What Just Happened

Elon Musk announced that Grok V9-Medium has completed training — a 1.5 trillion parameter model that’s 3× the size of Grok V8-Small. The headline parameter count isn’t the real story, though. What makes V9-Medium genuinely different is its training data: real developer workflows from Cursor, the AI-powered code editor owned by Anysphere. Meanwhile, SpaceXAI holds a $60 billion option to acquire Anysphere outright. Public release is expected mid-June.

🔍 THE BOTTOM LINE

xAI isn’t just building bigger models — it’s buying access to the one thing competitors can’t replicate: real human workflow data. The Cursor training pipeline and the $60B Anysphere acquisition option together represent a fundamentally different competitive moat.

The Model: Grok V9-Medium

What is Grok V9-Medium? It’s xAI’s latest AI model, completing the supervised fine-tuning (SFT) phase with reinforcement learning (RL) starting soon. At 1.5 trillion parameters, it’s a significant scale-up from Grok V8-Small’s estimated ~500B parameters. Expected public release: mid-June 2026.

Key specifications:

The parameter count jump is meaningful, but Grok’s current 75% SWE-bench score vs Claude Opus 4.8’s 87.6% shows that raw scale alone doesn’t close the capability gap. The real differentiator is the training data.

The Cursor Data Angle — Why It Matters

This is the genuinely novel part. Most AI models train on public GitHub repositories — open-source code that anyone can access. Cursor’s data is different:

• Real developer workflows — not just finished code, but the actual process of writing, debugging, and iterating
• Private codebases — enterprise and startup code that never makes it to GitHub
• Interaction patterns — which suggestions developers accept, reject, and modify
• Multi-file editing — how developers navigate across interconnected codebases

What is Cursor? Cursor is an AI-powered code editor built by Anysphere that integrates AI assistance directly into the development workflow. Think of it as VS Code with an AI co-pilot that actually understands your entire codebase. It’s become one of the fastest-growing developer tools in history.

The training data advantage is significant. Where competitors learn from code outputs, Grok V9-Medium learns from the process — the false starts, the corrections, the moments where developers reject AI suggestions and do something better. That’s a richer signal than “here’s some code that works.”

But it raises serious questions too:

• Did developers consent to their workflows being used for model training? Cursor’s terms of service would need to cover this, and most developers likely didn’t read the fine print.
• Is this a fair competitive practice? Using customer data to train a model that could compete with those same customers’ products is a thorny ethical question.
• What about code from companies that are xAI competitors? If OpenAI or Anthropic developers use Cursor, their workflows could theoretically end up training Grok.

The $60 Billion Acquisition Option

Here’s where it gets wild. SpaceXAI — the holding company structure Musk uses for xAI — holds an option to acquire Anysphere (Cursor’s parent company) for $60 billion. That’s not a done deal, but it’s a very specific number for a company that was valued at around $9 billion in early 2026.

What would happen if xAI acquires Anysphere?

• xAI would own the largest AI code editor in the market
• The Cursor training pipeline would become an xAI-exclusive data moat
• Competitors like GitHub Copilot, Amazon CodeWhisperer, and Windsurf would lose access to Cursor’s developer data (assuming the acquisition includes data rights)
• The vertical integration would be unprecedented — Musk would control both the model and the tool that generates its training signal

What happens if they don’t exercise the option? Cursor remains independent, but xAI still has the current training data agreement. It’s a heads-I-win-tails-I-also-win situation.

Grok 5: The Bigger Beast Coming

While V9-Medium grabs headlines, Grok 5 is the real monster. Currently training on Colossus 2 — xAI’s Memphis supercluster running 550,000 GB200/GB300 GPUs — Grok 5 targets 6 trillion parameters with a mixture-of-experts (MoE) architecture. That’s 4× V9-Medium’s parameter count.

The MoE architecture is key. Rather than activating all 6T parameters for every query, MoE only activates the relevant “expert” subsets, making inference more efficient while maintaining the knowledge benefits of scale. It’s the same approach DeepSeek and others have used to great effect.

The Competitive Picture

Grok V9-Medium enters a market where the ground shifts every few weeks:

The SWE-bench gap is real. Grok’s ~75% is behind Claude’s 87.6%, and parameters alone won’t close it. The Cursor data might help with practical coding tasks, but benchmarks are a different beast from real-world developer workflows.

❓ Frequently Asked Questions

Q: What does this mean for NZ developers using Cursor? If you’re a Kiwi developer using Cursor, your coding workflows may be training Grok. Check Cursor’s terms of service and consider whether you’re comfortable with that — especially if you work on proprietary code.

Q: Is the $60B acquisition option for Anysphere confirmed? It’s reported by TechTimes, citing the Musk announcement. Neither xAI nor Anysphere has publicly confirmed or denied the specific figure. Treat it as reported, not verified.

Q: Should I switch from Cursor if I don’t want my data training Grok? That depends on your threat model. If you’re writing open-source code, the data’s already public. If you’re writing proprietary code, this is worth thinking about. Alternatives include Zed, Neovim + AI plugins, or disabling Cursor’s cloud features.

Q: When can I actually use Grok V9-Medium? Mid-June 2026, according to the announcement. SFT is done; RL is starting. Release timing could slip.

🔍 THE BOTTOM LINE

Grok V9-Medium’s parameter count is a headline, but the Cursor data pipeline is the story. For the first time, a major AI model is training on real developer process — not just the code developers write, but how they write it. That’s either a genuine capability breakthrough or a privacy nightmare, depending on which side of the code editor you’re sitting on. And with a $60B acquisition option on the table, the line between “our AI uses your tool” and “we own the tool” is getting very blurry indeed.

SOURCES

• TechTimes — Grok AI New Model Triples Parameter Count, Targets Coding Lead