Ten of twelve. That is the number of original co-founders who have left xAI since March 2023. Simultaneously, xAI is training Grok 5 on Colossus 2 — the world's first gigawatt-scale AI supercomputer, housing over 555,000 NVIDIA GPUs across three buildings in Memphis, Tennessee. Those two facts are both true at the same time, and the tension between them is the actual story about where xAI stands heading into mid-2026.
Elon Musk posted in March 2026 that "xAI was not built right first time around, so is being rebuilt from the foundations up." He said this six weeks after Tesla invested $2 billion in xAI's Series E. He said this while Grok 5 — a 6-trillion-parameter model with a reported 10% probability of AGI, per Musk's own claim — is actively in training. The question worth sitting with is not whether Grok 5 will be impressive. It probably will be, at that scale. The question is whether compute can substitute for the institutional knowledge that walked out the door.
This article looks at both sides of that equation with the numbers that are publicly available — not the press releases, but the actual math on what xAI has and what it lost.
- Colossus 2: What the Infrastructure Actually Represents
- The Founder Exodus: Who Left, Why, and What It Means
- Grok 5: The Real Specifications Behind the Claims
- "Built Wrong": What Musk's Admission Actually Signals
- The Space Infrastructure Bet: Where xAI Is Staking Its Long-Term Position
- Where Grok Actually Stands Against Claude, GPT-5, and Gemini
- My Take
- FAQ
Colossus 2: What the Infrastructure Actually Represents
Thesis: xAI's compute advantage is real and historically unprecedented. The question is whether it's sufficient on its own.
Colossus 2 became operational in January 2026 as the first AI training cluster in the world to cross the 1-gigawatt threshold. It spans three buildings in Memphis, Tennessee, with a total target of 555,000 NVIDIA GPUs — roughly equal to an $18 billion GPU investment at current market rates. Capacity is being upgraded to 1.5GW by April 2026. For context, OpenAI's Stargate project targets 500,000 GPUs; Google's TPU clusters are estimated at around 100,000 equivalent units. xAI is not bluffing on the infrastructure side.
Grok 5 is training on this cluster right now. The model uses a Mixture-of-Experts architecture with 6 trillion total parameters — the largest publicly announced AI model ever. MoE architecture means only a subset of parameters activate per query, keeping inference costs manageable despite the scale. Training began in late September or early October 2025. The original Q1 2026 public release window was not met; the current estimate is a public beta between March and April 2026, with full API access in Q2 2026.
The Tesla Terafab project, confirmed by Musk to launch in late March 2026, includes infrastructure directly tied to xAI's Grok training operations — adding another layer of compute and power to an already exceptional setup. The merger with SpaceX, valued at $1.25 trillion, means Grok's compute ambitions now extend to plans for AI data centers in sun-synchronous low Earth orbit, where solar power is available 24/7 without the land-use, cooling, and regulatory constraints of terrestrial data centers.
Verdict on infrastructure: The compute case for Grok 5 is as strong as any model in the world. Google's Project Suncatcher analysis suggested that space-based AI data centers could reach cost parity with terrestrial facilities by 2035, based specifically on SpaceX's launch cost trajectory. That timeline, combined with xAI's ownership of both the model and the launch provider, is the long bet Musk is making. On a long enough runway, it's not an unreasonable one.
The Founder Exodus: Who Left, Why, and What It Means
Thesis: Losing 10 of 12 founders is not normal organizational evolution. The specifics of who left matter more than the headline number.
The departures did not happen in one event. They accumulated over roughly 18 months, accelerating dramatically in February and March 2026. The sequence, based on public reporting, runs as follows: Kyle Kosic (infrastructure lead) left for OpenAI in mid-2024. Christian Szegedy (Google veteran) departed February 2025. Igor Babuschkin left to found a venture firm in August 2025. Greg Yang stepped back citing health issues in January 2026. Jimmy Ba and Tony Wu — both research leads central to Grok's model quality — left the same week in February 2026. Guodong Zhang (head of the Imagine team) and Zihang Dai followed in March 2026. Remaining: Manuel Kroiss and Ross Nordeen.
