Unauthorized Claude Harnesses and xAI Access (January 2026)
- Investing.com — Musk pushes out more xAI co-founders amid coding division struggles (March 2026)