What does it mean when a company valued at $250 billion cannot hold onto a single one of its founders?
That is the question hanging over xAI this weekend. Manuel Kroiss, who led the company's pretraining team, has told people he is departing. Ross Nordeen — described by Business Insider as Elon Musk's "right-hand operator" and the last co-founder standing — left on Friday. With those two exits, the sweep is complete. All eleven co-founders of xAI, the AI lab Elon Musk assembled in 2023 with the explicit goal of building an AI smarter than any currently in existence, have now left the company.
This is not ordinary attrition. The people Musk recruited to build xAI were not mid-level engineers hired through a recruiting pipeline. They were elite AI researchers: Jimmy Ba, who co-authored the Adam optimization paper — the most-cited paper in AI with over 95,000 citations — departed in February after tensions over model performance targets. Igor Babuschkin, the chief engineer who came from Google DeepMind, left in mid-2025. Christian Szegedy, formerly of Google, resigned in early 2025. Tony Wu, who led reasoning, Greg Yang from Microsoft Research, Toby Pohlen and Guodong Zhang from DeepMind, Zihang Dai, Kyle Kosic — all gone. The cohort Musk assembled represented a rare concentration of research talent, the kind that takes years to build and a single bad culture to destroy.
The timing of the final departures is hard to separate from what Musk himself said on March 13. Speaking publicly, he acknowledged that xAI's AI coding tools simply did not work and that the underlying system needed to be rebuilt from the foundations up. "Not built right the first time around," was his exact phrasing. For the researchers who spent two years building those systems, hearing their company's founder describe the product as a fundamental failure — weeks after that same company was acquired for $250 billion by SpaceX — removed whatever residual reason existed to stay for the rebuild.
The AI talent market in 2026 gives them every reason to leave and every opportunity to land elsewhere. Meta has reportedly offered packages worth up to $300 million over four years to retain top AI researchers. OpenAI, Google DeepMind, and Anthropic are all aggressively expanding their research teams. OpenAI is preparing next-generation model releases and has been pulling talent from across the industry. Dario Amodei's Anthropic just confirmed a leaked frontier model and is chasing OpenAI's revenue lead with accelerated shipping cadences. Google DeepMind, under Demis Hassabis, recently shipped Gemini 3.1 Flash Live and is building real-time multimodal inference infrastructure for agentic deployments. Every serious competitor is running hot. The eleven researchers who left xAI represent exactly the kind of talent all of them would pay to acquire.
The structural situation at xAI makes the departures harder to reverse. In February 2026, SpaceX acquired xAI in an all-stock transaction that valued xAI at $250 billion and SpaceX at $1 trillion — the largest corporate merger by valuation in history. The deal placed xAI, X, and SpaceX under a single corporate umbrella, with SpaceX now targeting a mid-2026 IPO at a potential $1.75 trillion valuation. Tesla, separately, invested $2 billion in xAI's Series E round. Tesla shareholders are currently suing Musk for breach of fiduciary duty over that investment, arguing that he directed shareholder capital into his own private AI venture and then admitted it needed to be rebuilt.
What xAI still has: the Colossus supercomputer, built with more than 200,000 NVIDIA H100 GPUs, remains one of the largest AI training clusters in the world. Grok, the company's LLM, has distribution through X's hundreds of millions of users. The SpaceX merger provides capital access and engineering resources at a scale most AI labs cannot match. Infrastructure and compute are not nothing. NVIDIA's GPU supply chain, the core asset underlying every frontier model today — GPT, Claude, Gemini, Llama — can be deployed by whoever controls the hardware. xAI controls a lot of hardware.
But there is a meaningful distinction between having compute and knowing what to do with it. The difference between frontier AI labs in 2026 is not primarily GPU count. It is the quality of the research culture that decides how those GPUs are used — what experiments to run, which fine-tuning approaches to pursue, what architecture choices to make at the weights level, where to focus inference optimization. Those decisions require researchers who have built frontier models before, who understand the failure modes, who can move fast precisely because they have already made the expensive mistakes. That institutional knowledge was concentrated in xAI's co-founding team. It is now distributed across every other major lab in the industry.
Musk has rebuilt companies before from catastrophic positions. SpaceX nearly failed three times before reaching orbit. Tesla was weeks from bankruptcy in 2008. His track record in hardware-driven businesses, where technical risk is the primary variable and determination can substitute for lost talent, is extraordinary. The question is whether AI research operates by the same rules. Tesla and SpaceX build physical systems that respond to engineering iteration. A large language model's capabilities are shaped by the researchers making decisions about pretraining data, compute allocation, RLHF pipelines, and architectural experimentation. Musk's management style — which produces exceptional outcomes when paired with the right domain — appears to work less well in environments where the most valuable people have abundant options and low tolerance for instability.
The researchers who co-founded xAI chose to be there. They were not recruited under false pretenses; they came with full knowledge of who Musk was and what he had built. That every one of them ultimately chose to leave — not during a funding crisis, not during a period of competitive failure, but at the moment of a $250 billion valuation — is the most precise measurement available of what happened inside that company. Capital and compute cannot fix what they left behind.
Deep Dive
If this collapse of research talent at xAI concerns you, it's part of a broader pattern of AI labs competing for talent and resources that is reshaping the entire industry. Two recent pieces from this newsletter that give essential context:
Anthropic's Secret Weapon Just Leaked — And It Changes Everything — The accidental exposure of Anthropic's most capable unreleased model, and what it reveals about how Dario Amodei is positioning the company's research advantage against OpenAI.
Zuckerberg Just Declared War on Intel and x86 — And His Weapon Is a Chip Called AGI — Meta's partnership with Arm to co-develop the Arm AGI CPU goes deeper than a product launch: it is Zuckerberg building a hardware stack that reduces his dependence on both NVIDIA and Intel, permanently shifting the economics of AI infrastructure.
Originally on The Signal — free AI newsletter. Subscribe: newsletter.uddit.site
Top comments (0)