| Co-founder | Role | When Left | Reason Cited |
|---|---|---|---|
| Kyle Kosic | Infrastructure Lead | Mid-2024 | Joined OpenAI |
| Christian Szegedy | Research (ex-Google) | Feb 2025 | Departed |
| Igor Babuschkin | Research | Aug 2025 | Founded venture firm |
| Greg Yang | Research | Jan 2026 | Health issues (Lyme) |
| Jimmy Ba | Research / Safety Lead | Feb 2026 | Reported tensions over model performance |
| Tony Wu | Research Lead | Feb 2026 | "Time for my next chapter" |
| Guodong Zhang | Head of Imagine Team | March 2026 | Blamed for coding product issues |
| Zihang Dai | Co-founder | March 2026 | Departed |
| Manuel Kroiss + Ross Nordeen — only 2 of 12 original co-founders remaining |
The FT reported that researchers are leaving because of burnout from Musk's "extremely hardcore" work demands — and because Anthropic and OpenAI are offering better terms to departing talent. 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 assist their own development work. That detail, reported by Let's Data Science citing internal sources, is the most revealing single data point in this whole story: xAI engineers were using a competitor's product to do their jobs.
Verdict on talent: Losing infrastructure leads and research safety leads in the same period is not equivalent to losing product managers. Jimmy Ba's research was foundational to Grok's architecture. Hiring Andrew Milich and Jason Ginsberg from Cursor — two people who know how to build coding tools developers love — is a reasonable response to the coding gap. It doesn't replace model quality researchers. Those are different skill sets.
Grok 5: The Real Specifications Behind the Claims
Thesis: The 6 trillion parameter headline is real. The AGI claim is not verifiable and likely isn't meant to be.
What is confirmed about Grok 5: 6 trillion total parameters in a Mixture-of-Experts architecture, active training on Colossus 2, native multimodal capability (text, images, video, audio), real-time X data access baked into the architecture, 1.5 million token context window, and training on Tesla's autonomous driving video data as a proprietary data source not available to competitors. Elon Musk confirmed the 6T parameter figure at the Baron Capital conference in November 2025.
What is marketing: the "10% probability of AGI" claim. AGI is defined as a system capable of human-level reasoning across any domain. No current AI system is close to that threshold by any rigorous measure. The claim functions as investor signaling, not technical specification. The AI research community has been broadly skeptical — R&D World noted that "parameter scaling alone cannot address" the fundamental challenges between current LLMs and genuine general intelligence. That's the accurate framing.
The more interesting Grok 5 claim is the Tesla video data angle. Grok 5 is reportedly training on real-world video from Tesla's fleet — billions of hours of driving data capturing object recognition, spatial reasoning, and real-world physics that text-trained models simply don't see. If video prediction generalizes to broader reasoning (which is the hypothesis, not yet verified), xAI may have found an asymmetric data advantage. If it doesn't generalize, the 6T parameters produces a very expensive model with the same reasoning ceiling as everyone else.
Verdict on Grok 5 specs: The architecture is genuinely ambitious and the training data is legitimately differentiated. The AGI probability is a press release. Independent benchmarks on real tasks — not xAI-selected benchmarks — are the only thing worth waiting for.
"Built Wrong": What Musk's Admission Actually Signals
Thesis: "We need to rebuild from the foundations" six weeks after extracting $2 billion from Tesla shareholders is a specific kind of statement. It has implications.
On March 13, 2026, Musk posted on X: "xAI was not built right first time around, so is being rebuilt from the foundations up." This statement landed six weeks after Tesla's $2 billion investment in xAI's Series E, and after SpaceX acquired xAI in a deal valuing the combined entity at $1.25 trillion. Tesla shareholders are already suing Musk for breach of fiduciary duty over xAI's founding — arguing he diverted AI talent and resources from Tesla to benefit his private company. Musk admitting the company needed a fundamental rebuild, after extracting billions from Tesla and SpaceX investors, adds a specific legal dimension to those challenges.
The specific trigger for the "rebuild" framing was xAI's coding product performance. Grok Code Fast 1, xAI's dedicated coding model, was falling behind Anthropic's Claude Code and OpenAI's Codex on benchmarks that professional developers actually use. Musk brought in SpaceX and Tesla "fixers" to audit teams — and let go of employees whose work was deemed inadequate. The hiring of Milich and Ginsberg from Cursor, who report directly to Musk, is the response: people who built the tool that professional developers reach for first, tasked with making Grok competitive in that specific category by mid-2026.
Verdict on the rebuild: The "not built right" admission is honest in a way that most AI lab communications aren't — which is worth acknowledging. It also raises a question that hasn't been answered publicly: what specifically was built wrong? Organizational structure? Model training approach? Product architecture? The vagueness matters, because the fix for a structural problem looks different from the fix for a technical one.
The Space Infrastructure Bet: Where xAI Is Staking Its Long-Term Position
Thesis: The space data center idea is not crazy. The timeline is aggressive. The competitive question is who builds them, not whether they get built.
Google's Project Suncatcher, a team studying solar power satellites for AI compute, published findings suggesting space-based data centers could reach cost parity with terrestrial facilities by approximately 2035, based on SpaceX's launch cost improvement curve. The same report noted that solar power is available 24/7 in low Earth orbit, eliminating the battery and intermittency problems that constrain terrestrial renewable power for AI compute. Nvidia had the same idea independently. xAI, now merged with SpaceX, would be positioned as both the AI model provider and the launch infrastructure provider for this transition — a structural advantage no other AI lab has.
Google is already moving: the first two satellites of their Project Suncatcher constellation are scheduled to launch in 2026. The technology is not theoretical. The competitive question in this scenario isn't whether space-based AI compute gets built — it's who builds it. xAI with SpaceX has an argument to be the default infrastructure provider. That's the actual strategic logic behind the $1.25 trillion merger, separate from Grok's current performance against Claude or GPT-5.
Musk's claim that xAI will be so far ahead in three years that rivals won't be visible without the James Webb telescope is the compressed version of this argument. Three years is aggressive. The 2035 estimate from Google's own team is the more conservative version. Somewhere in that range is probably where the actual transition happens — and xAI's position at the intersection of the AI model and the launch infrastructure is genuinely unusual.
Where Grok Actually Stands Against Claude, GPT-5, and Gemini
Thesis: Grok 4.20 is competitive in specific categories. The headline gap against Anthropic and OpenAI is in coding — which is specifically what the rebuild targets.
An independent third-party estimate cited in the transcript places xAI approximately seven months behind Anthropic, Google, and OpenAI in overall AI capability. That framing — which comes from researcher Peter Wilderford, not from any lab — is directionally useful even if the specific number is approximate. xAI's own Colossus advantage and the Grok 4.20 series have closed some of that gap, particularly in real-time search and financial benchmarks.
On LMArena, Grok 4.20 ranks second for search — very close to Claude Opus 4.6. On the Alpha Arena financial trading benchmark, Grok 4.20 places near the top, achieving 12.11% average returns in real-time stock trading. For video: Grok Imagine ranks in the top 10 for text-to-video, and first for video editing tasks. For coding: Claude Code and OpenAI Codex lead on SWE-Bench Verified, the benchmark professional developers trust most. The gap is not on all dimensions — it's concentrated in the specific area that professional developer workflows prioritize most.
| Category | Grok 4.20 Position | Current Leader |
|---|---|---|
| Real-time search | #2 (LMArena) | Claude Opus 4.6 |
| Financial trading (Alpha Arena) | Near top, 12.11% avg returns | Competitive |
| Video editing | #1 (Grok Imagine) | Grok leads |
| Text-to-video | Top 10 | Competitive |
| AI coding (SWE-Bench) | Below Claude, Codex | Anthropic / OpenAI |
| General reasoning | Est. 7 months behind top tier | Anthropic / OpenAI / Google |
Verdict on competitive position: Grok 4.20 is not a weak model. It's competitive on real-time search in a way that matters for specific workflows — as discussed in the Perplexity vs ChatGPT research comparison, real-time search capability is a distinct and valuable category. The coding gap is real and it's specifically what the Cursor hires are targeting. Whether mid-2026 is enough time to close that gap is the question nobody has answered.
My Take
The "xAI is seven months behind" framing is probably the wrong unit of measurement. That estimate treats AI capability as a single number on a linear scale — which is how labs want you to think about it, because it makes their benchmark wins look decisive. The actual picture is more fragmented: Grok leads on video editing, is competitive on real-time search, and is materially behind on coding. Saying xAI is "seven months behind" on coding is different from saying xAI is "seven months behind" in general — and the distinction matters because the Cursor hires were chosen specifically to address the coding gap, not the general capabilities gap.
The detail I keep coming back to is that xAI engineers were using Claude through the Cursor IDE to do their own work — before Anthropic cut off their access under its commercial terms. I've covered a lot of AI lab announcements on this site. I have not seen another case where a lab's engineers were apparently dependent on a competitor's product to develop their own. That's not a security failure. It's a capability signal. Hiring Andrew Milich and Jason Ginsberg from Cursor is a rational response — but it's a response to a problem that the 6 trillion parameters of Grok 5 cannot directly fix.
The space bet is the part of this story that I think gets insufficient analytical attention. Most coverage frames it as Musk making another bold claim. But Google's own Suncatcher team validated the basic economics, citing SpaceX's launch cost trajectory specifically. If the cost of putting compute into space reaches parity with terrestrial data centers around 2035, the question of who wins in AI gets significantly reframed — and xAI is the only lab that also owns the rocket company. That's a real structural advantage, assuming everything else goes reasonably well between now and then.
The honest question this all raises: can you rebuild the foundational culture of a lab by importing talent from other companies, while simultaneously training the most expensive model ever built on infrastructure that didn't exist two years ago? The answer is probably yes — eventually. The question is whether the timeline pressure from Anthropic, OpenAI, and Google's own scaling efforts leaves enough runway to get there.
🔑 Key Takeaways
- 10 of 12 original xAI co-founders have departed — only Manuel Kroiss and Ross Nordeen remain
- Grok 5 is in active training on Colossus 2 — the world's first 1GW+ AI supercomputer, 555,000 GPUs
- 6 trillion parameters, MoE architecture, Tesla video data as proprietary training source — the specs are real
- The AGI probability claim (10%) is investor signaling, not technical specification — treat it accordingly
- Anthropic cut off xAI's Claude access after discovering xAI engineers were using it via Cursor to do their own work
- Grok 4.20 leads on real-time search (#2 LMArena) and video editing (#1), lags on coding benchmarks
- Cursor's Andrew Milich and Jason Ginsberg hired specifically to close the coding gap by mid-2026
- The space data center bet is supported by Google's own Suncatcher team projections — ~2035 cost parity
- xAI is the only AI lab that also owns the launch infrastructure for that transition (SpaceX merger)
- Full Grok 5 release moved from Q1 to Q2 2026; public beta estimated March–April 2026
FAQ
- 💰 Claude API vs GPT-5 API: The Exact Token Volume Where One Saves You More Money (2026)
- 💻 I Switched to DeepSeek-R1 for Coding: 7 Things It Does Better Than ChatGPT-4o
- 🔍 I Used Perplexity Pro for 30 Days — 5 Things Surprised Me
- 🤖 Self-Hosting Llama 4 vs GPT-4o API: The Exact Volume Where It Saves Money
- 📊 I Replaced My SEO Workflow With AI Agents for 30 Days: The Brutal Truth
📚 Sources
TechCrunch — Half of xAI's Founding Team Has Left (Feb 2026) ·
Electrek — Musk Admits xAI Built Wrong (March 2026) ·
Let's Data Science — xAI Rebuild Analysis ·
AdwaitX — Grok 5 Colossus 2 Training Analysis (Feb 2026) ·
Reuters via Yahoo Finance — xAI Founder Departures (March 2026)
All data sourced from public reporting and official announcements. No proprietary access to xAI internal information. Benchmark data from LMArena and Alpha Arena public leaderboards.
The question nobody has answered publicly is this: at what point does compute stop compensating for the institutional knowledge that walked out with 10 of the 12 people who designed the system? The engineers who left know exactly which shortcuts were taken in Grok's architecture, which training decisions created the coding gap, and which design choices made the rebuild necessary. Grok 5 will almost certainly be a capable model — the infrastructure is real and the scale is historic. Whether it can close a gap against Anthropic and OpenAI whose teams are also scaling, also hiring, and also have not lost their founding researchers is a different question. The answer arrives in Q2 2026, when the benchmarks land — not before.
0 Comments